From eaddf18ee1d44457c65d76d02e16e90a25f9f828 Mon Sep 17 00:00:00 2001 From: Daniil Lyakhov Date: Wed, 4 Oct 2023 07:59:34 +0200 Subject: [PATCH] [PTQ][Torch][Native] Experimental tensor statistics migration (#2117) ### Changes * Torch Init range moved to experimental TensorCollector * Torch PTQ migrated to experimental TensorCollector * Threesigma/MAD aggregators are implemented as experimental tensor collectors for Torch/OV backends * `no_outlier_map` function refactored to not depend on shape of input tensor to make it possible to unify Torch and OV backends as they are using different shape statistics ### Reason for changes * To use all experimental reducers/aggregators (tensor collectors) in Torch/Torch PTQ backend ### Related tickets 109665 ### Tests * Tests are updated * Test register_unnamed_inputs * Test MedianAbsoluteDeviationAggregator/PrecentileAggregator/PostAggregateHook * Test mixed min max TensorCollector creation function TODO: * Confirm that new `no_outlier_map` performs as good as previous implementation --- nncf/common/tensor_statistics/collectors.py | 137 ++-- nncf/common/tensor_statistics/statistics.py | 7 + .../common/tensor_statistics/collectors.py | 353 ++++++++--- nncf/onnx/graph/node_utils.py | 4 +- nncf/onnx/statistics/collectors.py | 97 ++- nncf/openvino/graph/node_utils.py | 40 +- nncf/openvino/statistics/collectors.py | 118 ++-- .../algorithms/bias_correction/algorithm.py | 2 +- .../algorithms/bias_correction/backend.py | 5 +- .../bias_correction/onnx_backend.py | 7 +- .../bias_correction/openvino_backend.py | 7 +- .../algorithms/channel_alignment/algorithm.py | 6 +- .../algorithms/channel_alignment/backend.py | 4 +- .../channel_alignment/openvino_backend.py | 4 +- .../fast_bias_correction/algorithm.py | 2 +- .../fast_bias_correction/backend.py | 5 +- .../fast_bias_correction/onnx_backend.py | 7 +- .../fast_bias_correction/openvino_backend.py | 7 +- .../fast_bias_correction/torch_backend.py | 12 +- .../algorithms/min_max/openvino_backend.py | 16 +- .../algorithms/min_max/torch_backend.py | 109 ++-- .../algorithms/smooth_quant/algorithm.py | 16 +- .../algorithms/smooth_quant/backend.py | 10 +- .../smooth_quant/openvino_backend.py | 12 +- nncf/scopes.py | 6 +- nncf/tensorflow/quantization/init_range.py | 4 +- .../tensor_statistics/collectors.py | 55 +- .../tensorflow/tensor_statistics/reduction.py | 8 +- nncf/torch/quantization/algo.py | 14 +- nncf/torch/quantization/init_range.py | 125 ++-- nncf/torch/statistics/aggregator.py | 4 +- nncf/torch/tensor.py | 3 + nncf/torch/tensor_statistics/algo.py | 38 +- nncf/torch/tensor_statistics/collectors.py | 592 +++++++++++++----- nncf/torch/tensor_statistics/statistics.py | 13 +- .../test_reducers_and_aggregators.py | 231 +++++-- .../experimental/test_statistic_collector.py | 204 +++++- tests/common/test_statistics_aggregator.py | 18 +- .../quantization/test_quantizer_config.py | 8 + ...y => test_calculation_quantizer_params.py} | 0 .../quantization/test_quantizer_config.py | 8 + .../test_reducers_and_aggregators.py | 14 +- .../openvino/native/test_model_transformer.py | 8 +- tests/openvino/native/test_node_utils.py | 10 +- .../native/test_statistic_collector.py | 40 ++ .../test_calculate_quantizer_parameters.py | 2 +- .../test_templates/test_channel_alignment.py | 9 +- .../test_templates/test_quantizer_config.py | 13 +- .../test_templates/test_smooth_quant.py | 6 +- .../test_tensor_statistics.py | 8 +- tests/torch/ptq/helpers.py | 4 +- .../ptq/test_calculation_quantizer_params.py | 3 +- tests/torch/ptq/test_ptq_params.py | 22 +- tests/torch/ptq/test_quantizer_config.py | 30 +- .../ptq/test_reducers_and_aggregators.py | 103 +++ tests/torch/ptq/test_statistic_collector.py | 55 ++ tests/torch/quantization/test_range_init.py | 369 ++++++++--- .../test_tensor_statistics.py | 171 ++--- tests/torch/test_statistics_aggregator.py | 2 +- 59 files changed, 2328 insertions(+), 859 deletions(-) rename tests/openvino/native/quantization/{test_fq_configurations.py => test_calculation_quantizer_params.py} (100%) create mode 100644 tests/torch/ptq/test_reducers_and_aggregators.py create mode 100644 tests/torch/ptq/test_statistic_collector.py diff --git a/nncf/common/tensor_statistics/collectors.py b/nncf/common/tensor_statistics/collectors.py index 907ae30fec8..ae9c2e536b2 100644 --- a/nncf/common/tensor_statistics/collectors.py +++ b/nncf/common/tensor_statistics/collectors.py @@ -12,7 +12,7 @@ from abc import ABC from abc import abstractmethod from collections import deque -from typing import Callable, List, Optional, Tuple, Union +from typing import List, Optional, Tuple, Union import numpy as np @@ -21,14 +21,13 @@ from nncf.common.tensor import TensorType from nncf.common.tensor_statistics.reduction import get_per_channel_history -ReductionShape = Tuple[int] -MaskedReduceFN = Callable[[NNCFTensor, Union[int, tuple, list], NNCFTensor, bool], NNCFTensor] +ReductionAxes = Tuple[int] class TensorStatisticCollectorBase(ABC): """Collector estimate statistics at the quantization point based on the provided reduction shape.""" - def __init__(self, reduction_shape: Optional[ReductionShape] = None, num_samples: Optional[int] = None): + def __init__(self, reduction_shape: Optional[ReductionAxes] = None, num_samples: Optional[int] = None): """ Initializes Tensor Statistic Collector @@ -101,7 +100,7 @@ class OfflineTensorStatisticCollector(TensorStatisticCollectorBase): """Collects statistics in offline regime by storing the data and aggregating it afterwards.""" def __init__( - self, reduction_shape: Optional[ReductionShape] = None, num_samples: int = None, window_size: int = None + self, reduction_shape: Optional[ReductionAxes] = None, num_samples: int = None, window_size: int = None ): super().__init__(reduction_shape, num_samples) self._samples = deque(maxlen=window_size) @@ -117,7 +116,7 @@ class NNCFCollectorTensorProcessor(ABC): @staticmethod @abstractmethod - def reduce_min(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def reduce_min(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: """ Computes minimum of elements across dimensions of NNCFTensor. @@ -130,7 +129,7 @@ def reduce_min(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = Fa @staticmethod @abstractmethod - def reduce_max(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def reduce_max(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: """ Computes maximum of elements across dimensions of NNCFTensor. @@ -175,7 +174,7 @@ def max(x1: NNCFTensor, x2: NNCFTensor) -> NNCFTensor: @staticmethod @abstractmethod - def mean(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTensor: + def mean(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims=False) -> NNCFTensor: """ Computes the mean of elements across given dimensions of NNCFTensor. @@ -188,7 +187,7 @@ def mean(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTe @staticmethod @abstractmethod - def median(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTensor: + def median(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims=False) -> NNCFTensor: """ Computes the median of elements across given dimensions of NNCFTensor. @@ -199,9 +198,11 @@ def median(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCF :return: Reduced NNCFTensor. """ - @staticmethod + @classmethod @abstractmethod - def masked_mean(x: NNCFTensor, axis: Union[int, tuple, list], mask: NNCFTensor, keepdims=False) -> NNCFTensor: + def masked_mean( + cls, x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], mask: NNCFTensor, keepdims=False + ) -> NNCFTensor: """ Computes the masked mean of elements across given dimensions of NNCFTensor. @@ -214,9 +215,11 @@ def masked_mean(x: NNCFTensor, axis: Union[int, tuple, list], mask: NNCFTensor, :return: Reduced NNCFTensor. """ - @staticmethod + @classmethod @abstractmethod - def masked_median(x: NNCFTensor, axis: Union[int, tuple, list], mask: NNCFTensor, keepdims=False) -> NNCFTensor: + def masked_median( + cls, x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], mask: NNCFTensor, keepdims=False + ) -> NNCFTensor: """ Computes the masked median of elements across given dimensions of NNCFTensor. @@ -251,6 +254,16 @@ def unstack(x: NNCFTensor, axis: int = 0) -> List[NNCFTensor]: :return: List of NNCFTensor. """ + @staticmethod + @abstractmethod + def squeeze(x: NNCFTensor, dim: Optional[Union[int, Tuple[int, ...]]] = None) -> NNCFTensor: + """ + Remove axes of length one from x. + + :param x: NNCFTensor to squeeze. + :param axis: Selects a subset of the entries of length one in the shape. + """ + @staticmethod @abstractmethod def sum(tensor: NNCFTensor) -> TensorElementsType: @@ -264,18 +277,42 @@ def sum(tensor: NNCFTensor) -> TensorElementsType: @staticmethod @abstractmethod def quantile( - tensor: NNCFTensor, quantile: Union[float, List[float]], axis: Union[int, tuple, list], keepdims: bool = False + tensor: NNCFTensor, + quantile: Union[float, List[float]], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, ) -> List[TensorElementsType]: """ - Compute the quantile-th percentile(s) of the data along the specified axis. + Compute the quantile(s) of the data along the specified axis. :param tensor: Given NNCFTensor. - :params quantile: Percentile or sequence of percentiles to compute, which must be between + :params quantile: Quantile or sequence of quantiles to compute, which must be between 0 and 1 inclusive. + :param axis: Axis or axes along which the quantiles are computed. + :param keepdims: If True, the axes which are reduced are left in the result + as dimensions with size one. + :returns: List of the quantile(s) of the tensor elements. + """ + + @classmethod + @abstractmethod + def percentile( + cls, + tensor: NNCFTensor, + percentile: Union[float, List[float]], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, + ) -> List[TensorElementsType]: + """ + Compute the percentile(s) of the data along the specified axis. + + :param tensor: Given NNCFTensor. + :params percentile: percentile or sequence of percentiles to compute, which must be between + 0 and 100 inclusive. :param axis: Axis or axes along which the percentiles are computed. :param keepdims: If True, the axes which are reduced are left in the result as dimensions with size one. - :returns: List of the quantile-th percentile(s) of the tensor elements. + :returns: List of the percentile(s) of the tensor elements. """ @staticmethod @@ -289,27 +326,47 @@ def mean_per_channel(x: NNCFTensor, axis: int) -> NNCFTensor: :return: Reduced NNCFTensor. """ - @classmethod + @staticmethod + def logical_or(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: + """ + Computes the element-wise logical OR of the given input tensors. + Zeros are treated as False and nonzeros are treated as True. + + :param input_: The input tensor. + :param other: The tensor to compute or with. + :return: Result of elementwise or operation between input_ and other tensor. + """ + + @staticmethod + def less(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: + """ + Return the truth value of (x1 < x2) element-wise. + + :param input_: The input tensor. + :param other: The tensor to compute or with. + :return: Result of elementwise less operation between input_ and other tensor. + """ + + @staticmethod @abstractmethod - def no_outliers_map(cls, x: NNCFTensor, fn: MaskedReduceFN, axis: int = 0, alpha: float = 0.01) -> NNCFTensor: + def sub(a: NNCFTensor, b: NNCFTensor) -> NNCFTensor: + """ + Returns result of a substract b operation. """ - Computes quantiles [alpha, 1 - alpha] on given tensor, masks all elements that - are smaller that alpha and bigger than 1 - alpha quantile and applies - given masked reduction function fn. - :param tensor: Given NNCFTensor. - :param fn: Masked reduce operation from the same NNCFCollectorTensorProcessor class. - :param axis: Axis along which the reduction function is computed. - :params alpha: Minimal percentile to filter outliers outside the range - [quantile(alpha), quantile(1 - alpha)]. Must be between 0 and 1. inclusive. - :returns: Result of given masked reduction function on filtered from outliers NNCFTensor. + @classmethod + @abstractmethod + def zero_elements(cls, x: NNCFTensor) -> NNCFTensor: + """ + Returns binary mask from the input x which equal true for all elemets that are smaller than + corresponding machine epsilon. """ class MinMaxStatisticCollector(OnlineTensorStatisticCollector): """Collector estimates min of minimum values and max of maximum values.""" - def __init__(self, use_abs_max: bool, reduction_shape: ReductionShape, num_samples: int = None): + def __init__(self, use_abs_max: bool, reduction_shape: ReductionAxes, num_samples: int = None): super().__init__(reduction_shape, num_samples) self._use_abs_max = use_abs_max self._tensor_processor = self._get_processor() @@ -353,7 +410,7 @@ def __init__( self, use_per_sample_stats: bool, use_abs_max: bool, - reduction_shape: ReductionShape, + reduction_shape: ReductionAxes, num_samples: int = None, window_size: int = None, ): @@ -407,7 +464,7 @@ def __init__( use_abs_max: bool, use_means_of_mins: bool, use_means_of_maxs: bool, - reduction_shape: ReductionShape, + reduction_shape: ReductionAxes, num_samples: int = None, window_size: int = None, ): @@ -447,17 +504,15 @@ class MeanStatisticCollector(OfflineTensorStatisticCollector): Collector that aggregates statistics as mean along a pre-assigned axis. """ - def __init__( - self, reduction_shape: ReductionShape, num_samples: Optional[int] = None, window_size: Optional[int] = None - ) -> None: + def __init__(self, channel_axis: int, num_samples: Optional[int] = None, window_size: Optional[int] = None) -> None: """ - :param reduction_shape: The shape for the reduction while statistics collection. - For the MeanStatisticCollector this parameter contains the main axis. + :param channel_axis: The main axis for the reduction while statistics collection. :param num_samples: Optional parameter for statistic collection that regulates the number of samples that will be processed. :param window_size: Optional maximum length for the statistic collection """ - super().__init__(reduction_shape, num_samples) + super().__init__(num_samples=num_samples) + self._channel_axis = channel_axis self._tensor_processor = self._get_processor() self._all_values = deque(maxlen=window_size) self._all_shapes = deque(maxlen=window_size) @@ -468,10 +523,10 @@ def _get_processor(): pass def _register_input_common(self, x: NNCFTensor): - if self._reduction_shape == 0: + if self._channel_axis == 0: self._all_values.append(self._tensor_processor.batch_mean(x)) else: - self._all_values.append(self._tensor_processor.mean_per_channel(x, self._reduction_shape)) + self._all_values.append(self._tensor_processor.mean_per_channel(x, self._channel_axis)) self._all_shapes.append(x.shape) def _reset(self): @@ -536,7 +591,7 @@ class PercentileStatisticCollector(OfflineTensorStatisticCollector): def __init__( self, percentiles_to_collect: List[float], - reduction_shape: Optional[ReductionShape] = None, + reduction_shape: Optional[ReductionAxes] = None, num_samples: int = None, window_size: int = None, ): @@ -561,7 +616,7 @@ class MeanPercentileStatisticCollector(OfflineTensorStatisticCollector): def __init__( self, percentiles_to_collect: List[float], - reduction_shape: Optional[ReductionShape] = None, + reduction_shape: Optional[ReductionAxes] = None, num_samples: int = None, window_size: int = None, ): diff --git a/nncf/common/tensor_statistics/statistics.py b/nncf/common/tensor_statistics/statistics.py index 0f6d0d1aad3..1bc62db3464 100644 --- a/nncf/common/tensor_statistics/statistics.py +++ b/nncf/common/tensor_statistics/statistics.py @@ -20,6 +20,8 @@ class TensorStatistic(ABC): """Base class that stores statistic data""" + TENSOR_STATISTIC_OUTPUT_KEY = "tensor_statistic_output" + @staticmethod @abstractmethod def tensor_eq(tensor1: TensorType, tensor2: TensorType, rtol=1e-6) -> bool: @@ -63,6 +65,9 @@ def __eq__(self, other: "MeanTensorStatistic") -> bool: class MedianMADTensorStatistic(TensorStatistic): + MEDIAN_VALUES_STAT = "median_values" + MAD_VALUES_STAT = "mad_values" + def __init__(self, median_values, mad_values): self.median_values = median_values self.mad_values = mad_values @@ -74,6 +79,8 @@ def __eq__(self, other: "MedianMADTensorStatistic") -> bool: class PercentileTensorStatistic(TensorStatistic): + PERCENTILE_VS_VALUE_DICT = "percentile_vs_values_dict" + def __init__(self, percentile_vs_values_dict): self.percentile_vs_values_dict = percentile_vs_values_dict diff --git a/nncf/experimental/common/tensor_statistics/collectors.py b/nncf/experimental/common/tensor_statistics/collectors.py index 08bf166f32d..68cd4442c0a 100644 --- a/nncf/experimental/common/tensor_statistics/collectors.py +++ b/nncf/experimental/common/tensor_statistics/collectors.py @@ -13,16 +13,22 @@ from abc import abstractmethod from collections import defaultdict from collections import deque -from typing import Any, Dict, List, Optional, Set, Tuple, TypeVar, Union +from typing import Any, Dict, List, Optional, Set, Tuple, Type, TypeVar, Union from nncf.common.tensor import TensorType from nncf.common.tensor_statistics.collectors import NNCFCollectorTensorProcessor from nncf.common.tensor_statistics.collectors import NNCFTensor -from nncf.common.tensor_statistics.collectors import ReductionShape +from nncf.common.tensor_statistics.collectors import ReductionAxes +from nncf.common.tensor_statistics.statistics import MeanTensorStatistic +from nncf.common.tensor_statistics.statistics import MedianMADTensorStatistic +from nncf.common.tensor_statistics.statistics import MinMaxTensorStatistic +from nncf.common.tensor_statistics.statistics import PercentileTensorStatistic +from nncf.common.tensor_statistics.statistics import RawTensorStatistic from nncf.common.tensor_statistics.statistics import TensorStatistic from nncf.quantization.advanced_parameters import AggregatorType InplaceInsertionFNType = TypeVar("InplaceInsertionFNType") +AggregationAxes = Tuple[int, ...] class TensorReducerBase(ABC): @@ -31,16 +37,16 @@ class TensorReducerBase(ABC): the specified rule. Could handle tensors inplace or out of place. """ - def __init__(self, reduction_shape: Optional[ReductionShape] = None, inplace: bool = False): + def __init__(self, reduction_axes: Optional[ReductionAxes] = None, inplace: bool = False): """ - :param reduction_shape: Reduction shape for reduction calculation. Equal to list(range(len(input.shape))) + :param reduction_axes: Reduction axes for reduction calculation. Equal to list(range(len(input.shape))) if empty. :param inplace: Whether should be calculated inplace or out of place. - """ - self._reduction_shape = reduction_shape + self._reduction_axes = reduction_axes self._tensor_processor: NNCFCollectorTensorProcessor = self._get_processor() self._inplace = inplace + self._keepdims = True @property def inplace(self): @@ -95,34 +101,44 @@ def __call__(self, x: List[NNCFTensor]): def __eq__(self, __o: object) -> bool: return ( isinstance(__o, self.__class__) - and self._reduction_shape == __o._reduction_shape + and self._reduction_axes == __o._reduction_axes and self._inplace == __o.inplace ) def __hash__(self) -> int: - return hash((self.__class__.__name__, self.inplace, self._reduction_shape)) + return hash((self.__class__.__name__, self.inplace, self._reduction_axes)) - def _get_reduction_shape(self, tensor: NNCFTensor) -> Union[int, Tuple[int, ...]]: - if self._reduction_shape is not None: - return self._reduction_shape + def _get_reduction_axes(self, tensor: NNCFTensor) -> ReductionAxes: + if self._reduction_axes is not None: + return self._reduction_axes return tuple(range(len(tensor.shape))) -class TensorAggregatorBase: +class AggregatorBase: """ - Tensor aggregator is designed to receive (register) calculated statistics and + Aggregator is designed to receive (register) calculated statistics and aggregate them in terms of NNCFCollectorTensorProcessor operations. """ - def __init__(self, tensor_processor: NNCFCollectorTensorProcessor, num_samples: Optional[int] = None): + def __init__( + self, + tensor_processor: NNCFCollectorTensorProcessor, + aggregation_axes: Optional[AggregationAxes] = None, + num_samples: Optional[int] = None, + ): """ :param tensor_processor: Backend-specific tensor processor. + :param aggregation_axes: Axes along which to operate. + Registered statistics are stacked along zero axis, + axes >=1 correspond to recieved statistic axes shifted left by 1. :param num_samples: Maximum number of samples to collect. Aggregator skips tensor registration if tensor registration was called num_samples times before. Aggregator never skips registration if num_samples is None. """ self._tensor_processor = tensor_processor + self._aggregation_axes = (0,) if aggregation_axes is None else aggregation_axes + self._keepdims = False self._num_samples = num_samples self._collected_samples = 0 self._container = [] @@ -187,8 +203,8 @@ class TensorCollector: def __init__(self, statistic_container: Optional[TensorStatistic] = None) -> None: self._reducers: Set[TensorReducerBase] = set() - self._aggregators: Dict[Tuple[int, int], TensorAggregatorBase] = {} - self._stat_container_kwargs_map: Dict[str, Tuple[int, int]] = {} + self._aggregators: Dict[Tuple[int, int, int], AggregatorBase] = {} + self._stat_container_kwargs_map: Dict[str, Tuple[int, int, int]] = {} self._stat_container = statistic_container self._enabled = True @@ -224,7 +240,7 @@ def register_statistic_branch( self, container_key: str, reducer: TensorReducerBase, - aggregator: TensorAggregatorBase, + aggregator: AggregatorBase, reducer_output_port_id: int = 0, ) -> None: """ @@ -290,6 +306,14 @@ def register_inputs(self, inputs: Dict[int, List[NNCFTensor]]) -> None: if reducer_hash in reduced_inputs: aggregator.register_reduced_input(reduced_inputs[reducer_hash][reducer_port_id]) + def register_input_for_all_reducers(self, input_: NNCFTensor) -> None: + """ + Registers given input_ in each avaliable statistic collection branch. + + :param input_: Tensor input to register. + """ + self.register_inputs({hash(reducer): [input_] for reducer in self._reducers}) + def _aggregate(self) -> None: result = {} for ( @@ -315,7 +339,7 @@ def get_statistics(self) -> Union[TensorStatistic, Dict[str, Any]]: if not self._stat_container: return kwargs - return self._stat_container(**kwargs) + return self._build_statistic_container(self._stat_container, kwargs) def get_inplace_fn_info(self) -> List[Tuple[Any, int]]: """ @@ -338,7 +362,7 @@ def any_stat_out_of_place(self) -> bool: """ return any(not reducer.inplace for reducer in self._reducers) - def replace_aggregator(self, key: Tuple[int, int, int], aggregator: TensorAggregatorBase) -> None: + def replace_aggregator(self, key: Tuple[int, int, int], aggregator: AggregatorBase) -> None: """ Friend method that replaces aggregator instance on equivalent one. Key should be valid for for given aggregator and a statistic branch @@ -361,8 +385,8 @@ def get_tensor_collector_inputs( ) -> Dict[int, List[NNCFTensor]]: """ Static method that converts all model outputs and collected output_info - to a layout required for `register_input` method. This method is not a part of - `register_input` to avoid all inputs passing to `TensorCollector.register_input` method. + to a layout required for `register_inputs` method. This method is not a part of + `register_inputs` to avoid all inputs passing to `TensorCollector.register_inputs` method. :param outputs: Target model outputs. :param output_info: Output info collected by a `TensorCollector.get_output_info` method. @@ -373,6 +397,39 @@ def get_tensor_collector_inputs( target_inputs[reducer] = [outputs[name] for name in names] return target_inputs + @staticmethod + def _build_statistic_container(statistic_container_cls: Type[TensorStatistic], kwargs: Dict[Any, Any]): + if issubclass(statistic_container_cls, MinMaxTensorStatistic): + return statistic_container_cls( + min_values=kwargs[MinMaxTensorStatistic.MIN_STAT], max_values=kwargs[MinMaxTensorStatistic.MAX_STAT] + ) + if issubclass(statistic_container_cls, MeanTensorStatistic): + return statistic_container_cls( + mean_values=kwargs[MeanTensorStatistic.MEAN_STAT], shape=kwargs[MeanTensorStatistic.SHAPE_STAT] + ) + if issubclass(statistic_container_cls, RawTensorStatistic): + return statistic_container_cls(values=kwargs[RawTensorStatistic.VALUES_STATS]) + if issubclass(statistic_container_cls, MedianMADTensorStatistic): + return statistic_container_cls( + median_values=kwargs[MedianMADTensorStatistic.TENSOR_STATISTIC_OUTPUT_KEY][ + MedianMADTensorStatistic.MEDIAN_VALUES_STAT + ], + mad_values=kwargs[MedianMADTensorStatistic.TENSOR_STATISTIC_OUTPUT_KEY][ + MedianMADTensorStatistic.MAD_VALUES_STAT + ], + ) + if issubclass(statistic_container_cls, PercentileTensorStatistic): + if PercentileTensorStatistic.TENSOR_STATISTIC_OUTPUT_KEY in kwargs: + percentile_vs_values_dict = kwargs[PercentileTensorStatistic.TENSOR_STATISTIC_OUTPUT_KEY] + else: + percentile_vs_values_dict = {} + for (_, percentile), value in kwargs.items(): + percentile_vs_values_dict[percentile] = value + return statistic_container_cls(percentile_vs_values_dict=percentile_vs_values_dict) + raise RuntimeError( + f"Statistic collector class {statistic_container_cls} is not supported by the TensorCollector class." + ) + class MergedTensorCollector(TensorCollector): """ @@ -388,7 +445,7 @@ def __init__(self, tensor_collectors: List[TensorCollector]) -> None: :param tensor_collectors: Tensor collectors to merge. """ super().__init__() - aggregators: Dict[Tuple[int, int], List[Tuple[TensorCollector, TensorAggregatorBase]]] = defaultdict(list) + aggregators: Dict[Tuple[int, int, int], List[Tuple[TensorCollector, AggregatorBase]]] = defaultdict(list) for tensor_collector in tensor_collectors: if not tensor_collector.enabled: continue @@ -424,68 +481,69 @@ def _reduce_out_of_place(self, x: List[TensorType]) -> List[TensorType]: class MinReducer(TensorReducerBase): def _reduce_out_of_place(self, x: List[NNCFTensor]) -> List[NNCFTensor]: x = x[0] - reduction_shape = self._get_reduction_shape(x) - return [self._tensor_processor.reduce_min(x, reduction_shape, keepdims=True)] + reduction_axes = self._get_reduction_axes(x) + return [self._tensor_processor.reduce_min(x, reduction_axes, keepdims=self._keepdims)] class MaxReducer(TensorReducerBase): def _reduce_out_of_place(self, x: List[NNCFTensor]) -> List[NNCFTensor]: x = x[0] - reduction_shape = self._get_reduction_shape(x) - return [self._tensor_processor.reduce_max(x, reduction_shape, keepdims=True)] + reduction_axes = self._get_reduction_axes(x) + return [self._tensor_processor.reduce_max(x, reduction_axes, keepdims=self._keepdims)] class AbsMaxReducer(TensorReducerBase): def _reduce_out_of_place(self, x: List[NNCFTensor]) -> List[NNCFTensor]: x = self._tensor_processor.abs(x[0]) - reduction_shape = self._get_reduction_shape(x) - return [self._tensor_processor.reduce_max(x, reduction_shape, keepdims=True)] + reduction_axes = self._get_reduction_axes(x) + return [self._tensor_processor.reduce_max(x, reduction_axes, keepdims=self._keepdims)] class MeanReducer(TensorReducerBase): def _reduce_out_of_place(self, x: List[NNCFTensor]) -> List[NNCFTensor]: x = x[0] - reduction_shape = self._get_reduction_shape(x) - return [self._tensor_processor.mean(x, reduction_shape, keepdims=True)] + reduction_axes = self._get_reduction_axes(x) + return [self._tensor_processor.mean(x, reduction_axes, keepdims=self._keepdims)] class QuantileReducerBase(TensorReducerBase): def __init__( self, - reduction_shape: Optional[ReductionShape] = None, + reduction_axes: Optional[ReductionAxes] = None, quantile: Optional[Union[float, Tuple[float]]] = None, inplace: bool = False, ): - super().__init__(reduction_shape, False) + super().__init__(reduction_axes=reduction_axes, inplace=False) self._quantile = (0.01, 0.99) if quantile is None else quantile def __eq__(self, __o: object) -> bool: return super().__eq__(__o) and self._quantile == __o._quantile def __hash__(self) -> int: - return hash((self.__class__.__name__, self.inplace, self._reduction_shape, tuple(self._quantile))) + return hash((self.__class__.__name__, self.inplace, self._reduction_axes, tuple(self._quantile))) class QuantileReducer(QuantileReducerBase): def _reduce_out_of_place(self, x: List[NNCFTensor]) -> List[NNCFTensor]: x = x[0] - reduction_shape = self._get_reduction_shape(x) - return self._tensor_processor.quantile(x, self._quantile, reduction_shape, keepdims=True) + reduction_axes = self._get_reduction_axes(x) + return self._tensor_processor.quantile(x, self._quantile, reduction_axes, keepdims=self._keepdims) class AbsQuantileReducer(QuantileReducerBase): def __init__( self, - reduction_shape: Optional[ReductionShape] = None, - quantile: Union[float, List[float]] = 0.99, + reduction_axes: Optional[ReductionAxes] = None, + quantile: Optional[Union[float, List[float]]] = None, inplace: bool = False, ): - super().__init__(reduction_shape, quantile, False) + quantile = (0.99,) if quantile is None else quantile + super().__init__(reduction_axes=reduction_axes, quantile=quantile, inplace=False) def _reduce_out_of_place(self, x: List[NNCFTensor]) -> List[NNCFTensor]: x = self._tensor_processor.abs(x[0]) - reduction_shape = self._get_reduction_shape(x) - return self._tensor_processor.quantile(x, [self._quantile], reduction_shape, keepdims=True) + reduction_axes = self._get_reduction_axes(x) + return self._tensor_processor.quantile(x, self._quantile, reduction_axes, keepdims=self._keepdims) class BatchMeanReducer(TensorReducerBase): @@ -497,19 +555,26 @@ def _reduce_out_of_place(self, x: List[NNCFTensor]) -> List[NNCFTensor]: class MeanPerChReducer(TensorReducerBase): - def __init__(self, channel_dim: int = 1, inplace: bool = False): - super().__init__(channel_dim, inplace) + def __init__(self, channel_axis: int = 1, inplace: bool = False): + super().__init__(inplace=inplace) + self._channel_axis = channel_axis def _reduce_out_of_place(self, x: List[NNCFTensor]) -> List[NNCFTensor]: - return [self._tensor_processor.mean_per_channel(x[0], self._reduction_shape)] + return [self._tensor_processor.mean_per_channel(x[0], self._channel_axis)] + + def __eq__(self, __o: object) -> bool: + return super().__eq__(__o) and self._channel_axis == __o._channel_axis + + def __hash__(self) -> int: + return hash((self.__class__.__name__, self.inplace, self._reduction_axes, self._channel_axis)) ##################################################Aggregators################################################## -class NoopAggregator(TensorAggregatorBase): +class NoopAggregator(AggregatorBase): def __init__(self, num_samples: Optional[int]): - super().__init__(None, num_samples) + super().__init__(None, num_samples=num_samples) def _register_reduced_input_impl(self, x: TensorType) -> None: self._container.append(x.tensor) @@ -518,9 +583,9 @@ def _aggregate_impl(self): return self._container -class ShapeAggregator(TensorAggregatorBase): +class ShapeAggregator(AggregatorBase): def __init__(self): - super().__init__(None, 1) + super().__init__(None, num_samples=1) def _register_reduced_input_impl(self, x: TensorType) -> None: self._container = x @@ -529,74 +594,124 @@ def _aggregate_impl(self): return self._container.shape -class MinAggregator(TensorAggregatorBase): - def _register_reduced_input_impl(self, x: TensorType) -> None: - if not self._container: - self._container = x - else: - self._container = self._tensor_processor.min(x, self._container) +class TensorAggregatorBase(AggregatorBase, ABC): + def __init__( + self, + tensor_processor: NNCFCollectorTensorProcessor, + aggregation_axes: Optional[AggregationAxes] = None, + num_samples: Optional[int] = None, + window_size=None, + ): + super().__init__(tensor_processor, aggregation_axes=aggregation_axes, num_samples=num_samples) + self._window_size = window_size + self._container = deque(maxlen=window_size) - def _aggregate_impl(self): - return self._container.tensor +class OnlineAggregatorBase(TensorAggregatorBase, ABC): + """ + Base class for aggregators which are using aggregation function fn with following property: + fn([x1, x2, x3]) == fn([fn([x1, x2]), x3]) where x1, x2, x3 are samples to aggregate. + Online aggregation fn([fn([x1, x2]), x3]) allows to keep memory stamp low as only + one sample is stored during statistic collection. + """ -class MaxAggregator(TensorAggregatorBase): - def _register_reduced_input_impl(self, x: TensorType) -> None: - if not self._container: - self._container = x + def _register_reduced_input_impl(self, x: NNCFTensor) -> None: + online_aggregation_axes = tuple(dim - 1 for dim in self._aggregation_axes if dim != 0) + if online_aggregation_axes: + reduced = self._aggregation_fn(x, axis=online_aggregation_axes, keepdims=self._keepdims) else: - self._container = self._tensor_processor.max(x, self._container) + reduced = x + if 0 in self._aggregation_axes: + if self._container: + reduced = self._aggregation_fn( + self._tensor_processor.stack([reduced, self._container]), axis=0, keepdims=False + ) + self._container = reduced + else: + self._container.append(reduced) + + def _aggregate_impl(self) -> NNCFTensor: + if 0 in self._aggregation_axes: + if self._keepdims: + return self._tensor_processor.stack([self._container]).tensor + return self._container.tensor + return self._tensor_processor.stack(self._container).tensor + + @abstractmethod + def _aggregation_fn(self, stacked_value: NNCFTensor, axis: AggregationAxes, keepdims: bool) -> NNCFTensor: + pass - def _aggregate_impl(self): - return self._container.tensor + +class MinAggregator(OnlineAggregatorBase): + def _aggregation_fn(self, stacked_value: NNCFTensor, axis: AggregationAxes, keepdims: bool) -> NNCFTensor: + return self._tensor_processor.reduce_min(stacked_value, axis=axis, keepdims=keepdims) + + +class MaxAggregator(OnlineAggregatorBase): + def _aggregation_fn(self, stacked_value: NNCFTensor, axis: AggregationAxes, keepdims: bool) -> NNCFTensor: + return self._tensor_processor.reduce_max(stacked_value, axis=axis, keepdims=keepdims) class OfflineAggregatorBase(TensorAggregatorBase, ABC): - def __init__( - self, tensor_processor, use_per_sample_stats: bool = False, num_samples: Optional[int] = None, window_size=None - ): - super().__init__(tensor_processor, num_samples) - self._window_size = window_size - self._container = deque(maxlen=window_size) - self._use_per_sample_stats = use_per_sample_stats + """ + Base class for aggregators which are using aggregation function fn which + does not fulfill property fn([x1, x2, x3]) == fn([fn([x1, x2]), x3]) + where x1, x2, x3 are samples to aggregate. Child aggregators collect + all samples in a container and aggregate them in one step. + """ def _register_reduced_input_impl(self, x: TensorType) -> None: - if self._use_per_sample_stats: - self._container.extend(self._tensor_processor.unstack(x)) - else: - self._container.append(x) + self._container.append(x) - def _offline_aggregation_impl(self, fn): + def _aggregate_impl(self) -> NNCFTensor: stacked_val = self._tensor_processor.stack(self._container) - return fn(stacked_val, axis=0, keepdims=False).tensor + return self._aggregation_fn(stacked_val, axis=self._aggregation_axes, keepdims=self._keepdims).tensor + + @abstractmethod + def _aggregation_fn(self, stacked_value: NNCFTensor, axis: AggregationAxes, keepdims: bool) -> NNCFTensor: + pass class MeanAggregator(OfflineAggregatorBase): - def _aggregate_impl(self): - return self._offline_aggregation_impl(self._tensor_processor.mean) + def _aggregation_fn(self, stacked_value: NNCFTensor, axis: AggregationAxes, keepdims: bool) -> NNCFTensor: + return self._tensor_processor.mean(stacked_value, axis=axis, keepdims=keepdims) class MedianAggregator(OfflineAggregatorBase): - def _aggregate_impl(self): - return self._offline_aggregation_impl(self._tensor_processor.median) + def _aggregation_fn(self, stacked_value: NNCFTensor, axis: AggregationAxes, keepdims: bool) -> NNCFTensor: + return self._tensor_processor.median(stacked_value, axis=axis, keepdims=keepdims) class NoOutliersAggregatorBase(OfflineAggregatorBase, ABC): def __init__( self, - tensor_processor, - use_per_sample_stats: bool = False, + tensor_processor: NNCFCollectorTensorProcessor, + aggregation_axes: Optional[AggregationAxes] = None, num_samples: Optional[int] = None, window_size=None, quantile: float = 0.01, ): - super().__init__(tensor_processor, use_per_sample_stats, num_samples, window_size) + super().__init__(tensor_processor, aggregation_axes=aggregation_axes, num_samples=num_samples) + self._window_size = window_size + self._container = deque(maxlen=window_size) self._quantile = quantile - def _offline_aggregation_impl(self, fn) -> List[NNCFTensor]: - stacked_val = self._tensor_processor.stack(self._container) - result = self._tensor_processor.no_outliers_map(stacked_val, fn, axis=0, alpha=self._quantile) - return result.tensor + def _aggregate_impl(self) -> NNCFTensor: + stacked_samples = self._tensor_processor.stack(self._container) + low_values, high_values = self._tensor_processor.quantile( + stacked_samples, quantile=(self._quantile, 1 - self._quantile), axis=self._aggregation_axes + ) + tp = self._tensor_processor + outliers_mask = tp.logical_or(tp.less(stacked_samples, low_values), tp.less(high_values, stacked_samples)) + return self._aggregation_fn( + stacked_samples=stacked_samples, mask=outliers_mask, axis=self._aggregation_axes, keepdims=self._keepdims + ).tensor + + @abstractmethod + def _aggregation_fn( + self, stacked_samples: NNCFTensor, mask: NNCFTensor, axis: AggregationAxes, keepdims: bool + ) -> NNCFTensor: + pass def __eq__(self, __o: object) -> bool: return super().__eq__(__o) and self._quantile == __o._quantile @@ -606,13 +721,71 @@ def __hash__(self) -> int: class MeanNoOutliersAggregator(NoOutliersAggregatorBase): - def _aggregate_impl(self) -> Any: - return self._offline_aggregation_impl(self._tensor_processor.masked_mean) + def _aggregation_fn( + self, stacked_samples: NNCFTensor, mask: NNCFTensor, axis: AggregationAxes, keepdims: bool + ) -> NNCFTensor: + return self._tensor_processor.masked_mean(stacked_samples, axis=axis, mask=mask, keepdims=keepdims) class MedianNoOutliersAggregator(NoOutliersAggregatorBase): - def _aggregate_impl(self) -> Any: - return self._offline_aggregation_impl(self._tensor_processor.masked_median) + def _aggregation_fn( + self, stacked_samples: NNCFTensor, mask: NNCFTensor, axis: AggregationAxes, keepdims: bool + ) -> NNCFTensor: + return self._tensor_processor.masked_median(stacked_samples, axis=axis, mask=mask, keepdims=keepdims) + + +class MedianAbsoluteDeviationAggregator(TensorAggregatorBase): + def _register_reduced_input_impl(self, x: TensorType) -> None: + return self._container.append(x) + + def _aggregate_impl(self) -> Dict[str, NNCFTensor]: + stacked_val = self._tensor_processor.stack(self._container) + + mask = self._tensor_processor.zero_elements(stacked_val) + median_per_ch = self._tensor_processor.masked_median( + stacked_val, mask=mask, axis=self._aggregation_axes, keepdims=True + ) + + mad_values = self._tensor_processor.median( + self._tensor_processor.abs(self._tensor_processor.sub(stacked_val, median_per_ch)), + axis=self._aggregation_axes, + keepdims=self._keepdims, + ) + if not self._keepdims: + median_per_ch = self._tensor_processor.squeeze(median_per_ch, self._aggregation_axes) + return { + MedianMADTensorStatistic.MEDIAN_VALUES_STAT: median_per_ch.tensor, + MedianMADTensorStatistic.MAD_VALUES_STAT: mad_values.tensor, + } + + +class PercentileAggregator(TensorAggregatorBase): + def __init__( + self, + tensor_processor: NNCFCollectorTensorProcessor, + percentiles_to_collect: List[float], + aggregation_axes: Optional[AggregationAxes] = None, + num_samples: Optional[int] = None, + window_size=None, + ): + super().__init__(tensor_processor, aggregation_axes=aggregation_axes, num_samples=num_samples) + self._percentiles_to_collect = percentiles_to_collect + self._window_size = window_size + self._container = deque(maxlen=window_size) + + def _register_reduced_input_impl(self, x: TensorType) -> None: + return self._container.append(x) + + def _aggregate_impl(self) -> Dict[float, NNCFTensor]: + stacked_val = self._tensor_processor.stack(self._container) + + percentiles = self._tensor_processor.percentile( + stacked_val, self._percentiles_to_collect, axis=self._aggregation_axes, keepdims=self._keepdims + ) + retval = {} + for idx, percentile in enumerate(self._percentiles_to_collect): + retval[percentile] = percentiles[idx].tensor + return retval AGGREGATORS_MAP = { diff --git a/nncf/onnx/graph/node_utils.py b/nncf/onnx/graph/node_utils.py index 5312f2fdded..6575dff6f1c 100644 --- a/nncf/onnx/graph/node_utils.py +++ b/nncf/onnx/graph/node_utils.py @@ -18,7 +18,7 @@ from nncf.common.graph.graph import NNCFNode from nncf.common.graph.transformations.commands import TargetType from nncf.common.logging.logger import nncf_logger -from nncf.common.tensor_statistics.collectors import ReductionShape +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.onnx.graph.metatypes import onnx_metatypes as om from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXDequantizeLinearMetatype from nncf.onnx.graph.onnx_graph import ONNXGraph @@ -131,7 +131,7 @@ def transpose_axis(shape: List[int], axis: int) -> int: return range(len(shape) - 1, -1, -1)[axis] # Iterate backward throug axis -def get_reduction_shape(shape: List[int], axis: int) -> ReductionShape: +def get_reduction_shape(shape: List[int], axis: int) -> ReductionAxes: """ Returns reduction shape for shape and axis. diff --git a/nncf/onnx/statistics/collectors.py b/nncf/onnx/statistics/collectors.py index 4afb2396a4b..2ce98915edc 100644 --- a/nncf/onnx/statistics/collectors.py +++ b/nncf/onnx/statistics/collectors.py @@ -9,7 +9,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from typing import Any, Callable, Deque, List, Optional, Union +from typing import Deque, List, Optional, Tuple, Union import numpy as np @@ -26,17 +26,18 @@ from nncf.onnx.tensor import ONNXNNCFTensor +# pylint: disable=too-many-public-methods class ONNXNNCFCollectorTensorProcessor(NNCFCollectorTensorProcessor): """ A realization of the processing methods for ONNXNNCFTensors. """ @staticmethod - def reduce_min(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def reduce_min(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: return ONNXNNCFTensor(np.amin(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod - def reduce_max(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def reduce_max(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: return ONNXNNCFTensor(np.amax(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod @@ -52,16 +53,20 @@ def max(x1: NNCFTensor, x2: NNCFTensor) -> NNCFTensor: return ONNXNNCFTensor(np.maximum(x1.tensor, x2.tensor)) @staticmethod - def mean(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTensor: + def mean(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims=False) -> NNCFTensor: return ONNXNNCFTensor(np.mean(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod - def median(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTensor: + def median(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims=False) -> NNCFTensor: return ONNXNNCFTensor(np.median(x.tensor, axis=axis, keepdims=keepdims)) @classmethod def masked_mean( - cls, x: NNCFTensor, axis: Optional[Union[int, tuple, list]], mask: Optional[NNCFTensor], keepdims: bool = False + cls, + x: NNCFTensor, + axis: Optional[Union[int, Tuple[int, ...], List[int]]], + mask: Optional[NNCFTensor], + keepdims: bool = False, ) -> NNCFTensor: if mask is None: return cls.mean(x, axis=axis, keepdims=keepdims) @@ -70,32 +75,24 @@ def masked_mean( @classmethod def masked_median( - cls, x: NNCFTensor, axis: Optional[Union[int, tuple, list]], mask: Optional[NNCFTensor], keepdims: bool = False + cls, + x: NNCFTensor, + axis: Optional[Union[int, Tuple[int, ...], List[int]]], + mask: Optional[NNCFTensor], + keepdims: bool = False, ) -> NNCFTensor: if mask is None: return cls.median(x, axis=axis, keepdims=keepdims) masked_x = np.ma.array(x.tensor, mask=mask.tensor) return ONNXNNCFTensor(np.ma.median(masked_x, axis=axis, keepdims=keepdims).data) - @classmethod - def no_outliers_map( - cls, - x: NNCFTensor, - fn: Callable[[NNCFTensor, int, NNCFTensor], Any], - axis: int = 0, - alpha: float = 0.01, - keepdims: bool = False, - ) -> NNCFTensor: - if len(x.shape) == 1: - return fn(x, axis=None, mask=None, keepdims=keepdims) - - x = x.tensor - if axis: - x = np.moveaxis(x, axis, 0) + @staticmethod + def logical_or(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: + return ONNXNNCFTensor(np.logical_or(input_.tensor, other.tensor)) - low_values, high_values = np.quantile(x, [alpha, 1 - alpha], 0) - outliers_mask = np.logical_or(x < low_values, high_values < x) - return fn(ONNXNNCFTensor(x), axis=0, mask=ONNXNNCFTensor(outliers_mask), keepdims=keepdims) + @staticmethod + def less(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: + return ONNXNNCFTensor(input_.tensor < other.tensor) @staticmethod def stack(x: Union[List[NNCFTensor], Deque[NNCFTensor]], axis: int = 0) -> NNCFTensor: @@ -106,17 +103,34 @@ def stack(x: Union[List[NNCFTensor], Deque[NNCFTensor]], axis: int = 0) -> NNCFT def unstack(x: NNCFTensor, axis: int = 0) -> List[NNCFTensor]: return [ONNXNNCFTensor(np.squeeze(e, axis)) for e in np.split(x.tensor, x.tensor.shape[axis], axis=axis)] + @staticmethod + def squeeze(x: NNCFTensor, dim: Optional[Union[int, Tuple[int, ...]]] = None) -> NNCFTensor: + raise NotImplementedError() + @staticmethod def sum(tensor: NNCFTensor) -> TensorElementsType: return np.sum(tensor.tensor) @staticmethod def quantile( - tensor: NNCFTensor, quantile: Union[float, List[float]], axis: Union[int, tuple, list], keepdims: bool = False + tensor: NNCFTensor, + quantile: Union[float, List[float]], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, ) -> List[TensorElementsType]: result = np.quantile(tensor.tensor, quantile, axis, keepdims=keepdims) return [ONNXNNCFTensor(x) for x in result] + @classmethod + def percentile( + cls, + tensor: NNCFTensor, + percentile: Union[float, List[float]], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, + ) -> List[TensorElementsType]: + raise NotImplementedError() + @staticmethod def mean_per_channel(x: NNCFTensor, axis: int) -> NNCFTensor: if len(x.shape) < 3: @@ -129,47 +143,64 @@ def mean_per_channel(x: NNCFTensor, axis: int) -> NNCFTensor: def batch_mean(x: NNCFTensor) -> NNCFTensor: return ONNXNNCFTensor(np.mean(x.tensor, axis=0, keepdims=True)) + @staticmethod + def sub(a: NNCFTensor, b: NNCFTensor) -> NNCFTensor: + raise NotImplementedError() + + @staticmethod + def zero_elements(x: NNCFTensor) -> NNCFTensor: + raise NotImplementedError() + class ONNXMinMaxStatisticCollector(MinMaxStatisticCollector): @staticmethod def _get_processor() -> NNCFCollectorTensorProcessor: - return ONNXNNCFCollectorTensorProcessor() + return ONNXNNCFCollectorTensorProcessor def _register_input(self, x: ONNXNNCFTensor): self._register_input_common(x) def _get_statistics(self) -> ONNXMinMaxTensorStatistic: - return ONNXMinMaxTensorStatistic(self._min_values.tensor, self._max_values.tensor) + return ONNXMinMaxTensorStatistic( + min_values=self._min_values.tensor, + max_values=self._max_values.tensor, + ) class ONNXMeanMinMaxStatisticCollector(MeanMinMaxStatisticCollector): @staticmethod def _get_processor() -> NNCFCollectorTensorProcessor: - return ONNXNNCFCollectorTensorProcessor() + return ONNXNNCFCollectorTensorProcessor def _register_input(self, x: ONNXNNCFTensor): self._register_input_common(x) def _get_statistics(self) -> ONNXMinMaxTensorStatistic: - return ONNXMinMaxTensorStatistic(self._min_aggregate().tensor, self._max_aggregate().tensor) + return ONNXMinMaxTensorStatistic( + min_values=self._min_aggregate().tensor, + max_values=self._max_aggregate().tensor, + ) class ONNXMeanStatisticCollector(MeanStatisticCollector): @staticmethod def _get_processor() -> NNCFCollectorTensorProcessor: - return ONNXNNCFCollectorTensorProcessor() + return ONNXNNCFCollectorTensorProcessor def _register_input(self, x: ONNXNNCFTensor): self._register_input_common(x) def _get_statistics(self) -> ONNXMeanTensorStatistic: - return ONNXMeanTensorStatistic(self._mean_aggregate().tensor, self._shape()) + return ONNXMeanTensorStatistic( + mean_values=self._mean_aggregate().tensor, + shape=self._shape(), + ) class ONNXRawStatisticCollector(RawStatisticCollector): @staticmethod def _get_processor() -> NNCFCollectorTensorProcessor: - return ONNXNNCFCollectorTensorProcessor() + return ONNXNNCFCollectorTensorProcessor def _register_input(self, x: ONNXNNCFTensor): self._register_input_common(x) diff --git a/nncf/openvino/graph/node_utils.py b/nncf/openvino/graph/node_utils.py index 9c9d41137cf..33efa0e1de2 100644 --- a/nncf/openvino/graph/node_utils.py +++ b/nncf/openvino/graph/node_utils.py @@ -17,6 +17,7 @@ from nncf.common.graph.graph import NNCFGraph from nncf.common.graph.graph import NNCFNode +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.openvino.graph.layer_attributes import OVLayerAttributes from nncf.openvino.graph.metatypes.groups import OPERATIONS_WITH_BIAS from nncf.openvino.graph.metatypes.groups import OPERATIONS_WITH_WEIGHTS @@ -161,7 +162,7 @@ def get_ov_model_reduce_node_name(output_name: str, reduce_node_name: str, port_ def get_inplace_reduce_op( - op: Type[ov.Node], reduce_node_name: str, reduction_axes: Optional[Tuple[int, ...]], use_abs: bool + op: Type[ov.Node], reduce_node_name: str, reduction_axes: Optional[ReductionAxes], use_abs: bool ) -> InplaceInsertionFnType: """ Returns inplace insertion function that adds reduce node to a passed node. @@ -169,6 +170,7 @@ def get_inplace_reduce_op( :param op: OpenVINO reduction operation type to insert. :param reduce_node_name: Reduce node name. :param reduction_axes: Target reduction axes for the reduction node. + Reduce along all axes in case reduction_axes are None. :param use_abs: Wheather reduce absolute values of input tensors or not. :returns: Inplace insertion function to use in ModelTransformer. """ @@ -200,37 +202,43 @@ def get_reduce_op(node: ov.Node, output_port_id: int) -> ov.Node: return get_reduce_op -def get_inplace_min_op(node_name: str, reduction_shape: Tuple[int, ...]) -> InplaceInsertionFnType: +def get_inplace_min_op(node_name: str, reduction_axes: Optional[ReductionAxes]) -> InplaceInsertionFnType: """ Returns inplace min function that adds reduce min node to a passed node. :param node_name: Min reduce node name. - :param reduction_shape: Target reduction axes for the reduction node. + :param reduction_axes: Target reduction axes for the reduction node. + Reduce along all axes in case reduction_axes are None. :returns: Inplace insertion function to use in ModelTransformer. """ - return get_inplace_reduce_op(opset.reduce_min, node_name, reduction_shape, False) + return get_inplace_reduce_op(opset.reduce_min, node_name, reduction_axes, False) -def get_inplace_max_op(node_name: str, reduction_shape: Tuple[int, ...], use_abs_max: bool) -> InplaceInsertionFnType: +def get_inplace_max_op( + node_name: str, reduction_axes: Optional[ReductionAxes], use_abs_max: bool +) -> InplaceInsertionFnType: """ Returns inplace max function that adds reduce max node to a passed node. :param node_name: Max reduce node name. - :param reduction_shape: Target reduction axes for the reduction node. + :param reduction_axes: Target reduction axes for the reduction node. + Reduce along all axes in case reduction_axes are None. :param use_abs: Wheather reduce absolute values of input tensors or not. :returns: Inplace insertion function to use in ModelTransformer. """ - return get_inplace_reduce_op(opset.reduce_max, node_name, reduction_shape, use_abs_max) + return get_inplace_reduce_op(opset.reduce_max, node_name, reduction_axes, use_abs_max) -def get_inplace_mean_op(node_name: str, reduction_shape: Tuple[int, ...]) -> InplaceInsertionFnType: +def get_inplace_mean_op(node_name: str, reduction_axes: Optional[ReductionAxes]) -> InplaceInsertionFnType: """ Returns inplace mean function that adds reduce mean node to a passed node. :param node_name: Mean reduce node name. + :param reduction_axes: Target reduction axes for the reduction node. + Reduce along all axes in case reduction_axes are None. :returns: Inplace insertion function to use in ModelTransformer. """ - return get_inplace_reduce_op(opset.reduce_mean, node_name, reduction_shape, False) + return get_inplace_reduce_op(opset.reduce_mean, node_name, reduction_axes, False) def get_inplace_batch_mean_op(node_name: str) -> InplaceInsertionFnType: @@ -302,7 +310,7 @@ def get_reduce_op(node: ov.Node, output_port_id: int) -> ov.Node: return get_reduce_op -def get_partial_shape_safe(node, port_id) -> int: +def get_partial_shape_safe(node, port_id) -> Tuple[int, ...]: partial_shape = node.get_output_partial_shape(port_id) if partial_shape.rank.is_dynamic or not partial_shape.all_non_negative: raise RuntimeError( @@ -373,18 +381,18 @@ def get_matmul_channel_axes(weights_port_id: int, ndims: int, transpose: bool) - return channel_axes -def get_channel_agnostic_reduction_shape(channel_axes: List[int], shape: List[int]) -> Tuple[int]: +def get_channel_agnostic_reduction_axes(channel_axes: List[int], shape: List[int]) -> Optional[ReductionAxes]: """ - Returns filtered reduction shape without axes that corresponds channels. + Returns filtered reduction axes without axes that corresponds channels. :param channel_axes: List of the channel axes. :param shape: Shape that need to be filtered. - :return: Reduction shape in tuple format. + :return: Reduction axes in tuple format. """ - reduction_shape = list(range(len(shape))) + reduction_axes = list(range(len(shape))) for channel_axis in sorted(channel_axes, reverse=True): - del reduction_shape[channel_axis] - return tuple(reduction_shape) + del reduction_axes[channel_axis] + return tuple(reduction_axes) def create_bias_tensor(node_without_bias: NNCFNode, graph: NNCFGraph, value: Any) -> np.ndarray: diff --git a/nncf/openvino/statistics/collectors.py b/nncf/openvino/statistics/collectors.py index 61ef776fab7..4672541d86b 100644 --- a/nncf/openvino/statistics/collectors.py +++ b/nncf/openvino/statistics/collectors.py @@ -9,7 +9,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from typing import Any, Callable, Deque, List, Optional, Union +from typing import Deque, List, Optional, Tuple, Union import numpy as np @@ -43,17 +43,18 @@ from nncf.quantization.advanced_parameters import StatisticsType +# pylint: disable=too-many-public-methods class OVNNCFCollectorTensorProcessor(NNCFCollectorTensorProcessor): """ A realization of the processing methods for OVNNCFTensors. """ @staticmethod - def reduce_min(x: NNCFTensor, axis: Union[int, tuple], keepdims: bool = True) -> NNCFTensor: + def reduce_min(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = True) -> NNCFTensor: return OVNNCFTensor(np.amin(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod - def reduce_max(x: NNCFTensor, axis: Union[int, tuple], keepdims: bool = True) -> NNCFTensor: + def reduce_max(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = True) -> NNCFTensor: return OVNNCFTensor(np.amax(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod @@ -69,30 +70,44 @@ def max(x1: NNCFTensor, x2: NNCFTensor) -> NNCFTensor: return OVNNCFTensor(np.maximum(x1.tensor, x2.tensor)) @staticmethod - def mean(x: NNCFTensor, axis: Union[int, tuple], keepdims: bool = False) -> NNCFTensor: + def mean(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: return OVNNCFTensor(np.mean(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod - def median(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def median(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: return OVNNCFTensor(np.median(x.tensor, axis=axis, keepdims=keepdims)) @classmethod def masked_mean( - cls, x: NNCFTensor, axis: Optional[Union[int, tuple, list]], mask: Optional[NNCFTensor], keepdims: bool = False + cls, + x: NNCFTensor, + axis: Optional[Union[int, Tuple[int, ...], List[int]]], + mask: Optional[NNCFTensor], + keepdims: bool = False, ) -> NNCFTensor: if mask is None: return cls.mean(x, axis=axis, keepdims=keepdims) masked_x = np.ma.array(x.tensor, mask=mask.tensor) - return OVNNCFTensor(np.ma.mean(masked_x, axis=axis, keepdims=False).data) + result = np.ma.mean(masked_x, axis=axis, keepdims=keepdims) + if isinstance(result, np.ma.MaskedArray): + return OVNNCFTensor(result.data) + return OVNNCFTensor(result) @classmethod def masked_median( - cls, x: NNCFTensor, axis: Optional[Union[int, tuple, list]], mask: Optional[NNCFTensor], keepdims: bool = False + cls, + x: NNCFTensor, + axis: Optional[Union[int, Tuple[int, ...], List[int]]], + mask: Optional[NNCFTensor], + keepdims: bool = False, ) -> NNCFTensor: if mask is None: return cls.median(x, axis=axis, keepdims=keepdims) masked_x = np.ma.array(x.tensor, mask=mask.tensor) - return OVNNCFTensor(np.ma.median(masked_x, axis=axis, keepdims=keepdims).data) + result = np.ma.median(masked_x, axis=axis, keepdims=keepdims) + if isinstance(result, np.ma.MaskedArray): + return OVNNCFTensor(result.data) + return OVNNCFTensor(result) @staticmethod def mean_per_channel(x: NNCFTensor, axis: int) -> NNCFTensor: @@ -102,30 +117,18 @@ def mean_per_channel(x: NNCFTensor, axis: int) -> NNCFTensor: t = x.reshape(x.shape[0], x.shape[1], -1) return OVNNCFTensor(np.mean(t, axis=(0, 2))) - @classmethod - def no_outliers_map( - cls, - x: NNCFTensor, - fn: Callable[[NNCFTensor, int, NNCFTensor], Any], - axis: int = 0, - alpha: float = 0.01, - keepdims: bool = False, - ) -> NNCFTensor: - if len(x.shape) == 1: - return fn(x, axis=None, mask=None, keepdims=keepdims) - - x = x.tensor - if axis: - x = np.moveaxis(x, axis, 0) - - low_values, high_values = np.quantile(x, [alpha, 1 - alpha], 0) - outliers_mask = np.logical_or(x < low_values, high_values < x) - return fn(OVNNCFTensor(x), axis=0, mask=OVNNCFTensor(outliers_mask), keepdims=keepdims) - @staticmethod def batch_mean(x: NNCFTensor) -> NNCFTensor: return OVNNCFTensor(np.mean(x.tensor, axis=0, keepdims=True)) + @staticmethod + def logical_or(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: + return OVNNCFTensor(np.logical_or(input_.tensor, other.tensor)) + + @staticmethod + def less(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: + return OVNNCFTensor(input_.tensor < other.tensor) + @staticmethod def stack(x: Union[List[NNCFTensor], Deque[NNCFTensor]], axis: int = 0) -> NNCFTensor: x = [t.tensor for t in x] @@ -135,17 +138,45 @@ def stack(x: Union[List[NNCFTensor], Deque[NNCFTensor]], axis: int = 0) -> NNCFT def unstack(x: NNCFTensor, axis: int = 0) -> List[NNCFTensor]: return [OVNNCFTensor(np.squeeze(e, axis)) for e in np.split(x.tensor, x.tensor.shape[axis], axis=axis)] + @staticmethod + def squeeze(x: NNCFTensor, dim: Optional[Union[int, Tuple[int, ...]]] = None) -> NNCFTensor: + return OVNNCFTensor(np.squeeze(x.tensor, axis=dim)) + @staticmethod def sum(tensor: NNCFTensor) -> TensorElementsType: return np.sum(tensor.tensor) @staticmethod def quantile( - tensor: NNCFTensor, quantile: Union[float, List[float]], axis: Union[int, tuple, list], keepdims: bool = False + tensor: NNCFTensor, + quantile: Union[float, List[float]], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, ) -> List[NNCFTensor]: result = np.quantile(tensor.tensor, quantile, axis, keepdims=keepdims) return [OVNNCFTensor(x) for x in result] + @classmethod + def percentile( + cls, + tensor: NNCFTensor, + percentile: Union[float, List[float]], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, + ) -> List[TensorElementsType]: + quantile = np.true_divide(percentile, 100) + return cls.quantile(tensor, quantile=quantile, axis=axis, keepdims=keepdims) + + @staticmethod + def sub(a: NNCFTensor, b: NNCFTensor) -> NNCFTensor: + return NNCFTensor(a.tensor - b.tensor) + + @staticmethod + def zero_elements(x: NNCFTensor) -> NNCFTensor: + np_tensor = x.tensor + eps = np.finfo(np_tensor.dtype).eps + return NNCFTensor(np.abs(np_tensor) < eps) + class OVNoopReducer(NoopReducer): def get_output_names(self, target_node_name: str, port_id: int) -> List[str]: @@ -157,7 +188,7 @@ def _get_processor(self): return OVNNCFCollectorTensorProcessor def get_inplace_fn(self): - return get_inplace_min_op(self.name, self._reduction_shape) + return get_inplace_min_op(self.name, self._reduction_axes) def get_output_names(self, target_node_name: str, port_id: int) -> List[str]: return get_reducer_output_node_names(self.name, target_node_name, port_id, self.output_port_id, self.inplace) @@ -168,7 +199,7 @@ def _get_processor(self): return OVNNCFCollectorTensorProcessor def get_inplace_fn(self): - return get_inplace_max_op(self.name, self._reduction_shape, False) + return get_inplace_max_op(self.name, self._reduction_axes, False) def get_output_names(self, target_node_name: str, port_id: int) -> List[str]: return get_reducer_output_node_names(self.name, target_node_name, port_id, self.output_port_id, self.inplace) @@ -179,7 +210,7 @@ def _get_processor(self): return OVNNCFCollectorTensorProcessor def get_inplace_fn(self): - return get_inplace_max_op(self.name, self._reduction_shape, True) + return get_inplace_max_op(self.name, self._reduction_axes, True) def get_output_names(self, target_node_name: str, port_id: int) -> List[str]: return get_reducer_output_node_names(self.name, target_node_name, port_id, self.output_port_id, self.inplace) @@ -190,7 +221,7 @@ def _get_processor(self): return OVNNCFCollectorTensorProcessor def get_inplace_fn(self): - return get_inplace_mean_op(self.name, self._reduction_shape) + return get_inplace_mean_op(self.name, self._reduction_axes) def get_output_names(self, target_node_name: str, port_id: int) -> List[str]: return get_reducer_output_node_names(self.name, target_node_name, port_id, self.output_port_id, self.inplace) @@ -212,7 +243,7 @@ def _get_processor(self): return OVNNCFCollectorTensorProcessor def get_inplace_fn(self): - return get_inplace_mean_per_ch(self.name, self._reduction_shape) + return get_inplace_mean_per_ch(self.name, self._reduction_axes) def get_output_names(self, target_node_name: str, port_id: int) -> List[str]: return get_reducer_output_node_names(self.name, target_node_name, port_id, self.output_port_id, self.inplace) @@ -240,19 +271,30 @@ def get_output_names(self, target_node_name: str, port_id: int) -> List[str]: return get_reducer_output_node_names(self.name, target_node_name, port_id, self.output_port_id, self.inplace) -def get_mean_stat_collector(num_samples, channel_axis, window_size=None, inplace=True): +def get_mean_statistic_collector( + num_samples: int, channel_axis: int, window_size: Optional[int] = None, inplace: bool = True +) -> TensorCollector: + """ + Mean statistic collector builder. + + :param num_samples: Maximum number of samples to collect. + :param channel_axis: Channel axis to use during reduction phase. + :param window_size: Number of samples from the end of the list of collected samples to aggregate. + Aggregates all available collected statistics in case parameter is None. + :param inplace: Whether the mean reducer should be calculated inplace or out of place. + :return: Mean statistic collector. + """ # TODO(dlyakhov): use inplace OVBatchMeanReducer and OVMeanPerChanelReducer # after migration on openvino-dev=2023.0 inplace = False if channel_axis == 0: reducer = OVBatchMeanReducer(inplace) else: - reducer = OVMeanPerChanelReducer(channel_axis, inplace) + reducer = OVMeanPerChanelReducer(channel_axis=channel_axis, inplace=inplace) noop_reducer = OVNoopReducer() kwargs = { "tensor_processor": OVNNCFCollectorTensorProcessor, - "use_per_sample_stats": False, "num_samples": num_samples, "window_size": window_size, } diff --git a/nncf/quantization/algorithms/bias_correction/algorithm.py b/nncf/quantization/algorithms/bias_correction/algorithm.py index c3e1e10a6f6..67a5d5855fd 100644 --- a/nncf/quantization/algorithms/bias_correction/algorithm.py +++ b/nncf/quantization/algorithms/bias_correction/algorithm.py @@ -509,7 +509,7 @@ def get_statistic_points(self, model: TModel, graph: NNCFGraph) -> StatisticPoin TargetType.POST_LAYER_OPERATION, node_name, port_id=OUTPUT_PORT_OF_NODE ) stat_collector = self._backend_entity.mean_statistic_collector( - reduction_shape=channel_axis, num_samples=self.subset_size, inplace=self.inplace_statistics + channel_axis=channel_axis, num_samples=self.subset_size, inplace=self.inplace_statistics ) statistic_container.add_statistic_point( StatisticPoint( diff --git a/nncf/quantization/algorithms/bias_correction/backend.py b/nncf/quantization/algorithms/bias_correction/backend.py index a85f2fffb0a..a3f626ffe25 100644 --- a/nncf/quantization/algorithms/bias_correction/backend.py +++ b/nncf/quantization/algorithms/bias_correction/backend.py @@ -21,7 +21,6 @@ from nncf.common.graph.transformations.commands import TargetType from nncf.common.graph.transformations.commands import TransformationCommand from nncf.common.tensor import NNCFTensor -from nncf.common.tensor_statistics.collectors import ReductionShape from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase from nncf.common.utils.registry import Registry @@ -87,7 +86,7 @@ def output_insertion_command(nncf_graph: NNCFGraph, target_point: TargetPoint) - @staticmethod @abstractmethod def mean_statistic_collector( - reduction_shape: ReductionShape, + channel_axis: int, inplace: bool, num_samples: Optional[int] = None, window_size: Optional[int] = None, @@ -95,7 +94,7 @@ def mean_statistic_collector( """ Returns backend-specific mean statistic collector. - :param reduction_shape: Channel axis for the statistics aggregation. + :param channel_axis: Channel axis for the statistics aggregation. :param inplace: Whether to calculate statistic inplace or not. :param num_samples: Maximum number of samples to collect. :param window_size: The maximum size of the samples queue. diff --git a/nncf/quantization/algorithms/bias_correction/onnx_backend.py b/nncf/quantization/algorithms/bias_correction/onnx_backend.py index 0e9ad720a10..364c93acc5a 100644 --- a/nncf/quantization/algorithms/bias_correction/onnx_backend.py +++ b/nncf/quantization/algorithms/bias_correction/onnx_backend.py @@ -17,7 +17,6 @@ from nncf.common.graph import NNCFGraph from nncf.common.graph import NNCFNode from nncf.common.graph.transformations.commands import TargetType -from nncf.common.tensor_statistics.collectors import ReductionShape from nncf.common.utils.backend import BackendType from nncf.onnx.graph.model_utils import remove_fq_from_inputs from nncf.onnx.graph.node_utils import get_bias_value @@ -43,7 +42,7 @@ class ONNXBiasCorrectionAlgoBackend(BiasCorrectionAlgoBackend): @property def tensor_processor(self) -> ONNXNNCFCollectorTensorProcessor: - return ONNXNNCFCollectorTensorProcessor() + return ONNXNNCFCollectorTensorProcessor @property def types_to_insert_bias(self): @@ -77,12 +76,12 @@ def output_insertion_command(nncf_graph: NNCFGraph, target_point: ONNXTargetPoin @staticmethod def mean_statistic_collector( - reduction_shape: ReductionShape, + channel_axis: int, inplace: bool, num_samples: Optional[int] = None, window_size: Optional[int] = None, ) -> ONNXMeanStatisticCollector: - return ONNXMeanStatisticCollector(reduction_shape, num_samples, window_size) + return ONNXMeanStatisticCollector(channel_axis, num_samples, window_size) @staticmethod def raw_statistic_collector(inplace: bool, num_samples: int = None) -> ONNXMeanStatisticCollector: diff --git a/nncf/quantization/algorithms/bias_correction/openvino_backend.py b/nncf/quantization/algorithms/bias_correction/openvino_backend.py index c57f9df6a20..7af72dec173 100644 --- a/nncf/quantization/algorithms/bias_correction/openvino_backend.py +++ b/nncf/quantization/algorithms/bias_correction/openvino_backend.py @@ -17,7 +17,6 @@ from nncf.common.graph import NNCFGraph from nncf.common.graph import NNCFNode from nncf.common.graph.transformations.commands import TargetType -from nncf.common.tensor_statistics.collectors import ReductionShape from nncf.common.utils.backend import BackendType from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.openvino.graph.metatypes.groups import FAKE_QUANTIZE_OPERATIONS @@ -31,7 +30,7 @@ from nncf.openvino.graph.transformations.commands import OVOutputInsertionCommand from nncf.openvino.graph.transformations.commands import OVTargetPoint from nncf.openvino.statistics.collectors import OVNNCFCollectorTensorProcessor -from nncf.openvino.statistics.collectors import get_mean_stat_collector +from nncf.openvino.statistics.collectors import get_mean_statistic_collector from nncf.openvino.statistics.collectors import get_raw_stat_collector from nncf.openvino.tensor import OVNNCFTensor from nncf.quantization.algorithms.bias_correction.backend import ALGO_BACKENDS @@ -65,12 +64,12 @@ def output_insertion_command(nncf_graph: NNCFGraph, target_point: OVTargetPoint) @staticmethod def mean_statistic_collector( - reduction_shape: ReductionShape, + channel_axis: int, inplace: bool, num_samples: Optional[int] = None, window_size: Optional[int] = None, ) -> TensorCollector: - return get_mean_stat_collector(num_samples, reduction_shape, window_size, inplace) + return get_mean_statistic_collector(num_samples, channel_axis, window_size, inplace) @staticmethod def raw_statistic_collector(inplace: bool, num_samples: int = None) -> TensorCollector: diff --git a/nncf/quantization/algorithms/channel_alignment/algorithm.py b/nncf/quantization/algorithms/channel_alignment/algorithm.py index c5cf65dfefd..0170606c64e 100644 --- a/nncf/quantization/algorithms/channel_alignment/algorithm.py +++ b/nncf/quantization/algorithms/channel_alignment/algorithm.py @@ -374,11 +374,11 @@ def get_statistic_points(self, model: TModel, graph: NNCFGraph) -> StatisticPoin for conv_in, add_in, _ in self._get_node_pairs(graph): target_point, node_in = self._get_target_point_and_node_in(conv_in, add_in) channel_axis = conv_in.metatype.output_channel_axis - reduction_shape = list(range(len(graph.get_output_edges(node_in)[0].tensor_shape))) - reduction_shape.remove(channel_axis) + reduction_axes = list(range(len(graph.get_output_edges(node_in)[0].tensor_shape))) + reduction_axes.remove(channel_axis) statistic_collector = self._backend_entity.get_statistic_collector( - tuple(reduction_shape), self._quantile, self.subset_size, self.inplace_statistics + tuple(reduction_axes), self._quantile, self.subset_size, self.inplace_statistics ) statistic_container.add_statistic_point( StatisticPoint( diff --git a/nncf/quantization/algorithms/channel_alignment/backend.py b/nncf/quantization/algorithms/channel_alignment/backend.py index a02b788cf98..f78f2db08a6 100644 --- a/nncf/quantization/algorithms/channel_alignment/backend.py +++ b/nncf/quantization/algorithms/channel_alignment/backend.py @@ -98,12 +98,12 @@ def get_weights_port_ids_for_node(node: NNCFNode) -> Tuple[int, int]: @staticmethod @abstractmethod def get_statistic_collector( - reduction_shape, q: float, num_samples: int, inplace: bool + reduction_axes, q: float, num_samples: int, inplace: bool ) -> TensorStatisticCollectorBase: """ Get backend-specific tensor collector that collects medians of minimal and maximal quantiles. - :param reduction_shape: Target reduction shape for the reduction. + :param reduction_axes: Target reduction axes for the reduction. :param q: Minimal quantile for the tensor collector. :param num_samples: Num samples to collect by the tensor collector. :param inplace: Should statistic be calculated inplace or out of place. diff --git a/nncf/quantization/algorithms/channel_alignment/openvino_backend.py b/nncf/quantization/algorithms/channel_alignment/openvino_backend.py index 0db705b676b..1a3667be9c5 100644 --- a/nncf/quantization/algorithms/channel_alignment/openvino_backend.py +++ b/nncf/quantization/algorithms/channel_alignment/openvino_backend.py @@ -78,10 +78,10 @@ def get_add_metatypes(): @staticmethod def get_statistic_collector( - reduction_shape, q: float, num_samples: int, inplace: bool + reduction_axes, q: float, num_samples: int, inplace: bool ) -> TensorStatisticCollectorBase: tensor_collector = TensorCollector(OVMinMaxTensorStatistic) - quantile_reducer = OVQuantileReducer(reduction_shape, (q, 1 - q), inplace) + quantile_reducer = OVQuantileReducer(reduction_axes, (q, 1 - q), inplace) for port_id, container_key in enumerate([OVMinMaxTensorStatistic.MIN_STAT, OVMinMaxTensorStatistic.MAX_STAT]): aggregator = MedianAggregator(OVNNCFCollectorTensorProcessor, num_samples=num_samples) diff --git a/nncf/quantization/algorithms/fast_bias_correction/algorithm.py b/nncf/quantization/algorithms/fast_bias_correction/algorithm.py index c83fe3b8e26..7ed9182d004 100644 --- a/nncf/quantization/algorithms/fast_bias_correction/algorithm.py +++ b/nncf/quantization/algorithms/fast_bias_correction/algorithm.py @@ -271,7 +271,7 @@ def _add_statistic_point(self, container: StatisticPointsContainer, point: Targe :param axis: Channel axis for the statistics calculation. """ stat_collector = self._backend_entity.mean_statistic_collector( - reduction_shape=axis, num_samples=self.subset_size, inplace=self.inplace_statistics + channel_axis=axis, num_samples=self.subset_size, inplace=self.inplace_statistics ) container.add_statistic_point( StatisticPoint(target_point=point, tensor_collector=stat_collector, algorithm=self._algorithm_key) diff --git a/nncf/quantization/algorithms/fast_bias_correction/backend.py b/nncf/quantization/algorithms/fast_bias_correction/backend.py index ca25adc2fb2..38618fe2efa 100644 --- a/nncf/quantization/algorithms/fast_bias_correction/backend.py +++ b/nncf/quantization/algorithms/fast_bias_correction/backend.py @@ -21,7 +21,6 @@ from nncf.common.graph.transformations.commands import TargetType from nncf.common.graph.transformations.commands import TransformationCommand from nncf.common.tensor import NNCFTensor -from nncf.common.tensor_statistics.collectors import ReductionShape from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase from nncf.common.utils.registry import Registry @@ -79,7 +78,7 @@ def model_extraction_command(inputs: List[str], outputs: List[str]) -> Transform @staticmethod @abstractmethod def mean_statistic_collector( - reduction_shape: ReductionShape, + channel_axis: int, inplace: bool, num_samples: Optional[int] = None, window_size: Optional[int] = None, @@ -87,7 +86,7 @@ def mean_statistic_collector( """ Returns backend-specific mean statistic collector. - :param reduction_shape: Channel axes for the statistics aggregation. + :param channel_axis: Channel axes for the statistics aggregation. :param inplace: Whether to calculate statistic inplace or not. :param num_samples: Maximum number of samples to collect. :param window_size: The maximum size of the samples queue. diff --git a/nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py b/nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py index 733018e9bd2..02bb54106c6 100644 --- a/nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py +++ b/nncf/quantization/algorithms/fast_bias_correction/onnx_backend.py @@ -17,7 +17,6 @@ from nncf.common.graph import NNCFGraph from nncf.common.graph import NNCFNode from nncf.common.graph.transformations.commands import TargetType -from nncf.common.tensor_statistics.collectors import ReductionShape from nncf.common.utils.backend import BackendType from nncf.onnx.graph.node_utils import get_bias_value from nncf.onnx.graph.node_utils import is_any_weight_quantized @@ -42,7 +41,7 @@ def types_to_insert_bias(self): @property def tensor_processor(self) -> ONNXNNCFCollectorTensorProcessor: - return ONNXNNCFCollectorTensorProcessor() + return ONNXNNCFCollectorTensorProcessor @staticmethod def target_point(target_type: TargetType, target_node_name: str, port_id: int) -> ONNXTargetPoint: @@ -64,12 +63,12 @@ def model_extraction_command(inputs: List[str], outputs: List[str]) -> ONNXModel @staticmethod def mean_statistic_collector( - reduction_shape: ReductionShape, + channel_axis: int, inplace: bool, num_samples: Optional[int] = None, window_size: Optional[int] = None, ) -> ONNXMeanStatisticCollector: - return ONNXMeanStatisticCollector(reduction_shape, num_samples, window_size) + return ONNXMeanStatisticCollector(channel_axis, num_samples, window_size) @staticmethod def get_sub_input_output_names(subgraph: onnx.ModelProto) -> Tuple[str, str]: diff --git a/nncf/quantization/algorithms/fast_bias_correction/openvino_backend.py b/nncf/quantization/algorithms/fast_bias_correction/openvino_backend.py index 9436a7544f1..61ebb5a695b 100644 --- a/nncf/quantization/algorithms/fast_bias_correction/openvino_backend.py +++ b/nncf/quantization/algorithms/fast_bias_correction/openvino_backend.py @@ -17,7 +17,6 @@ from nncf.common.graph import NNCFGraph from nncf.common.graph import NNCFNode from nncf.common.graph.transformations.commands import TargetType -from nncf.common.tensor_statistics.collectors import ReductionShape from nncf.common.utils.backend import BackendType from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.openvino.graph.metatypes.groups import FAKE_QUANTIZE_OPERATIONS @@ -28,7 +27,7 @@ from nncf.openvino.graph.transformations.commands import OVModelExtractionCommand from nncf.openvino.graph.transformations.commands import OVTargetPoint from nncf.openvino.statistics.collectors import OVNNCFCollectorTensorProcessor -from nncf.openvino.statistics.collectors import get_mean_stat_collector +from nncf.openvino.statistics.collectors import get_mean_statistic_collector from nncf.openvino.tensor import OVNNCFTensor from nncf.quantization.algorithms.fast_bias_correction.backend import ALGO_BACKENDS from nncf.quantization.algorithms.fast_bias_correction.backend import FastBiasCorrectionAlgoBackend @@ -56,12 +55,12 @@ def model_extraction_command(inputs: List[str], outputs: List[str]) -> OVModelEx @staticmethod def mean_statistic_collector( - reduction_shape: ReductionShape, + channel_axis: int, inplace: bool, num_samples: Optional[int] = None, window_size: Optional[int] = None, ) -> TensorCollector: - return get_mean_stat_collector(num_samples, reduction_shape, window_size, inplace) + return get_mean_statistic_collector(num_samples, channel_axis, window_size, inplace) @staticmethod def get_sub_input_output_names(subgraph: ov.Model) -> Tuple[str, str]: diff --git a/nncf/quantization/algorithms/fast_bias_correction/torch_backend.py b/nncf/quantization/algorithms/fast_bias_correction/torch_backend.py index b7316724db0..cb9b0026e3f 100644 --- a/nncf/quantization/algorithms/fast_bias_correction/torch_backend.py +++ b/nncf/quantization/algorithms/fast_bias_correction/torch_backend.py @@ -18,8 +18,8 @@ from nncf.common.graph import NNCFNode from nncf.common.graph.definitions import NNCFGraphNodeType from nncf.common.graph.transformations.commands import TargetType -from nncf.common.tensor_statistics.collectors import ReductionShape from nncf.common.utils.backend import BackendType +from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.quantization.algorithms.fast_bias_correction.backend import ALGO_BACKENDS from nncf.quantization.algorithms.fast_bias_correction.backend import FastBiasCorrectionAlgoBackend from nncf.torch.graph.transformations.command_creation import create_bias_correction_command @@ -32,8 +32,8 @@ from nncf.torch.model_analyzer import is_quantized_weights from nncf.torch.nncf_network import NNCFNetwork from nncf.torch.tensor import PTNNCFTensor -from nncf.torch.tensor_statistics.collectors import PTMeanStatisticCollector from nncf.torch.tensor_statistics.collectors import PTNNCFCollectorTensorProcessor +from nncf.torch.tensor_statistics.collectors import get_mean_statistic_collector @ALGO_BACKENDS.register(BackendType.TORCH) @@ -45,7 +45,7 @@ class PTFastBiasCorrectionAlgoBackend(FastBiasCorrectionAlgoBackend): @property def tensor_processor(self) -> PTNNCFCollectorTensorProcessor: - return PTNNCFCollectorTensorProcessor() + return PTNNCFCollectorTensorProcessor @staticmethod def target_point(target_type: TargetType, target_node_name: str, port_id: int) -> PTTargetPoint: @@ -67,12 +67,12 @@ def model_extraction_command(inputs: List[str], outputs: List[str]) -> PTModelEx @staticmethod def mean_statistic_collector( - reduction_shape: ReductionShape, + channel_axis: int, inplace: bool, num_samples: Optional[int] = None, window_size: Optional[int] = None, - ) -> PTMeanStatisticCollector: - return PTMeanStatisticCollector(reduction_shape, num_samples, window_size) + ) -> TensorCollector: + return get_mean_statistic_collector(num_samples, channel_axis, window_size) @staticmethod def get_sub_input_output_names(subgraph: NNCFNetwork) -> Tuple[str, str]: diff --git a/nncf/quantization/algorithms/min_max/openvino_backend.py b/nncf/quantization/algorithms/min_max/openvino_backend.py index 4394c1870fa..4ad4e309dc8 100644 --- a/nncf/quantization/algorithms/min_max/openvino_backend.py +++ b/nncf/quantization/algorithms/min_max/openvino_backend.py @@ -20,14 +20,14 @@ from nncf.common.hardware.config import HWConfig from nncf.common.quantization.structs import QuantizationMode from nncf.common.quantization.structs import QuantizerConfig -from nncf.common.tensor_statistics.collectors import ReductionShape +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.common.utils.backend import BackendType from nncf.experimental.common.tensor_statistics.collectors import AGGREGATORS_MAP from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.openvino.graph.layer_attributes import OVLayerAttributes from nncf.openvino.graph.metatypes import openvino_metatypes as om from nncf.openvino.graph.metatypes.groups import OPERATIONS_WITH_WEIGHTS -from nncf.openvino.graph.node_utils import get_channel_agnostic_reduction_shape +from nncf.openvino.graph.node_utils import get_channel_agnostic_reduction_axes from nncf.openvino.graph.node_utils import get_weight_channel_axes from nncf.openvino.graph.transformations.commands import OVQuantizerInsertionCommand from nncf.openvino.graph.transformations.commands import OVTargetPoint @@ -122,9 +122,9 @@ def unify_statistics(statistics: List[OVMinMaxTensorStatistic]) -> OVMinMaxTenso return OVMinMaxTensorStatistic(min_values=min_values, max_values=max_values) @staticmethod - def _get_reduction_shape_and_use_abs_max( + def _get_reduction_axes_and_use_abs_max( nncf_graph: NNCFGraph, target_point: OVTargetPoint, quantizer_config: QuantizerConfig - ) -> Tuple[ReductionShape, bool]: + ) -> Tuple[ReductionAxes, bool]: use_abs_max = quantizer_config.mode == QuantizationMode.SYMMETRIC if not quantizer_config.per_channel: return None, use_abs_max @@ -140,7 +140,7 @@ def _get_reduction_shape_and_use_abs_max( # TODO (l-bat): Disable quantizer propogation through layout changing operations channel_axis = 1 # OpenVINO activations have channel first layout: [N, C, Z, Y, X] - axes = get_channel_agnostic_reduction_shape([channel_axis], shape) + axes = get_channel_agnostic_reduction_axes([channel_axis], shape) return axes, use_abs_max assert isinstance(node.layer_attributes, OVLayerAttributes) @@ -148,7 +148,7 @@ def _get_reduction_shape_and_use_abs_max( if quantizer_config.per_channel: channel_axes = get_weight_channel_axes(node, target_point.port_id) - axes = get_channel_agnostic_reduction_shape(channel_axes, const_shape) + axes = get_channel_agnostic_reduction_axes(channel_axes, const_shape) else: axes = tuple(range(len(const_shape))) return axes, use_abs_max @@ -162,7 +162,7 @@ def get_statistic_collector( inplace: bool, num_samples: int = None, ) -> TensorCollector: - reduction_shape, use_abs_max = OVMinMaxAlgoBackend._get_reduction_shape_and_use_abs_max( + reduction_axes, use_abs_max = OVMinMaxAlgoBackend._get_reduction_axes_and_use_abs_max( nncf_graph, target_point, quantizer_config ) @@ -181,7 +181,7 @@ def get_statistic_collector( f"Aggregator type: {params.aggregator_type} is not supported for OpenVino PTQ backend yet." ) - kwargs = {"reduction_shape": reduction_shape, "inplace": inplace} + kwargs = {"reduction_axes": reduction_axes, "inplace": inplace} if params.statistics_type in [StatisticsType.QUANTILE, StatisticsType.ABS_QUANTILE]: if container_key == OVMinMaxTensorStatistic.MIN_STAT: quantile = params.quantile_outlier_prob diff --git a/nncf/quantization/algorithms/min_max/torch_backend.py b/nncf/quantization/algorithms/min_max/torch_backend.py index e0aa5970b0c..0a8fe5778c5 100644 --- a/nncf/quantization/algorithms/min_max/torch_backend.py +++ b/nncf/quantization/algorithms/min_max/torch_backend.py @@ -9,7 +9,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from typing import Dict, List, Optional, Set, Tuple, Union +from typing import Dict, List, Optional, Set, Tuple import numpy as np import torch @@ -23,13 +23,13 @@ from nncf.common.graph.operator_metatypes import OperatorMetatype from nncf.common.graph.transformations.commands import TargetType from nncf.common.hardware.config import HWConfig -from nncf.common.quantization.initialization.range import RangeInitConfig from nncf.common.quantization.structs import QuantizationMode from nncf.common.quantization.structs import QuantizerConfig from nncf.common.utils.backend import BackendType +from nncf.experimental.common.tensor_statistics.collectors import AGGREGATORS_MAP +from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.parameters import ModelType from nncf.parameters import TargetDevice -from nncf.quantization.advanced_parameters import AggregatorType from nncf.quantization.advanced_parameters import StatisticsType from nncf.quantization.algorithms.min_max.backend import ALGO_BACKENDS from nncf.quantization.algorithms.min_max.backend import MinMaxAlgoBackend @@ -42,14 +42,13 @@ from nncf.torch.nncf_network import NNCFNetwork from nncf.torch.quantization.default_quantization import DEFAULT_PT_QUANT_TRAIT_TO_OP_DICT from nncf.torch.quantization.init_range import PTRangeInitCollectorParams -from nncf.torch.quantization.init_range import StatCollectorGenerator from nncf.torch.quantization.layers import QUANTIZATION_MODULES from nncf.torch.quantization.layers import AsymmetricQuantizer from nncf.torch.quantization.layers import BaseQuantizer from nncf.torch.quantization.layers import PTQuantizerSpec from nncf.torch.quantization.layers import get_scale_shape -from nncf.torch.tensor_statistics.collectors import PTMeanMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMinMaxStatisticCollector +from nncf.torch.tensor_statistics.collectors import PT_REDUCERS_MAP +from nncf.torch.tensor_statistics.collectors import PTNNCFCollectorTensorProcessor from nncf.torch.tensor_statistics.statistics import PTMinMaxTensorStatistic @@ -154,32 +153,47 @@ def get_statistic_collector( quantizer_config: QuantizerConfig, inplace: bool, num_samples: int = None, - ) -> Union[PTMinMaxStatisticCollector, PTMeanMinMaxStatisticCollector]: - if ( - range_estimator_params.min.statistics_type == StatisticsType.MIN - and range_estimator_params.min.aggregator_type == AggregatorType.MIN - and range_estimator_params.max.statistics_type == StatisticsType.MAX - and range_estimator_params.max.aggregator_type == AggregatorType.MAX + ) -> TensorCollector: + collector_params = PTMinMaxAlgoBackend._default_collector_params(nncf_graph, target_point, quantizer_config) + reduction_axes = collector_params.get_reduction_axes(per_sample_stats=False) + aggregation_axes = collector_params.get_aggregation_axes(per_sample_stats=False) + + collector = TensorCollector(PTMinMaxTensorStatistic) + for params, container_key in zip( + [range_estimator_params.min, range_estimator_params.max], + [PTMinMaxTensorStatistic.MIN_STAT, PTMinMaxTensorStatistic.MAX_STAT], ): - collector_name = "min_max" - - elif ( - range_estimator_params.min.statistics_type == StatisticsType.MIN - and range_estimator_params.min.aggregator_type == AggregatorType.MEAN - and range_estimator_params.max.statistics_type == StatisticsType.MAX - and range_estimator_params.max.aggregator_type == AggregatorType.MEAN - ): - collector_name = "mean_min_max" - - else: - raise RuntimeError( - "The following range estimator parameters are not supported by PyTorch backend by now: " - f"{str(range_estimator_params)}" + if not params.statistics_type in PT_REDUCERS_MAP: + raise RuntimeError( + f"Statistic type: {params.statistics_type} is not supported for Torch PTQ backend yet." + ) + + if not params.aggregator_type in AGGREGATORS_MAP: + raise RuntimeError( + f"Aggregator type: {params.aggregator_type} is not supported for Torch PTQ backend yet." + ) + + statistic_type = params.statistics_type + if statistic_type in [StatisticsType.QUANTILE, StatisticsType.ABS_QUANTILE]: + # TODO(dlyakhov): merge two quantile aggregators in one + if container_key == PTMinMaxTensorStatistic.MIN_STAT: + quantile = params.quantile_outlier_prob + else: + quantile = 1 - params.quantile_outlier_prob + reducer = PT_REDUCERS_MAP[statistic_type](reduction_axes=reduction_axes, quantile=[quantile]) + else: + if collector_params.use_abs_max and statistic_type == StatisticsType.MAX: + statistic_type = StatisticsType.ABS_MAX + reducer = PT_REDUCERS_MAP[statistic_type](reduction_axes=reduction_axes) + + aggregator = AGGREGATORS_MAP[params.aggregator_type]( + aggregation_axes=aggregation_axes, + num_samples=num_samples, + tensor_processor=PTNNCFCollectorTensorProcessor, ) - return PTMinMaxAlgoBackend._statistic_collector_builder( - collector_name, nncf_graph, target_point, quantizer_config, num_samples - ) + collector.register_statistic_branch(container_key, reducer, aggregator) + return collector @staticmethod def get_weight_tensor_port_ids(node: NNCFNode) -> List[Optional[int]]: @@ -222,37 +236,18 @@ def _get_input_scale_shape( return input_shape, scale_shape, channel_idx @staticmethod - def _default_collector_params_and_scale_shape( + def _default_collector_params( nncf_graph: NNCFGraph, target_point: PTTargetPoint, quantizer_config: QuantizerConfig - ) -> Tuple[PTRangeInitCollectorParams, Tuple[int, ...]]: - input_shape, scale_shape, channel_idx = PTMinMaxAlgoBackend._get_input_scale_shape( - nncf_graph, target_point, quantizer_config - ) - return ( - PTRangeInitCollectorParams( - is_weights=target_point.is_weight_target_point(), - mode=quantizer_config.mode, - per_channel=quantizer_config.per_channel, - input_shape=input_shape, - channel_idx=channel_idx, - ), - scale_shape, - ) - - @staticmethod - def _statistic_collector_builder( - collector_name: str, - nncf_graph: NNCFGraph, - target_point: PTTargetPoint, - quantizer_config: QuantizerConfig, - num_samples: int = None, - ) -> PTMeanMinMaxStatisticCollector: - collector_params, scale_shape = PTMinMaxAlgoBackend._default_collector_params_and_scale_shape( + ) -> PTRangeInitCollectorParams: + input_shape, _, channel_idx = PTMinMaxAlgoBackend._get_input_scale_shape( nncf_graph, target_point, quantizer_config ) - init_config = RangeInitConfig(collector_name, num_samples) - return StatCollectorGenerator.generate_stat_collector_for_range_init_config( - init_config, scale_shape, collector_params, num_samples + return PTRangeInitCollectorParams( + is_weights=target_point.is_weight_target_point(), + mode=quantizer_config.mode, + per_channel=quantizer_config.per_channel, + input_shape=input_shape, + channel_idx=channel_idx, ) @staticmethod diff --git a/nncf/quantization/algorithms/smooth_quant/algorithm.py b/nncf/quantization/algorithms/smooth_quant/algorithm.py index a9ccdef9e10..90c3baa7a13 100644 --- a/nncf/quantization/algorithms/smooth_quant/algorithm.py +++ b/nncf/quantization/algorithms/smooth_quant/algorithm.py @@ -232,11 +232,11 @@ def get_statistic_points(self, model: TModel, graph: NNCFGraph) -> StatisticPoin target_node_name=node_to_smooth.node_name, port_id=node_data["input_act_port"], ) - input_reduction_shape = self._calculate_input_reduction_shape( + input_reduction_axes = self._calculate_input_reduction_axes( graph, node_to_smooth, node_data["input_act_port"] ) stat_collector = self._backend_entity.get_abs_max_channel_collector( - self._subset_size, input_reduction_shape, self._inplace_statistics, STATISTIC_BRANCH_KEY + self._subset_size, input_reduction_axes, self._inplace_statistics, STATISTIC_BRANCH_KEY ) statistic_container.add_statistic_point( StatisticPoint( @@ -316,21 +316,21 @@ def _calculate_weight_scale(self, scale_value: TTensor, node: NNCFNode) -> TTens return self._backend_entity.calculate_weight_scale(scale_value, weights_size, channel_axis) return scale_value - def _calculate_input_reduction_shape(self, nncf_graph: NNCFGraph, node: NNCFNode, input_port: int) -> Tuple[int]: + def _calculate_input_reduction_axes(self, nncf_graph: NNCFGraph, node: NNCFNode, input_port: int) -> Tuple[int]: """ - Returns reduction shape for specified input. + Returns reduction axes for specified input. :param nncf_graph: NNCFGraph instance. :param node: NNCFNode to check. :param input_port: Specified input port id. - :return: Calculated reduction shape. + :return: Calculated reduction axes. """ shape = nncf_graph.get_input_edges(node)[input_port].tensor_shape - reduction_shape = tuple([0]) + reduction_axes = tuple([0]) if len(shape) > 1: channel_axis = self._backend_entity.get_activation_channel_axis(node, input_port) - reduction_shape = self._backend_entity.get_channel_agnostic_reduction_shape(channel_axis, shape) - return reduction_shape + reduction_axes = self._backend_entity.get_channel_agnostic_reduction_axes(channel_axis, shape) + return reduction_axes def _process_weight_statistics(self, node: NNCFNode, weights: TTensor, port_id: int) -> TTensor: """ diff --git a/nncf/quantization/algorithms/smooth_quant/backend.py b/nncf/quantization/algorithms/smooth_quant/backend.py index dda7fc44d2d..c6385d8b4f1 100644 --- a/nncf/quantization/algorithms/smooth_quant/backend.py +++ b/nncf/quantization/algorithms/smooth_quant/backend.py @@ -70,24 +70,24 @@ def get_input_ports_map(node: NNCFNode, nncf_graph: NNCFGraph) -> Dict[str, int] @staticmethod @abstractmethod - def get_channel_agnostic_reduction_shape(channel_axis: int, shape: Tuple[int]) -> Tuple[int]: + def get_channel_agnostic_reduction_axes(channel_axis: int, shape: Tuple[int]) -> Tuple[int]: """ - Returns filtered reduction shape without axes that corresponds channels. + Returns filtered reduction axes without axes that corresponds channels. :param channel_axes: List of the channel axes. :param shape: Shape that need to be filtered. - :return: Reduction shape in tuple format. + :return: Reduction axes in tuple format. """ @staticmethod @abstractmethod def get_abs_max_channel_collector( - num_samples: int, stats_reduction_shape: Tuple[int], inplace: bool, branch_key: str + num_samples: int, stats_reduction_axes: Tuple[int], inplace: bool, branch_key: str ) -> TensorCollector: """ Returns TensorCollector with MaxAggregator and AbsMaxReducer. - :param stats_reduction_shape: Calculated reduction shape. + :param stats_reduction_axes: Calculated reduction axes. :param inplace: Whether to calculate statistic inplace or not. :param branch_key: Specific string for branch key. :return: TensorCollector instance. diff --git a/nncf/quantization/algorithms/smooth_quant/openvino_backend.py b/nncf/quantization/algorithms/smooth_quant/openvino_backend.py index 415ea5a626a..12005168428 100644 --- a/nncf/quantization/algorithms/smooth_quant/openvino_backend.py +++ b/nncf/quantization/algorithms/smooth_quant/openvino_backend.py @@ -22,7 +22,7 @@ from nncf.experimental.common.tensor_statistics.collectors import MaxAggregator from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.openvino.graph.metatypes.openvino_metatypes import OVMatMulMetatype -from nncf.openvino.graph.node_utils import get_channel_agnostic_reduction_shape +from nncf.openvino.graph.node_utils import get_channel_agnostic_reduction_axes from nncf.openvino.graph.node_utils import get_weight_value from nncf.openvino.graph.transformations.command_creation import OVCommandCreator from nncf.openvino.graph.transformations.commands import OVMultiplyInsertionCommand @@ -61,16 +61,16 @@ def get_input_ports_map(node: NNCFNode, nncf_graph: NNCFGraph) -> Dict[str, int] return {"activation": activation_ports[0], "weight": weight_ports[0]} @staticmethod - def get_channel_agnostic_reduction_shape(channel_axis: int, shape: Tuple[int]) -> Tuple[int]: - return get_channel_agnostic_reduction_shape([channel_axis], shape) + def get_channel_agnostic_reduction_axes(channel_axis: int, shape: Tuple[int]) -> Tuple[int]: + return get_channel_agnostic_reduction_axes([channel_axis], shape) @staticmethod def get_abs_max_channel_collector( - num_samples: int, stats_reduction_shape: Tuple[int], inplace: bool, branch_key: str + num_samples: int, stats_reduction_axes: Tuple[int], inplace: bool, branch_key: str ) -> TensorCollector: collector = TensorCollector() - reducer = OVAbsMaxReducer(stats_reduction_shape, inplace) - aggregator = MaxAggregator(OVNNCFCollectorTensorProcessor, num_samples) + reducer = OVAbsMaxReducer(reduction_axes=stats_reduction_axes, inplace=inplace) + aggregator = MaxAggregator(tensor_processor=OVNNCFCollectorTensorProcessor, num_samples=num_samples) collector.register_statistic_branch(branch_key, reducer, aggregator) return collector diff --git a/nncf/scopes.py b/nncf/scopes.py index 4b6adcd03cf..7145d33a4a3 100644 --- a/nncf/scopes.py +++ b/nncf/scopes.py @@ -121,7 +121,7 @@ def get_ignored_node_names_from_ignored_scope( raise RuntimeError( f"Ignored nodes with name {list(skipped_names)} were not found in the NNCFGraph. " + error_msg ) - nncf_logger.info(f"{len(matched_by_names)} ignored nodes was found by name in the NNCFGraph") + nncf_logger.info(f"{len(matched_by_names)} ignored nodes were found by name in the NNCFGraph") matched_by_patterns = [] if ignored_scope.patterns: @@ -134,7 +134,7 @@ def get_ignored_node_names_from_ignored_scope( matched_by_patterns.extend(matches) if strict and not_matched_patterns: raise RuntimeError(f"No matches for ignored patterns {not_matched_patterns} in the NNCFGraph. " + error_msg) - nncf_logger.info(f"{len(matched_by_patterns)} ignored nodes was found by patterns in the NNCFGraph") + nncf_logger.info(f"{len(matched_by_patterns)} ignored nodes were found by patterns in the NNCFGraph") matched_by_types = [] if ignored_scope.types: @@ -148,6 +148,6 @@ def get_ignored_node_names_from_ignored_scope( raise RuntimeError( f"Nodes with ignored types {list(not_matched_types)} were not found in the NNCFGraph. " + error_msg ) - nncf_logger.info(f"{len(matched_by_types)} ignored nodes was found by types in the NNCFGraph") + nncf_logger.info(f"{len(matched_by_types)} ignored nodes were found by types in the NNCFGraph") return set(matched_by_names + matched_by_types + matched_by_patterns) diff --git a/nncf/tensorflow/quantization/init_range.py b/nncf/tensorflow/quantization/init_range.py index 2d6bff0a405..900f3248134 100644 --- a/nncf/tensorflow/quantization/init_range.py +++ b/nncf/tensorflow/quantization/init_range.py @@ -23,7 +23,7 @@ from nncf.common.quantization.initialization.range import RangeInitParams from nncf.common.quantization.structs import QuantizerGroup from nncf.common.scopes import should_consider_scope -from nncf.common.tensor_statistics.collectors import ReductionShape +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase from nncf.config.schemata.defaults import MAX_PERCENTILE from nncf.config.schemata.defaults import MIN_PERCENTILE @@ -99,7 +99,7 @@ def __init__(self, range_init_params: TFRangeInitParams): @staticmethod def generate_stat_collector( - reduction_shape: ReductionShape, + reduction_shape: ReductionAxes, collector_params: RangeInitCollectorParams, init_config: RangeInitConfig, num_samples_to_collect_override: int = None, diff --git a/nncf/tensorflow/tensor_statistics/collectors.py b/nncf/tensorflow/tensor_statistics/collectors.py index 0b47a605eff..a8c3da70a8b 100644 --- a/nncf/tensorflow/tensor_statistics/collectors.py +++ b/nncf/tensorflow/tensor_statistics/collectors.py @@ -9,7 +9,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -from typing import Any, Callable, Deque, List, Optional, Union +from typing import Deque, List, Optional, Tuple, Union import numpy as np import tensorflow as tf @@ -37,11 +37,11 @@ class TFNNCFCollectorTensorProcessor(NNCFCollectorTensorProcessor): """ @staticmethod - def reduce_min(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def reduce_min(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: return TFNNCFTensor(tf.reduce_min(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod - def reduce_max(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def reduce_max(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: return TFNNCFTensor(tf.reduce_max(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod @@ -57,19 +57,31 @@ def max(x1: tf.Tensor, x2: tf.Tensor) -> NNCFTensor: return TFNNCFTensor(tf.math.maximum(x1.tensor, x2.tensor)) @staticmethod - def mean(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTensor: + def mean(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims=False) -> NNCFTensor: return TFNNCFTensor(tf.math.reduce_mean(x.tensor, axis=axis, keepdims=keepdims)) @staticmethod - def median(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTensor: + def median(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims=False) -> NNCFTensor: + raise NotImplementedError() + + @classmethod + def masked_mean( + cls, x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], mask: NNCFTensor, keepdims=False + ) -> NNCFTensor: + raise NotImplementedError() + + @classmethod + def masked_median( + cls, x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], mask: NNCFTensor, keepdims=False + ) -> NNCFTensor: raise NotImplementedError() @staticmethod - def masked_mean(x: NNCFTensor, axis: Union[int, tuple, list], mask: NNCFTensor, keepdims=False) -> NNCFTensor: + def logical_or(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: raise NotImplementedError() @staticmethod - def masked_median(x: NNCFTensor, axis: Union[int, tuple, list], mask: NNCFTensor, keepdims=False) -> NNCFTensor: + def less(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: raise NotImplementedError() @staticmethod @@ -85,24 +97,43 @@ def unstack(x: NNCFTensor, axis: int = 0) -> List[NNCFTensor]: tensor_list = tf.unstack(tensor, axis=axis) return [TFNNCFTensor(t) for t in tensor_list] + @staticmethod + def squeeze(x: NNCFTensor, dim: Optional[Union[int, Tuple[int, ...]]] = None) -> NNCFTensor: + raise NotImplementedError() + @staticmethod def sum(tensor: NNCFTensor) -> TensorElementsType: return tf.reduce_sum(tensor.tensor).numpy() @staticmethod def quantile( - tensor: NNCFTensor, quantile: Union[float, List[float]], axis: Union[int, tuple, list], keepdims: bool = False + tensor: NNCFTensor, + quantile: Union[float, List[float]], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, ) -> List[NNCFTensor]: raise NotImplementedError() + @classmethod + def percentile( + cls, + tensor: NNCFTensor, + percentile: Union[float, List[float]], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, + ) -> List[TensorElementsType]: + raise NotImplementedError() + @staticmethod def mean_per_channel(x: NNCFTensor, axis: int) -> NNCFTensor: raise NotImplementedError() - @classmethod - def no_outliers_map( - cls, x: NNCFTensor, fn: Callable[[NNCFTensor, Optional[int]], Any], axis: int = 0, alpha: float = 0.01 - ): + @staticmethod + def sub(a: NNCFTensor, b: NNCFTensor) -> NNCFTensor: + raise NotImplementedError() + + @staticmethod + def zero_elements(x: NNCFTensor) -> NNCFTensor: raise NotImplementedError() diff --git a/nncf/tensorflow/tensor_statistics/reduction.py b/nncf/tensorflow/tensor_statistics/reduction.py index d43afddf947..00229d88be5 100644 --- a/nncf/tensorflow/tensor_statistics/reduction.py +++ b/nncf/tensorflow/tensor_statistics/reduction.py @@ -13,7 +13,7 @@ import tensorflow as tf -from nncf.common.tensor_statistics.collectors import ReductionShape +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.tensorflow.layers.data_layout import get_weight_shape @@ -28,7 +28,7 @@ def get_axes(ndims: int, per_channel: bool, channel_axes: Union[int, list, tuple def get_reduction_shape_activations( layer: tf.keras.layers.Layer, channel_axes: Union[int, tuple, list], use_per_sample_stats: bool -) -> ReductionShape: +) -> ReductionAxes: ndims = len(layer.get_input_shape_at(0)) channel_axes_ = channel_axes if isinstance(channel_axes, (list, tuple)) else [channel_axes] reduction_shape = get_axes(ndims, layer.per_channel, channel_axes_) @@ -39,7 +39,7 @@ def get_reduction_shape_activations( def get_reduction_shape_weights( layer: tf.keras.layers.Layer, weight_attr: str, channel_axes: Union[int, tuple, list], per_channel: bool -) -> ReductionShape: +) -> ReductionAxes: weight_shape = get_weight_shape(layer, weight_attr) ndims = len(weight_shape) channel_axes_ = channel_axes if isinstance(channel_axes, (list, tuple)) else [channel_axes] @@ -47,7 +47,7 @@ def get_reduction_shape_weights( return tuple(reduction_shape) -def convert_rs_to_pt_type(input_shape: Tuple[int], reduction_shape: ReductionShape) -> ReductionShape: +def convert_rs_to_pt_type(input_shape: Tuple[int], reduction_shape: ReductionAxes) -> ReductionAxes: if len(reduction_shape) == len(input_shape): pt_reduction_shape = [1] else: diff --git a/nncf/torch/quantization/algo.py b/nncf/torch/quantization/algo.py index 2ead247854b..673d2918311 100644 --- a/nncf/torch/quantization/algo.py +++ b/nncf/torch/quantization/algo.py @@ -57,6 +57,7 @@ from nncf.common.quantization.structs import WeightQuantizerId from nncf.common.schedulers import BaseCompressionScheduler from nncf.common.statistics import NNCFStatistics +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.common.utils.api_marker import api from nncf.common.utils.backend import BackendType from nncf.common.utils.backend import copy_model @@ -130,7 +131,6 @@ from nncf.torch.structures import AutoQPrecisionInitArgs from nncf.torch.structures import QuantizationPrecisionInitArgs from nncf.torch.tensor_statistics.algo import TensorStatisticsCollectionBuilder -from nncf.torch.tensor_statistics.collectors import ReductionShape from nncf.torch.tensor_statistics.statistics import MinMaxTensorStatistic from nncf.torch.tensor_statistics.statistics import TensorStatistic from nncf.torch.tensor_statistics.statistics import pt_convert_stat_to_min_max_tensor_stat @@ -582,7 +582,7 @@ def _parse_precision_init_params(self, initializer_config: Dict) -> Tuple[str, B def _get_minmax_values_for_quantizer_locations( self, quantizer_setup: SingleConfigQuantizerSetup, - tensor_statistics: Dict[PTTargetPoint, Dict[ReductionShape, TensorStatistic]], + tensor_statistics: Dict[PTTargetPoint, Dict[ReductionAxes, TensorStatistic]], target_model_graph: PTNNCFGraph, ) -> Dict[QuantizationPointId, MinMaxTensorStatistic]: retval = {} @@ -662,7 +662,7 @@ def _get_transformation_layout(self, target_model: NNCFNetwork) -> PTTransformat @staticmethod def get_statistics_for_quantizer_setup( target_model: NNCFNetwork, quantizer_setup: QuantizerSetupBase, range_init_params: PTRangeInitParams - ) -> Dict[PTTargetPoint, Dict[ReductionShape, TensorStatistic]]: + ) -> Dict[PTTargetPoint, Dict[ReductionAxes, TensorStatistic]]: if range_init_params is None: return {} observation_points_vs_collectors_dict = ( @@ -688,7 +688,7 @@ def get_statistics_for_quantizer_setup( def _get_statistics_for_final_range_init( self, target_model: NNCFNetwork, quantizer_setup: QuantizerSetupBase, range_init_params: PTRangeInitParams - ) -> Dict[PTTargetPoint, Dict[ReductionShape, TensorStatistic]]: + ) -> Dict[PTTargetPoint, Dict[ReductionAxes, TensorStatistic]]: return self.get_statistics_for_quantizer_setup(target_model, quantizer_setup, range_init_params) def _get_single_config_quantizer_setup(self, target_model) -> SingleConfigQuantizerSetup: @@ -1508,7 +1508,7 @@ def __init__( self, quantizer_setup: MultiConfigQuantizerSetup, initial_quantizer_setup: SingleConfigQuantizerSetup, - tensor_stats_for_all_setup_variations: Dict[PTTargetPoint, Dict[ReductionShape, TensorStatistic]], + tensor_stats_for_all_setup_variations: Dict[PTTargetPoint, Dict[ReductionAxes, TensorStatistic]], hw_config: HWConfig = None, ): should_init = bool(tensor_stats_for_all_setup_variations) @@ -1527,7 +1527,7 @@ def _get_single_config_quantizer_setup(self, target_model) -> SingleConfigQuanti def _get_statistics_for_final_range_init( self, target_model: NNCFNetwork, quantizer_setup: QuantizerSetupBase, range_init_params: PTRangeInitParams - ) -> Dict[PTTargetPoint, Dict[ReductionShape, TensorStatistic]]: + ) -> Dict[PTTargetPoint, Dict[ReductionAxes, TensorStatistic]]: return self._tensor_stats def _build_controller(self, model: NNCFNetwork) -> "ExperimentalQuantizationController": @@ -1583,7 +1583,7 @@ def __init__( quantizer_setup: MultiConfigQuantizerSetup, initial_quantizer_setup: SingleConfigQuantizerSetup, setup_to_module_id_translation_dict: Dict[QuantizationPointId, QuantizerId], - tensor_stats: Dict[PTTargetPoint, Dict[ReductionShape, TensorStatistic]], + tensor_stats: Dict[PTTargetPoint, Dict[ReductionAxes, TensorStatistic]], build_time_metric_info: QuantizationShareBuildTimeInfo, should_setup_adjust_pad_ops=False, hw_config: HWConfig = None, diff --git a/nncf/torch/quantization/init_range.py b/nncf/torch/quantization/init_range.py index 259e7ec1ada..28a5204ce22 100644 --- a/nncf/torch/quantization/init_range.py +++ b/nncf/torch/quantization/init_range.py @@ -14,6 +14,7 @@ from typing import Callable, Dict, List, Tuple import numpy as np +import torch from nncf.common.graph.layer_attributes import WeightedLayerAttributes from nncf.common.quantization.initialization.range import RangeInitCollectorParams @@ -27,9 +28,10 @@ from nncf.common.quantization.structs import QuantizerId from nncf.common.quantization.structs import WeightQuantizerId from nncf.common.scopes import should_consider_scope -from nncf.common.tensor_statistics.collectors import ReductionShape +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase from nncf.config.schemata.algo.quantization import RANGE_INIT_TYPES_VS_DESCRIPTIONS +from nncf.experimental.common.tensor_statistics.collectors import AggregationAxes from nncf.torch.graph.graph import PTNNCFGraph from nncf.torch.initialization import DataLoaderBaseRunner from nncf.torch.nncf_network import NNCFNetwork @@ -38,12 +40,12 @@ from nncf.torch.quantization.layers import get_scale_shape from nncf.torch.quantization.translator import PTTargetPointTranslator from nncf.torch.tensor_statistics.algo import TensorStatisticObservationPoint -from nncf.torch.tensor_statistics.collectors import PTMeanMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMeanPercentileStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMedianMADStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMixedMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTPercentileStatisticCollector +from nncf.torch.tensor_statistics.algo import create_register_input_hook +from nncf.torch.tensor_statistics.collectors import get_mean_percentile_statistic_collector +from nncf.torch.tensor_statistics.collectors import get_median_mad_statistic_collector +from nncf.torch.tensor_statistics.collectors import get_min_max_statistic_collector +from nncf.torch.tensor_statistics.collectors import get_mixed_min_max_statistic_collector +from nncf.torch.tensor_statistics.collectors import get_percentile_tensor_collector from nncf.torch.tensor_statistics.statistics import pt_convert_stat_to_min_max_tensor_stat @@ -103,28 +105,39 @@ def __init__( self._input_shape = input_shape self._channel_idx = channel_idx - def convert_reduction_shape(self, per_sample_stats) -> ReductionShape: + def get_reduction_axes(self, per_sample_stats: bool) -> ReductionAxes: """ - Calculates the reduction shape of the tensor. + Calculates the reduction axes of the tensor. :param per_sample_stats: Boolean flag that indicated whether statistics are collected per-sample or per-batch. :return: Shape to reduce to. """ ndims = len(self._input_shape) - reduction_shape = list(range(ndims)) # type: List[int] + reduction_axes = list(range(ndims)) # type: List[int] if self._per_channel: val = (ndims + self._channel_idx) % ndims - reduction_shape.remove(val) + reduction_axes.remove(val) + if not val and self.use_per_sample_stats(per_sample_stats): + raise RuntimeError("Batch dimension should be equal to zero") if self.use_per_sample_stats(per_sample_stats): - reduction_shape = reduction_shape[1:] # Assumes batch is the first dimension - return tuple(reduction_shape) + reduction_axes = reduction_axes[1:] # Assumes batch is the first dimension + return tuple(reduction_axes) + + def get_aggregation_axes(self, per_sample_stats: bool) -> AggregationAxes: + """ + Calculates the aggregation axes of the tensor. + + :param per_sample_stats: Boolean flag that indicated whether statistics are collected per-sample or per-batch. + :return: Shape to aggregate to. + """ + return (0, 1) if self.use_per_sample_stats(per_sample_stats) else (0,) class StatCollectorGenerator: @staticmethod def generate_collectors_for_range_init_statistics_collection( target_model_graph: PTNNCFGraph, quantizer_setup: QuantizerSetupBase, range_init_params: PTRangeInitParams - ) -> Dict[TensorStatisticObservationPoint, Dict[ReductionShape, TensorStatisticCollectorBase]]: + ) -> Dict[TensorStatisticObservationPoint, Dict[ReductionAxes, TensorStatisticCollectorBase]]: retval = {} for qp in quantizer_setup.quantization_points.values(): init_config = range_init_params.get_init_config_for_quantization_point(qp) @@ -154,8 +167,8 @@ def generate_collectors_for_range_init_statistics_collection( @staticmethod def generate_stat_collector_for_range_init_config( init_config: RangeInitConfig, - reduction_shape: ReductionShape = None, - collector_params=None, + scale_shape: ReductionAxes = None, + collector_params: PTRangeInitCollectorParams = None, num_samples_to_collect_override: int = None, ) -> TensorStatisticCollectorBase: num_samples = init_config.num_init_samples @@ -163,47 +176,73 @@ def generate_stat_collector_for_range_init_config( num_samples = num_samples_to_collect_override if init_config.init_type not in RANGE_INIT_TYPES_VS_DESCRIPTIONS: raise RuntimeError("Unknown range init type: {}".format(init_config.init_type)) + + use_per_sample_stats = collector_params.use_per_sample_stats(init_config.init_type == "mixed_min_max") + reduction_axes = collector_params.get_reduction_axes(use_per_sample_stats) + aggregation_axes = collector_params.get_aggregation_axes(use_per_sample_stats) + if init_config.init_type == "min_max": - reduction_shape_converted = collector_params.convert_reduction_shape(per_sample_stats=False) - return PTMinMaxStatisticCollector( - collector_params.use_abs_max, reduction_shape_converted, reduction_shape, num_samples + return get_min_max_statistic_collector( + use_abs_max=collector_params.use_abs_max, + reduction_axes=reduction_axes, + aggregation_axes=aggregation_axes, + scale_shape=scale_shape, + num_samples=num_samples, ) if init_config.init_type == "mixed_min_max": - reduction_shape_converted = collector_params.convert_reduction_shape(per_sample_stats=True) - return PTMixedMinMaxStatisticCollector( - collector_params.use_per_sample_stats(per_sample_stats=True), - collector_params.use_abs_max, - collector_params.use_means_of_mins, - collector_params.use_means_of_maxs, - reduction_shape_converted, - reduction_shape, - num_samples, + return get_mixed_min_max_statistic_collector( + use_abs_max=collector_params.use_abs_max, + reduction_axes=reduction_axes, + aggregation_axes=aggregation_axes, + scale_shape=scale_shape, + use_means_of_mins=collector_params.use_means_of_mins, + use_means_of_maxs=collector_params.use_means_of_maxs, + num_samples=num_samples, ) if init_config.init_type == "mean_min_max": - reduction_shape_converted = collector_params.convert_reduction_shape(per_sample_stats=False) - return PTMeanMinMaxStatisticCollector( - collector_params.use_per_sample_stats(per_sample_stats=False), - collector_params.use_abs_max, - reduction_shape_converted, - reduction_shape, - num_samples, + return get_mixed_min_max_statistic_collector( + use_abs_max=collector_params.use_abs_max, + reduction_axes=reduction_axes, + aggregation_axes=aggregation_axes, + scale_shape=scale_shape, + use_means_of_mins=True, + use_means_of_maxs=True, + num_samples=num_samples, ) if init_config.init_type == "threesigma": - return PTMedianMADStatisticCollector(reduction_shape, num_samples) + return get_median_mad_statistic_collector( + reduction_axes=reduction_axes, + aggregation_axes=aggregation_axes, + scale_shape=scale_shape, + num_samples=num_samples, + ) if init_config.init_type == "percentile": min_percentile = init_config.init_type_specific_params.get("min_percentile", 0.1) max_percentile = init_config.init_type_specific_params.get("max_percentile", 99.9) - return PTPercentileStatisticCollector([min_percentile, max_percentile], reduction_shape, num_samples) + return get_percentile_tensor_collector( + percentiles_to_collect=(min_percentile, max_percentile), + reduction_axes=reduction_axes, + aggregation_axes=aggregation_axes, + scale_shape=scale_shape, + num_samples=num_samples, + ) + if init_config.init_type == "mean_percentile": min_percentile = init_config.init_type_specific_params.get("min_percentile", 0.1) max_percentile = init_config.init_type_specific_params.get("max_percentile", 99.9) - return PTMeanPercentileStatisticCollector([min_percentile, max_percentile], reduction_shape, num_samples) + return get_mean_percentile_statistic_collector( + percentiles_to_collect=(min_percentile, max_percentile), + reduction_axes=reduction_axes, + aggregation_axes=aggregation_axes, + scale_shape=scale_shape, + num_samples=num_samples, + ) raise ValueError("Range init type not handled!") @classmethod def get_all_scale_shapes_with_params( cls, qp: QuantizationPointBase, target_nncf_graph: PTNNCFGraph - ) -> Dict[ReductionShape, PTRangeInitCollectorParams]: + ) -> Dict[ReductionAxes, PTRangeInitCollectorParams]: qconfigs = qp.get_all_configs_list() if qp.is_weight_quantization_point(): module_node = target_nncf_graph.get_node_by_name(qp.insertion_point.target_node_name) @@ -249,9 +288,13 @@ def __init__( self.hook_handles = [] self.batch_size = batch_size - def _get_fwd_hook(self, collector: TensorStatisticCollectorBase) -> Callable: + def _get_fwd_hook( + self, collector: TensorStatisticCollectorBase + ) -> Callable[["torch.Module", torch.Tensor, torch.Tensor], torch.Tensor]: + hook = create_register_input_hook(collector=collector) + def fwd_hook(module, input_, output): - collector.register_input(input_[0]) + hook(input_[0]) return fwd_hook diff --git a/nncf/torch/statistics/aggregator.py b/nncf/torch/statistics/aggregator.py index 6c2c48256c6..41fdc20c4fa 100644 --- a/nncf/torch/statistics/aggregator.py +++ b/nncf/torch/statistics/aggregator.py @@ -23,6 +23,7 @@ from nncf.torch.graph.transformations.commands import PTInsertionCommand from nncf.torch.nncf_network import NNCFNetwork from nncf.torch.tensor import PTNNCFTensor +from nncf.torch.tensor_statistics.algo import create_register_input_hook class PTStatisticsAggregator(StatisticsAggregator): @@ -41,6 +42,7 @@ def _get_transformation_layout_extra_outputs( ) -> TransformationLayout: transformation_layout = TransformationLayout() transformation_commands = [] + for _statistic_points in statistic_points.values(): for _statistic_point in _statistic_points: for collectors in _statistic_point.algorithm_to_tensor_collectors.values(): @@ -48,7 +50,7 @@ def _get_transformation_layout_extra_outputs( transformation_commands.append( PTInsertionCommand( _statistic_point.target_point, - collector.register_input, + create_register_input_hook(collector=collector), TransformationPriority.FP32_TENSOR_STATISTICS_OBSERVATION, ) ) diff --git a/nncf/torch/tensor.py b/nncf/torch/tensor.py index 6dd4e88b68a..986adb46aa4 100644 --- a/nncf/torch/tensor.py +++ b/nncf/torch/tensor.py @@ -30,3 +30,6 @@ def __init__(self, tensor: torch.tensor): @property def device(self) -> torch.device: return self._tensor.device + + def is_empty(self) -> bool: + return self.tensor.size == 0 diff --git a/nncf/torch/tensor_statistics/algo.py b/nncf/torch/tensor_statistics/algo.py index 0f9c56f14e5..55cf2ee4b1b 100644 --- a/nncf/torch/tensor_statistics/algo.py +++ b/nncf/torch/tensor_statistics/algo.py @@ -8,26 +8,31 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -from typing import Dict, Set +from typing import Callable, Dict, Set + +import torch from nncf.api.compression import CompressionStage from nncf.common.schedulers import StubCompressionScheduler from nncf.common.statistics import NNCFStatistics -from nncf.common.tensor_statistics.collectors import ReductionShape +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase from nncf.config import NNCFConfig +from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.torch.algo_selector import ZeroCompressionLoss from nncf.torch.compression_method_api import PTCompressionAlgorithmBuilder from nncf.torch.compression_method_api import PTCompressionAlgorithmController +from nncf.torch.dynamic_graph.context import no_nncf_trace from nncf.torch.graph.transformations.commands import PTInsertionCommand from nncf.torch.graph.transformations.commands import PTTargetPoint from nncf.torch.graph.transformations.commands import TransformationPriority from nncf.torch.graph.transformations.layout import PTTransformationLayout from nncf.torch.nncf_network import NNCFNetwork +from nncf.torch.tensor import PTNNCFTensor class TensorStatisticObservationPoint: - def __init__(self, target_point: PTTargetPoint, reduction_shapes: Set[ReductionShape] = None): + def __init__(self, target_point: PTTargetPoint, reduction_shapes: Set[ReductionAxes] = None): self.target_point = target_point self.reduction_shapes = reduction_shapes @@ -38,6 +43,28 @@ def __eq__(self, other: "TensorStatisticObservationPoint"): return self.target_point == other.target_point +def create_register_input_hook(collector: TensorCollector) -> Callable[[torch.Tensor], torch.Tensor]: + """ + Function to create regiter inputs hook function. + + :param collector: Collector to use in resulting hook. + :return: Register inputs hook function. + """ + + def register_inputs_hook(x: torch.Tensor) -> torch.Tensor: + """ + Register inputs hook function. + + :parameter x: tensor to register in hook. + :return: tensor to register in hook. + """ + with no_nncf_trace(): + collector.register_input_for_all_reducers(PTNNCFTensor(x)) + return x + + return register_inputs_hook + + class TensorStatisticsCollectionBuilder(PTCompressionAlgorithmBuilder): def __init__( self, @@ -54,9 +81,10 @@ def _get_transformation_layout(self, target_model: NNCFNetwork) -> PTTransformat layout = PTTransformationLayout() for op, rs_vs_collector in self._observation_points_vs_collectors.items(): for collector in rs_vs_collector.values(): - hook_obj = collector.register_input command = PTInsertionCommand( - op.target_point, hook_obj, TransformationPriority.FP32_TENSOR_STATISTICS_OBSERVATION + op.target_point, + create_register_input_hook(collector=collector), + TransformationPriority.FP32_TENSOR_STATISTICS_OBSERVATION, ) layout.register(command) return layout diff --git a/nncf/torch/tensor_statistics/collectors.py b/nncf/torch/tensor_statistics/collectors.py index 186bbb18393..4089fb77f2e 100644 --- a/nncf/torch/tensor_statistics/collectors.py +++ b/nncf/torch/tensor_statistics/collectors.py @@ -9,71 +9,107 @@ # See the License for the specific language governing permissions and # limitations under the License. -from typing import Any, Callable, Deque, List, Optional, Union +from functools import partial +from typing import Deque, List, Optional, Tuple, Type, Union +import numpy as np import torch -from nncf.common.tensor import NNCFTensor from nncf.common.tensor import TensorElementsType -from nncf.common.tensor_statistics.collectors import MeanMinMaxStatisticCollector -from nncf.common.tensor_statistics.collectors import MeanPercentileStatisticCollector -from nncf.common.tensor_statistics.collectors import MeanStatisticCollector -from nncf.common.tensor_statistics.collectors import MedianMADStatisticCollector -from nncf.common.tensor_statistics.collectors import MinMaxStatisticCollector -from nncf.common.tensor_statistics.collectors import MixedMinMaxStatisticCollector from nncf.common.tensor_statistics.collectors import NNCFCollectorTensorProcessor -from nncf.common.tensor_statistics.collectors import PercentileStatisticCollector -from nncf.common.tensor_statistics.collectors import ReductionShape -from nncf.common.tensor_statistics.reduction import np_percentile_reduce_like -from nncf.torch.dynamic_graph.context import no_nncf_trace +from nncf.common.tensor_statistics.collectors import NNCFTensor +from nncf.experimental.common.tensor_statistics.collectors import AbsMaxReducer +from nncf.experimental.common.tensor_statistics.collectors import AbsQuantileReducer +from nncf.experimental.common.tensor_statistics.collectors import BatchMeanReducer +from nncf.experimental.common.tensor_statistics.collectors import MaxAggregator +from nncf.experimental.common.tensor_statistics.collectors import MaxReducer +from nncf.experimental.common.tensor_statistics.collectors import MeanAggregator +from nncf.experimental.common.tensor_statistics.collectors import MeanPerChReducer +from nncf.experimental.common.tensor_statistics.collectors import MeanReducer +from nncf.experimental.common.tensor_statistics.collectors import MedianAbsoluteDeviationAggregator +from nncf.experimental.common.tensor_statistics.collectors import MinAggregator +from nncf.experimental.common.tensor_statistics.collectors import MinReducer +from nncf.experimental.common.tensor_statistics.collectors import NoopReducer +from nncf.experimental.common.tensor_statistics.collectors import PercentileAggregator +from nncf.experimental.common.tensor_statistics.collectors import QuantileReducer +from nncf.experimental.common.tensor_statistics.collectors import ShapeAggregator +from nncf.experimental.common.tensor_statistics.collectors import TensorAggregatorBase +from nncf.experimental.common.tensor_statistics.collectors import TensorCollector +from nncf.quantization.advanced_parameters import StatisticsType from nncf.torch.tensor import PTNNCFTensor -from nncf.torch.tensor_statistics.reduction import expand_like from nncf.torch.tensor_statistics.statistics import PTMeanTensorStatistic from nncf.torch.tensor_statistics.statistics import PTMedianMADTensorStatistic from nncf.torch.tensor_statistics.statistics import PTMinMaxTensorStatistic from nncf.torch.tensor_statistics.statistics import PTPercentileTensorStatistic +# pylint: disable=too-many-public-methods class PTNNCFCollectorTensorProcessor(NNCFCollectorTensorProcessor): """ A realization of the processing methods for PTNNCFTensors. """ @staticmethod - def reduce_min(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def reduce_min(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: return PTNNCFTensor(torch.amin(x.tensor, dim=axis, keepdim=keepdims)) @staticmethod - def reduce_max(x: NNCFTensor, axis: Union[int, tuple, list], keepdims: bool = False) -> NNCFTensor: + def reduce_max(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims: bool = False) -> NNCFTensor: return PTNNCFTensor(torch.amax(x.tensor, dim=axis, keepdim=keepdims)) @staticmethod def abs(x: NNCFTensor) -> NNCFTensor: return PTNNCFTensor(torch.abs(x.tensor)) - @staticmethod - def min(x1: NNCFTensor, x2: NNCFTensor) -> NNCFTensor: - return PTNNCFTensor(torch.min(x1.tensor, x2.tensor)) + @classmethod + def min(cls, *args) -> NNCFTensor: + stacked = cls.stack(args) + return cls.reduce_min(stacked, axis=0, keepdims=False) - @staticmethod - def max(x1: NNCFTensor, x2: NNCFTensor) -> NNCFTensor: - return PTNNCFTensor(torch.max(x1.tensor, x2.tensor)) + @classmethod + def max(cls, *args) -> NNCFTensor: + stacked = cls.stack(args) + return cls.reduce_max(stacked, axis=0, keepdims=False) @staticmethod - def mean(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTensor: + def mean(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims=False) -> NNCFTensor: return PTNNCFTensor(x.tensor.mean(dim=axis, keepdim=keepdims)) @staticmethod - def median(x: NNCFTensor, axis: Union[int, tuple, list], keepdims=False) -> NNCFTensor: - return PTNNCFTensor(x.tensor.median(dim=axis, keepdim=keepdims)) + def median(x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], keepdims=False) -> NNCFTensor: + # See https://github.com/pytorch/pytorch/issues/61582 + if not isinstance(axis, int): + device = x.tensor.device + result = torch.tensor(np.median(x.tensor.detach().cpu().numpy(), axis=axis, keepdims=keepdims)) + return PTNNCFTensor(result.type(x.tensor.dtype).to(device)) + return PTNNCFTensor(torch.quantile(x.tensor, q=0.5, dim=axis, keepdim=keepdims).values) - @staticmethod - def masked_mean(x: NNCFTensor, axis: Union[int, tuple, list], mask: NNCFTensor, keepdims=False) -> NNCFTensor: - raise NotImplementedError() + @classmethod + def masked_mean( + cls, x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], mask: NNCFTensor, keepdims=False + ) -> NNCFTensor: + if mask is None: + return cls.mean(x, axis=axis, keepdims=keepdims) + device = x.tensor.device + masked_x = np.ma.array(x.tensor.detach().cpu().numpy(), mask=mask.tensor.detach().cpu().numpy()) + result = np.ma.mean(masked_x, axis=axis, keepdims=keepdims).astype(masked_x.dtype) + if isinstance(result, np.ma.MaskedArray): + result = result.data + return PTNNCFTensor(torch.tensor(result).to(device=device)) - @staticmethod - def masked_median(x: NNCFTensor, axis: Union[int, tuple, list], mask: NNCFTensor, keepdims=False) -> NNCFTensor: - raise NotImplementedError() + @classmethod + def masked_median( + cls, x: NNCFTensor, axis: Union[int, Tuple[int, ...], List[int]], mask: NNCFTensor, keepdims=False + ) -> NNCFTensor: + # Implemented in numy as torch.masked.median is not implemented yet + if mask is None: + return cls.median(x, axis=axis, keepdims=keepdims) + device = x.tensor.device + masked_x = np.ma.array(x.tensor.detach().cpu().numpy(), mask=mask.tensor.detach().cpu().numpy()) + result = np.ma.median(masked_x, axis=axis, keepdims=keepdims).astype(masked_x.dtype) + if isinstance(result, np.ma.MaskedArray): + result = result.data + return PTNNCFTensor(torch.tensor(result).to(device=device)) @staticmethod def mean_per_channel(x: NNCFTensor, axis: int) -> NNCFTensor: @@ -87,6 +123,14 @@ def mean_per_channel(x: NNCFTensor, axis: int) -> NNCFTensor: def batch_mean(x: NNCFTensor) -> NNCFTensor: return PTNNCFTensor(torch.mean(x.tensor, axis=0, keepdims=True)) + @staticmethod + def logical_or(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: + return PTNNCFTensor(torch.logical_or(input_.tensor, other.tensor)) + + @staticmethod + def less(input_: NNCFTensor, other: NNCFTensor) -> NNCFTensor: + return PTNNCFTensor(input_.tensor < other.tensor) + @staticmethod def stack(x: Union[List[NNCFTensor], Deque[NNCFTensor]], axis: int = 0) -> NNCFTensor: x = [t.tensor for t in x] @@ -100,161 +144,395 @@ def unstack(x: NNCFTensor, axis: int = 0) -> List[NNCFTensor]: tensor_list = torch.unbind(tensor, dim=axis) return [PTNNCFTensor(t) for t in tensor_list] + @staticmethod + def squeeze(x: NNCFTensor, dim: Optional[Union[int, Tuple[int, ...]]] = None) -> NNCFTensor: + return PTNNCFTensor(torch.squeeze(x.tensor, dim=dim)) + @staticmethod def sum(tensor: NNCFTensor) -> TensorElementsType: return torch.sum(tensor.tensor).item() @staticmethod def quantile( - tensor: NNCFTensor, quantile: Union[float, List[float]], axis: Union[int, tuple, list], keepdims: bool = False + tensor: NNCFTensor, + quantile: Union[float, List[float], np.ndarray], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, ) -> List[NNCFTensor]: - raise NotImplementedError() + device = tensor.device + # See https://github.com/pytorch/pytorch/issues/61582 + if not isinstance(axis, int): + result = torch.tensor( + np.quantile(tensor.tensor.detach().cpu().numpy(), q=quantile, axis=axis, keepdims=keepdims) + ) + else: + result = torch.quantile(tensor.tensor, torch.tensor(quantile).type(tensor.tensor.dtype), axis, keepdims) + result = result.type(tensor.tensor.dtype).to(device) + return [PTNNCFTensor(x) for x in result] @classmethod - def no_outliers_map( - cls, x: NNCFTensor, fn: Callable[[NNCFTensor, Optional[int]], Any], axis: int = 0, alpha: float = 0.01 - ): - raise NotImplementedError() - - -class PTMinMaxStatisticCollector(MinMaxStatisticCollector): - def __init__( - self, use_abs_max: bool, reduction_shape: ReductionShape, output_shape: ReductionShape, num_samples: int = None - ): - super().__init__(use_abs_max, reduction_shape, num_samples) - self._output_shape = output_shape + def percentile( + cls, + tensor: NNCFTensor, + percentile: Union[float, List[float], np.ndarray], + axis: Union[int, Tuple[int, ...], List[int]], + keepdims: bool = False, + ) -> List[TensorElementsType]: + quantile = np.true_divide(percentile, 100) + return cls.quantile(tensor, quantile=quantile, axis=axis, keepdims=keepdims) @staticmethod - def _get_processor() -> NNCFCollectorTensorProcessor: - return PTNNCFCollectorTensorProcessor() - - def _register_input(self, x: torch.Tensor): - with no_nncf_trace(): - self._register_input_common(PTNNCFTensor(x)) - - def _get_statistics(self) -> PTMinMaxTensorStatistic: - min_values = self._min_values.tensor.view(self._output_shape) - max_values = self._max_values.tensor.view(self._output_shape) - return PTMinMaxTensorStatistic(min_values, max_values) - - -class PTMixedMinMaxStatisticCollector(MixedMinMaxStatisticCollector): - def __init__( - self, - use_per_sample_stats: bool, - use_abs_max: bool, - use_means_of_mins: bool, - use_means_of_maxs: bool, - reduction_shape: ReductionShape, - output_shape: ReductionShape, - num_samples: int = None, - window_size: int = None, - ): - super().__init__( - use_per_sample_stats, - use_abs_max, - use_means_of_mins, - use_means_of_maxs, - reduction_shape, - num_samples, - window_size, - ) - self._output_shape = output_shape + def sub(a: NNCFTensor, b: NNCFTensor) -> NNCFTensor: + return NNCFTensor(a.tensor - b.tensor) @staticmethod - def _get_processor() -> NNCFCollectorTensorProcessor: - return PTNNCFCollectorTensorProcessor() - - def _register_input(self, x: torch.Tensor): - with no_nncf_trace(): - self._register_input_common(PTNNCFTensor(x)) - - def _get_statistics(self) -> PTMinMaxTensorStatistic: - min_values = self._min_aggregate().tensor.view(self._output_shape) - max_values = self._max_aggregate().tensor.view(self._output_shape) - return PTMinMaxTensorStatistic(min_values, max_values) - - -class PTMeanMinMaxStatisticCollector(MeanMinMaxStatisticCollector): - def __init__( - self, - use_per_sample_stats: bool, - use_abs_max: bool, - reduction_shape: ReductionShape, - output_shape: ReductionShape, - num_samples: int = None, - window_size: int = None, - ): - super().__init__(use_per_sample_stats, use_abs_max, reduction_shape, num_samples, window_size) - self._output_shape = output_shape + def zero_elements(x: NNCFTensor) -> NNCFTensor: + pt_tensor = x.tensor + eps = torch.finfo(pt_tensor.dtype).eps + return NNCFTensor(pt_tensor.abs() < eps) - @staticmethod - def _get_processor() -> NNCFCollectorTensorProcessor: - return PTNNCFCollectorTensorProcessor() - def _register_input(self, x: torch.Tensor): - with no_nncf_trace(): - self._register_input_common(PTNNCFTensor(x)) +class PTReducerMixIn: + def _get_processor(self): + return PTNNCFCollectorTensorProcessor - def _get_statistics(self) -> PTMinMaxTensorStatistic: - min_values = self._min_aggregate().tensor.view(self._output_shape) - max_values = self._max_aggregate().tensor.view(self._output_shape) - return PTMinMaxTensorStatistic(min_values, max_values) + def get_inplace_fn(self): + return None + def get_output_names(self, target_node_name: str, port_id: int) -> List[str]: + return [] -class PTMedianMADStatisticCollector(MedianMADStatisticCollector): - def _register_input(self, x: torch.Tensor): - with no_nncf_trace(): - self._samples.append(x.detach().cpu().numpy()) - def _get_statistics(self) -> PTMedianMADTensorStatistic: - numpy_median, numpy_mad = self._prepare_statistics() - median_tensor = torch.from_numpy(numpy_median).to(dtype=torch.float) - mad_tensor = torch.from_numpy(numpy_mad).to(dtype=torch.float) +class PTNoopReducer(PTReducerMixIn, NoopReducer): + pass - median_tensor = expand_like(median_tensor, list(self._reduction_shape)) - mad_tensor = expand_like(mad_tensor, list(self._reduction_shape)) - return PTMedianMADTensorStatistic(median_tensor, mad_tensor) +class PTMinReducer(PTReducerMixIn, MinReducer): + pass -class PTPercentileStatisticCollector(PercentileStatisticCollector): - def _register_input(self, x: torch.Tensor): - with no_nncf_trace(): - self._samples.append(x.detach().cpu().numpy()) +class PTMaxReducer(PTReducerMixIn, MaxReducer): + pass - def _get_statistics(self) -> PTPercentileTensorStatistic: - percentile_vs_values_dict = self._prepare_statistics() - for key, val in percentile_vs_values_dict.items(): - torch_percentiles = torch.from_numpy(val).to(dtype=torch.float) - percentile_vs_values_dict[key] = expand_like(torch_percentiles, list(self._reduction_shape)) - return PTPercentileTensorStatistic(percentile_vs_values_dict) +class PTAbsMaxReducer(PTReducerMixIn, AbsMaxReducer): + pass -class PTMeanPercentileStatisticCollector(MeanPercentileStatisticCollector): - def _register_input(self, x: torch.Tensor): - with no_nncf_trace(): - for pct, val in self._all_pct_values.items(): - np_vals = np_percentile_reduce_like(x.cpu().numpy(), self._reduction_shape, pct) - torch_vals = torch.from_numpy(np_vals).to(dtype=torch.float) - val.append(torch_vals) - def _get_statistics(self) -> PTPercentileTensorStatistic: - mean_percentile_values = {} - for pct, val in self._all_pct_values.items(): - stacked_pct_vals = torch.stack(list(val)) - mean_percentile_values[pct] = stacked_pct_vals.mean(dim=0).view(self._reduction_shape) - return PTPercentileTensorStatistic(mean_percentile_values) +class PTMeanReducer(PTReducerMixIn, MeanReducer): + pass -class PTMeanStatisticCollector(MeanStatisticCollector): - @staticmethod - def _get_processor() -> NNCFCollectorTensorProcessor: - return PTNNCFCollectorTensorProcessor() +class PTQuantileReducer(PTReducerMixIn, QuantileReducer): + pass + + +class PTAbsQuantileReducer(PTReducerMixIn, AbsQuantileReducer): + pass + + +class PTBatchMeanReducer(PTReducerMixIn, BatchMeanReducer): + pass + + +class PTMeanPerChanelReducer(PTReducerMixIn, MeanPerChReducer): + pass + + +def _reshape_all(targets: Tuple[torch.Tensor, ...], target_shape: Tuple[int, ...]): + return map(lambda stat: torch.reshape(stat, target_shape), targets) + + +def _get_wrapped_min_max_tensor_statistic(target_shape: Tuple[int, ...]) -> Type[PTMinMaxTensorStatistic]: + """ + Returns PTMinMaxTensorStatistic type but all statistics are reshaped to target_shape. + + :param target_shape: Target shape of the tensor statistic + :return: PTMinMaxTensorStatistic type but all statistics are reshaped to target_shape. + """ - def _register_input(self, x: torch.Tensor): - with no_nncf_trace(): - self._register_input_common(PTNNCFTensor(x)) + class WrappedPTMinMaxTensorStatistic(PTMinMaxTensorStatistic): + def __init__(self, min_values, max_values): + min_values, max_values = _reshape_all((min_values, max_values), target_shape) + super().__init__(min_values, max_values) - def _get_statistics(self) -> PTMeanTensorStatistic: - return PTMeanTensorStatistic(self._mean_aggregate().tensor, self._shape()) + return WrappedPTMinMaxTensorStatistic + + +def _get_wrapped_percentile_tensor_statistic(target_shape: Tuple[int, ...]) -> Type[PTPercentileTensorStatistic]: + """ + Returns PTPercentileTensorStatistic type but all statistics are reshaped to target_shape. + + :param target_shape: Target shape of the tensor statistic + :return: PTPercentileTensorStatistic type but all statistics are reshaped to target_shape. + """ + + class WrappedPTPercentileTensorStatistic(PTPercentileTensorStatistic): + def __init__(self, percentile_vs_values_dict): + reshaped_percentiles = {} + for k, v in percentile_vs_values_dict.items(): + reshaped_percentiles[k] = torch.reshape(v, target_shape) + super().__init__(reshaped_percentiles) + + return WrappedPTPercentileTensorStatistic + + +def get_min_max_statistic_collector( + use_abs_max: bool, + reduction_axes: Tuple[int, ...], + aggregation_axes: Tuple[int, ...], + scale_shape: Tuple[int, ...], + num_samples: int, +) -> TensorCollector: + """ + Min max statistic collector builder. + + :param use_abs_max: Whether to use abs max reducer or max reducer. + :param reduction_axes: Axes to use in reduction functions. + :param aggregation_axes: Axes to use in aggregation functions. + :param scale_shape: Target shape for collected statistics. + :param num_samples: Maximum number of samples to collect. + :return: Min max statistic collector. + """ + + tensor_collector = TensorCollector(_get_wrapped_min_max_tensor_statistic(target_shape=scale_shape)) + + aggregator_kwargs = { + "tensor_processor": PTNNCFCollectorTensorProcessor, + "num_samples": num_samples, + "aggregation_axes": aggregation_axes, + } + min_reducer = PTMinReducer(reduction_axes) + min_aggregator = MinAggregator(**aggregator_kwargs) + tensor_collector.register_statistic_branch(PTMinMaxTensorStatistic.MIN_STAT, min_reducer, min_aggregator) + + max_reducer_cls = PTAbsMaxReducer if use_abs_max else PTMaxReducer + max_reducer = max_reducer_cls(reduction_axes) + max_aggregator = MaxAggregator(**aggregator_kwargs) + tensor_collector.register_statistic_branch(PTMinMaxTensorStatistic.MAX_STAT, max_reducer, max_aggregator) + return tensor_collector + + +def get_mixed_min_max_statistic_collector( + use_abs_max: bool, + reduction_axes: Tuple[int, ...], + aggregation_axes: Tuple[int, ...], + scale_shape: Tuple[int, ...], + use_means_of_mins: bool, + use_means_of_maxs: bool, + num_samples: int = None, + window_size: Optional[int] = None, +) -> TensorCollector: + """ + Mixed min max statistic collector builder. + + :param use_abs_max: Whether to use abs max reducer or max reducer. + :param reduction_axes: Axes to use in reduction functions. + :param aggregation_axes: Axes to use in aggregation functions. + :param scale_shape: Target shape for collected statistics. + :param use_means_of_mins: Whether to use mean or min aggregator for minimum statistic branch. + :param use_means_of_maxs: Whether to use mean or max aggregator for maximum statistic branch. + :param num_samples: Maximum number of samples to collect. + :param window_size: Number of samples from the end of the list of collected samples to aggregate. + Aggregates all available collected statistics in case parameter is None. + :return: Mixed min max statistic collector. + """ + tensor_collector = TensorCollector(_get_wrapped_min_max_tensor_statistic(target_shape=scale_shape)) + min_reducer = PTMinReducer(reduction_axes) + + kwargs = { + "tensor_processor": PTNNCFCollectorTensorProcessor, + "num_samples": num_samples, + "aggregation_axes": aggregation_axes, + "window_size": window_size, + } + min_aggregator_cls = MeanAggregator if use_means_of_mins else MinAggregator + min_aggregator = min_aggregator_cls(**kwargs) + tensor_collector.register_statistic_branch(PTMinMaxTensorStatistic.MIN_STAT, min_reducer, min_aggregator) + + max_reducer_cls = PTAbsMaxReducer if use_abs_max else PTMaxReducer + max_reducer = max_reducer_cls(reduction_axes) + max_aggregator_cls = MeanAggregator if use_means_of_maxs else MaxAggregator + max_aggregator = max_aggregator_cls(**kwargs) + tensor_collector.register_statistic_branch(PTMinMaxTensorStatistic.MAX_STAT, max_reducer, max_aggregator) + + return tensor_collector + + +def get_median_mad_statistic_collector( + reduction_axes: Tuple[int, ...], + aggregation_axes: Tuple[int, ...], + scale_shape: Tuple[int, ...], + num_samples: int, + window_size: Optional[int] = None, +) -> TensorCollector: + """ + Median Absolute Deviation statistic collector builder. + + :param reduction_axes: Axes to use in reduction functions. + :param aggregation_axes: Axes to use in aggregation functions. + :param scale_shape: Target shape for collected statistics. + :param num_samples: Maximum number of samples to collect. + :param window_size: Number of samples from the end of the list of collected samples to aggregate. + Aggregates all available collected statistics in case parameter is None. + :return: Median Absolute Deviation statistic collector. + + """ + + class WrappedPTMedianMADTensorStatistic(PTMedianMADTensorStatistic): + def __init__(self, median_values, mad_values): + median_values, mad_values = _reshape_all((median_values, mad_values), scale_shape) + super().__init__(median_values, mad_values) + + return _get_collection_without_reduction( + MedianAbsoluteDeviationAggregator, + WrappedPTMedianMADTensorStatistic, + reduction_axes=reduction_axes, + aggregation_axes=aggregation_axes, + num_samples=num_samples, + window_size=window_size, + ) + + +def get_percentile_tensor_collector( + percentiles_to_collect: Tuple[int, ...], + reduction_axes: Tuple[int, ...], + aggregation_axes: Tuple[int, ...], + scale_shape: Tuple[int, ...], + num_samples: int, + window_size: Optional[int] = None, +) -> TensorCollector: + """ + Percentile statistic collector builder. + + :param percentiles_to_collect: Percetiles to use on aggregation phase. + :param reduction_axes: Axes to use in reduction functions. + :param aggregation_axes: Axes to use in aggregation functions. + :param scale_shape: Target shape for collected statistics. + :param num_samples: Maximum number of samples to collect. + :param window_size: Number of samples from the end of the list of collected samples to aggregate. + Aggregates all available collected statistics in case parameter is None. + :return: Percentile statistic collector. + """ + return _get_collection_without_reduction( + partial(PercentileAggregator, percentiles_to_collect=percentiles_to_collect), + _get_wrapped_percentile_tensor_statistic(target_shape=scale_shape), + reduction_axes=reduction_axes, + aggregation_axes=aggregation_axes, + num_samples=num_samples, + window_size=window_size, + ) + + +def _get_collection_without_reduction( + aggregator_cls: TensorAggregatorBase, + statistic_cls: TensorAggregatorBase, + reduction_axes: Tuple[int, ...], + aggregation_axes: Tuple[int, ...], + num_samples: int, + window_size: Optional[int] = None, +) -> TensorCollector: + """ + Helper function to build a tensor collector which is reducing statistics exclusively during aggregation phase. + + :param aggregator_cls: Aggregator class to build the tensor collector. + :param aggregator_cls: Statistic class to build the tensor collector. + :param reduction_axes: Axes to use in reduction functions. + :param aggregation_axes: Axes to use in aggregation functions. + :param num_samples: Maximum number of samples to collect. + :param window_size: Number of samples from the end of the list of collected samples to aggregate. + Aggregates all available collected statistics in case parameter is None. + :return: Target statistic collector. + """ + tensor_collector = TensorCollector(statistic_cls) + reducer = PTNoopReducer() + aggregation_axes = list(set(list(aggregation_axes) + [dim + 1 for dim in reduction_axes])) + aggregator = aggregator_cls( + PTNNCFCollectorTensorProcessor, + aggregation_axes=aggregation_axes, + window_size=window_size, + num_samples=num_samples, + ) + + tensor_collector.register_statistic_branch( + PTMedianMADTensorStatistic.TENSOR_STATISTIC_OUTPUT_KEY, reducer, aggregator + ) + return tensor_collector + + +def get_mean_percentile_statistic_collector( + percentiles_to_collect: Tuple[int, ...], + reduction_axes: Tuple[int, ...], + aggregation_axes: Tuple[int, ...], + scale_shape: Tuple[int, ...], + num_samples: int, + window_size: Optional[int] = None, +) -> TensorCollector: + """ + Mean percentile statistic collector builder. + + :param percentiles_to_collect: Percetiles to use on reduction phase. + :param reduction_axes: Axes to use in reduction functions. + :param aggregation_axes: Axes to use in aggregation functions. + :param scale_shape: Target shape for collected statistics. + :param num_samples: Maximum number of samples to collect. + :param window_size: Number of samples from the end of the list of collected samples to aggregate. + Aggregates all available collected statistics in case parameter is None. + :return: Mean percentile statistic collector. + """ + tensor_collector = TensorCollector(_get_wrapped_percentile_tensor_statistic(target_shape=scale_shape)) + quantiles_to_collect = np.true_divide(percentiles_to_collect, 100) + reducer = PTQuantileReducer(reduction_axes=reduction_axes, quantile=quantiles_to_collect) + for output_port_id, p in enumerate(percentiles_to_collect): + aggregator = MeanAggregator( + PTNNCFCollectorTensorProcessor, + aggregation_axes=aggregation_axes, + num_samples=num_samples, + window_size=window_size, + ) + tensor_collector.register_statistic_branch( + (PTPercentileTensorStatistic.PERCENTILE_VS_VALUE_DICT, p), reducer, aggregator, output_port_id + ) + return tensor_collector + + +def get_mean_statistic_collector( + num_samples: int, channel_axis: int, window_size: Optional[int] = None +) -> TensorCollector: + """ + Mean statistic collector builder. + + :param num_samples: Maximum number of samples to collect. + :param channel_axis: Channel axis to use during reduction phase. + :param window_size: Number of samples from the end of the list of collected samples to aggregate. + Aggregates all available collected statistics in case parameter is None. + :return: Mean statistic collector. + """ + if channel_axis == 0: + reducer = PTBatchMeanReducer() + else: + reducer = PTMeanPerChanelReducer(channel_axis=channel_axis) + noop_reducer = PTNoopReducer() + + kwargs = { + "tensor_processor": PTNNCFCollectorTensorProcessor, + "num_samples": num_samples, + "window_size": window_size, + } + aggregate_mean = MeanAggregator(**kwargs) + aggregate_shape = ShapeAggregator() + + collector = TensorCollector(PTMeanTensorStatistic) + collector.register_statistic_branch(PTMeanTensorStatistic.MEAN_STAT, reducer, aggregate_mean) + collector.register_statistic_branch(PTMeanTensorStatistic.SHAPE_STAT, noop_reducer, aggregate_shape) + return collector + + +PT_REDUCERS_MAP = { + StatisticsType.MIN: PTMinReducer, + StatisticsType.MAX: PTMaxReducer, + StatisticsType.ABS_MAX: PTAbsMaxReducer, + StatisticsType.MEAN: PTMeanReducer, + StatisticsType.QUANTILE: PTQuantileReducer, + StatisticsType.ABS_QUANTILE: PTAbsQuantileReducer, +} diff --git a/nncf/torch/tensor_statistics/statistics.py b/nncf/torch/tensor_statistics/statistics.py index 7a251b19207..ba51df16ce9 100644 --- a/nncf/torch/tensor_statistics/statistics.py +++ b/nncf/torch/tensor_statistics/statistics.py @@ -9,6 +9,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from typing import Optional, Tuple + import torch from nncf.common.tensor_statistics.statistics import MeanTensorStatistic @@ -18,6 +20,10 @@ from nncf.common.tensor_statistics.statistics import TensorStatistic +def _reshape_all(targets: Tuple[torch.Tensor, ...], target_shape: Tuple[int, ...]): + return map(lambda stat: torch.reshape(stat, target_shape), targets) + + class PTMinMaxTensorStatistic(MinMaxTensorStatistic): @staticmethod def tensor_eq(tensor1: torch.Tensor, tensor2: torch.Tensor, rtol=1e-6) -> bool: @@ -49,8 +55,8 @@ def pt_convert_stat_to_min_max_tensor_stat(statistic: TensorStatistic) -> PTMinM # Using three-sigma approach to estimate min and max # Constant factor depends on the distribution form - assuming normal and the factor is 1.4826 return PTMinMaxTensorStatistic( - statistic.median_values - 3 * 1.4826230 * statistic.mad_values, - statistic.median_values + 3 * 1.4826230 * statistic.mad_values, + min_values=statistic.median_values - 3 * 1.4826230 * statistic.mad_values, + max_values=statistic.median_values + 3 * 1.4826230 * statistic.mad_values, ) if isinstance(statistic, PTPercentileTensorStatistic): if len(statistic.percentile_vs_values_dict.keys()) < 2: @@ -58,6 +64,7 @@ def pt_convert_stat_to_min_max_tensor_stat(statistic: TensorStatistic) -> PTMinM min_pct = min(statistic.percentile_vs_values_dict.keys()) max_pct = max(statistic.percentile_vs_values_dict.keys()) return PTMinMaxTensorStatistic( - statistic.percentile_vs_values_dict[min_pct], statistic.percentile_vs_values_dict[max_pct] + min_values=statistic.percentile_vs_values_dict[min_pct], + max_values=statistic.percentile_vs_values_dict[max_pct], ) raise ValueError("Unknown TensorStatistic to generate min-max stat from!") diff --git a/tests/common/experimental/test_reducers_and_aggregators.py b/tests/common/experimental/test_reducers_and_aggregators.py index cc54fe987ac..e307fb73f84 100644 --- a/tests/common/experimental/test_reducers_and_aggregators.py +++ b/tests/common/experimental/test_reducers_and_aggregators.py @@ -10,18 +10,26 @@ # limitations under the License. from abc import abstractmethod +from dataclasses import dataclass +from functools import partial from itertools import product +from typing import Any, List, Optional, Tuple import numpy as np import pytest +from nncf.common.graph.layer_attributes import Dtype +from nncf.common.tensor_statistics.collectors import NNCFCollectorTensorProcessor +from nncf.experimental.common.tensor_statistics.collectors import AggregationAxes from nncf.experimental.common.tensor_statistics.collectors import MaxAggregator from nncf.experimental.common.tensor_statistics.collectors import MeanAggregator from nncf.experimental.common.tensor_statistics.collectors import MeanNoOutliersAggregator +from nncf.experimental.common.tensor_statistics.collectors import MedianAbsoluteDeviationAggregator from nncf.experimental.common.tensor_statistics.collectors import MedianAggregator from nncf.experimental.common.tensor_statistics.collectors import MedianNoOutliersAggregator from nncf.experimental.common.tensor_statistics.collectors import MinAggregator from nncf.experimental.common.tensor_statistics.collectors import NoopAggregator +from nncf.experimental.common.tensor_statistics.collectors import PercentileAggregator from nncf.experimental.common.tensor_statistics.collectors import ShapeAggregator DEFALUT_3D_MEAN_VALUE = [[2503.125, -2493.75, 5009.375], [-4987.5, 7515.625, -7481.25], [10021.875, -9975.0, 12528.125]] @@ -43,17 +51,82 @@ default_test_quantile = 0.1 -def default_test_mean_no_outlier(tp, ps): - return MeanNoOutliersAggregator(tp, ps, quantile=default_test_quantile) +@dataclass +class OfflineAggregatorTestCase: + aggregation_axes: Optional[AggregationAxes] + min_ref: np.ndarray + max_ref: np.ndarray + + +OFFLINE_AGGREGATORS_TEST_CASES = [ + OfflineAggregatorTestCase( + aggregation_axes=None, + min_ref=np.array([[[-50000, -4, -8], [-12, -16, -20], [-24, -28, -32]]]), + max_ref=np.array([[[50000, 4, 8], [12, 16, 20], [24, 28, 32]]]), + ), + OfflineAggregatorTestCase( + aggregation_axes=(0,), + min_ref=np.array([[[-50000, -4, -8], [-12, -16, -20], [-24, -28, -32]]]), + max_ref=np.array([[[50000, 4, 8], [12, 16, 20], [24, 28, 32]]]), + ), + OfflineAggregatorTestCase( + aggregation_axes=(0, 2), + min_ref=np.array([[-50000, -28, -32]]), + max_ref=np.array([[50000, 28, 32]]), + ), + OfflineAggregatorTestCase( + aggregation_axes=(2,), + min_ref=np.array( + [ + [[-50000, 5, 10]], + [[-40000, 4, 8]], + [[-30000, 3, 6]], + [[-20000, 2, 4]], + [[-10000, 1, 2]], + [[0, 0, 0]], + [[-6, -7, -8]], + [[-12, -14, -16]], + [[-18, -21, -24]], + [[-24, -28, -32]], + ] + ), + max_ref=np.array( + [ + [[50000, -5, -10]], + [[40000, -4, -8]], + [[30000, -3, -6]], + [[20000, -2, -4]], + [[10000, -1, -2]], + [[0, 0, 0]], + [[6, 7, 8]], + [[12, 14, 16]], + [[18, 21, 24]], + [[24, 28, 32]], + ] + ), + ), +] -def default_test_median_no_outlier(tp, ps): - return MedianNoOutliersAggregator(tp, ps, quantile=default_test_quantile) +def default_test_mean_no_outlier(tensor_processor, aggregation_axes): + return MeanNoOutliersAggregator( + tensor_processor=tensor_processor, + aggregation_axes=aggregation_axes, + quantile=default_test_quantile, + ) + + +def default_test_median_no_outlier(tensor_processor, aggregation_axes): + return MedianNoOutliersAggregator( + tensor_processor=tensor_processor, + aggregation_axes=aggregation_axes, + quantile=default_test_quantile, + ) class TemplateTestReducersAggreagtors: @abstractmethod - def get_nncf_tensor(self, x: np.array): + def get_nncf_tensor(self, x: np.array, dtype: Optional[Dtype] = None): pass @pytest.fixture @@ -70,6 +143,18 @@ def reducers(self): def all_close(self, val, ref) -> bool: pass + @abstractmethod + def squeeze_tensor(self, ref_tensor: List[Any], axes: Optional[Tuple[int]] = None): + pass + + @abstractmethod + def cast_tensor(self, tensor, dtype: Dtype): + pass + + @abstractmethod + def expand_dims(self, tensor, dims: Tuple[int, ...]): + pass + def test_noop_reducer(self, reducers): reducer = reducers["noop"]() input_ = np.arange(24).reshape((1, 2, 3, 4)) @@ -87,27 +172,27 @@ def test_noop_reducer(self, reducers): ], ) def test_min_max_mean_reducers(self, reducer_name, ref, reducers): - reduction_shape = (1, 2) + reduction_axes = (1, 2) input_ = np.arange(-26, 10).reshape((4, 3, 3)) - for i, red_shape in enumerate([reduction_shape, None]): - reducer = reducers[reducer_name](red_shape, False) - val = reducer([self.get_nncf_tensor(input_)]) + for i, reduction_axes_ in enumerate([reduction_axes, None]): + reducer = reducers[reducer_name](reduction_axes=reduction_axes_, inplace=False) + val = reducer([self.get_nncf_tensor(input_, Dtype.FLOAT)]) assert len(val) == 1 - assert self.all_close(val[0].tensor, ref[i]) + assert self.all_close(val[0].tensor, self.cast_tensor(ref[i], Dtype.FLOAT)) @pytest.mark.parametrize( "reducer_name,ref", [("quantile", ([[[[-20000]]]], [[[[10000]]]])), ("abs_quantile", ([[[[20000]]]],))] ) def test_quantile_reducers(self, reducer_name, ref, reducers): - reduction_shape = (1, 2, 3) + reduction_axes = (1, 2, 3) input_ = np.arange(-26, 10).reshape((1, 4, 3, 3)) input_[0][0][0] = -20000 input_[0][0][1] = 10000 - reducer = reducers[reducer_name](reduction_shape, inplace=False) - val = reducer([self.get_nncf_tensor(input_)]) + reducer = reducers[reducer_name](reduction_axes=reduction_axes, inplace=False) + val = reducer([self.get_nncf_tensor(input_, dtype=Dtype.FLOAT)]) assert len(val) == len(ref) for i, ref_ in enumerate(ref): - assert self.all_close(val[i].tensor, ref_) + assert self.all_close(val[i].tensor, self.cast_tensor(ref_, Dtype.FLOAT)) @pytest.mark.parametrize( "reducer_name,ref", @@ -116,9 +201,9 @@ def test_quantile_reducers(self, reducer_name, ref, reducers): def test_batch_mean_mean_per_ch_reducers(self, reducer_name, ref, reducers): input_ = np.arange(-26, 10).reshape((4, 1, 3, 3)) reducer = reducers[reducer_name](inplace=False) - val = reducer([self.get_nncf_tensor(input_)]) + val = reducer([self.get_nncf_tensor(input_, Dtype.FLOAT)]) assert len(val) == 1 - assert self.all_close(val[0].tensor, ref) + assert self.all_close(val[0].tensor, self.cast_tensor(ref, Dtype.FLOAT)) def test_noop_aggregator(self): aggregator = NoopAggregator(None) @@ -146,24 +231,33 @@ def test_shape_aggregator(self): assert aggregator._collected_samples == 1 assert ref_shape == aggregator.aggregate() - def test_min_max_aggregators(self, tensor_processor): - min_aggregator = MinAggregator(tensor_processor) - max_aggregator = MaxAggregator(tensor_processor) + @pytest.mark.parametrize( + "offline_aggregators_test_desc", + OFFLINE_AGGREGATORS_TEST_CASES, + ) + def test_min_max_aggregators( + self, offline_aggregators_test_desc: OfflineAggregatorTestCase, tensor_processor: NNCFCollectorTensorProcessor + ): + aggregation_axes = offline_aggregators_test_desc.aggregation_axes + min_aggregator = MinAggregator(tensor_processor=tensor_processor, aggregation_axes=aggregation_axes) + max_aggregator = MaxAggregator(tensor_processor=tensor_processor, aggregation_axes=aggregation_axes) input_ = np.arange(3 * 3).reshape((1, 3, 3)) input_[0, 0, 0] = -10000 for i in range(-5, 5): min_aggregator.register_reduced_input(self.get_nncf_tensor(input_ * (-i))) max_aggregator.register_reduced_input(self.get_nncf_tensor(input_ * i)) - min_ref = [[[-50000, -4, -8], [-12, -16, -20], [-24, -28, -32]]] - assert self.all_close(min_ref, min_aggregator.aggregate()) - - max_ref = [[[50000, 4, 8], [12, 16, 20], [24, 28, 32]]] - assert self.all_close(max_ref, max_aggregator.aggregate()) + min_ref = offline_aggregators_test_desc.min_ref + max_ref = offline_aggregators_test_desc.max_ref + assert self.all_close( + min_aggregator.aggregate(), + min_ref, + ) + assert self.all_close(max_aggregator.aggregate(), max_ref) NO_OUTLIERS_TEST_PARAMS = [ (MeanAggregator, True, 1, 1404.5138888888905), - (MedianAggregator, True, 1, 15.5), + (MedianAggregator, True, 1, 24.0), ( MeanAggregator, False, @@ -179,8 +273,8 @@ def test_min_max_aggregators(self, tensor_processor): (MedianAggregator, True, 3, DEFALUT_3D_MEDIAN_VALUE), (MeanAggregator, False, 3, [DEFALUT_3D_MEAN_VALUE]), (MedianAggregator, False, 3, [DEFALUT_3D_MEDIAN_VALUE]), - (default_test_mean_no_outlier, True, 1, 1404.5138888888905), - (default_test_median_no_outlier, True, 1, 15.5), + (default_test_mean_no_outlier, True, 1, 20.0893), + (default_test_median_no_outlier, True, 1, 30.0), ( default_test_mean_no_outlier, False, @@ -211,19 +305,86 @@ def test_mean_median_agggregators(self, aggregator_cls, refs, tensor_processor, input_ = input_.reshape((1, 3, 3)) input_with_outliers = input_with_outliers.reshape((1, 3, 3)) - aggregator = aggregator_cls(tensor_processor, use_per_sample_stats) + aggregation_axes = (0, 1) if use_per_sample_stats else (0,) + aggregator = aggregator_cls(tensor_processor=tensor_processor, aggregation_axes=aggregation_axes) for i in range(1, 6): - aggregator.register_reduced_input(self.get_nncf_tensor(input_ * i)) + aggregator.register_reduced_input(self.get_nncf_tensor(input_ * i, Dtype.FLOAT)) # this registration is to make diff between mean and median bigger - aggregator.register_reduced_input(self.get_nncf_tensor(input_ * 10)) + aggregator.register_reduced_input(self.get_nncf_tensor(input_ * 10, Dtype.FLOAT)) is_median = isinstance(aggregator, (MedianAggregator, MedianNoOutliersAggregator)) # Outliers registration for i in range(2): # mult is needed to make outlier and no outlier aggreagators differs mult = 2.2 * i - 1 if not is_median else 1 - aggregator.register_reduced_input(self.get_nncf_tensor(input_with_outliers * mult)) + aggregator.register_reduced_input(self.get_nncf_tensor(input_with_outliers * mult, Dtype.FLOAT)) + if is_median and dims == 1 and use_per_sample_stats: + # To make no outliers and outliers versions return different output + aggregator.register_reduced_input( + self.get_nncf_tensor(np.full(input_with_outliers.shape, input_with_outliers[-1]), Dtype.FLOAT) + ) + ret_val = aggregator.aggregate() + + assert self.all_close(ret_val, self.cast_tensor(refs, Dtype.FLOAT)) + + REF_MAD_PERCENTILE_REF_VALUES = { + MedianAbsoluteDeviationAggregator: { + None: { + "median_values": np.array([4.5, 9.0, 13.5, 18.0, 22.5, 27.0, 31.5, 36.0, 40.5]), + "mad_values": np.array([2.5, 5.0, 7.5, 10.0, 12.5, 15.0, 17.5, 20.0, 22.5]), + }, + (0,): { + "median_values": np.array([4.5, 9.0, 13.5, 18.0, 22.5, 27.0, 31.5, 36.0, 40.5]), + "mad_values": np.array([2.5, 5.0, 7.5, 10.0, 12.5, 15.0, 17.5, 20.0, 22.5]), + }, + (0, 1): { + "median_values": np.array(18.0), + "mad_values": np.array(12.0), + }, + }, + PercentileAggregator: { + None: { + 5: np.array([0.4, 0.8, 1.2, 1.6, 2.0, 2.4, 2.8, 3.2, 3.6]), + 10: np.array([0.8, 1.6, 2.4, 3.2, 4.0, 4.8, 5.6, 6.4, 7.2]), + 90: np.array([7.2, 14.4, 21.6, 28.8, 36.0, 43.2, 50.4, 57.6, 64.8]), + 95: np.array([7.6, 15.2, 22.8, 30.4, 38.0, 45.6, 53.2, 60.8, 68.4]), + }, + (0,): { + 5: np.array([0.4, 0.8, 1.2, 1.6, 2.0, 2.4, 2.8, 3.2, 3.6]), + 10: np.array([0.8, 1.6, 2.4, 3.2, 4.0, 4.8, 5.6, 6.4, 7.2]), + 90: np.array([7.2, 14.4, 21.6, 28.8, 36.0, 43.2, 50.4, 57.6, 64.8]), + 95: np.array([7.6, 15.2, 22.8, 30.4, 38.0, 45.6, 53.2, 60.8, 68.4]), + }, + (0, 1): { + 5: np.array(0.0), + 10: np.array(0.0), + 90: np.array(48.0), + 95: np.array(56.0), + }, + }, + } + + @pytest.mark.parametrize( + "aggregator_cls", + [ + MedianAbsoluteDeviationAggregator, + partial( + PercentileAggregator, + percentiles_to_collect=[5, 10, 90, 95], + ), + ], + ) + @pytest.mark.parametrize("aggregation_axes", [None, (0,), (0, 1)]) + def test_mad_percentile_aggregators(self, aggregator_cls, tensor_processor, aggregation_axes): + aggregator = aggregator_cls(tensor_processor=tensor_processor, aggregation_axes=aggregation_axes) + input_ = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9], dtype=np.float32) + for i in range(9): + aggregator.register_reduced_input(self.get_nncf_tensor(input_ * i, Dtype.FLOAT)) + ret_val = aggregator.aggregate() - assert self.all_close(ret_val, refs) + ref_values = self.REF_MAD_PERCENTILE_REF_VALUES[aggregator.__class__][aggregation_axes] + assert len(ret_val) == len(ref_values) + for k, v in ref_values.items(): + assert self.all_close(ret_val[k], self.cast_tensor(v, Dtype.FLOAT)) @pytest.mark.parametrize( "reducer_name", @@ -240,16 +401,16 @@ def test_reducers_name_hash_equal(self, reducer_name, reducers): params = {} if reducer_name in ["min", "max", "abs_max", "mean"]: - params["reduction_shape"] = [None, (0, 1, 3), (1, 2, 3)] + params["reduction_axes"] = [None, (0, 1, 3), (1, 2, 3)] params["inplace"] = [False, True] elif reducer_name in ["quantile", "abs_quantile"]: - params["reduction_shape"] = [None, (0, 1, 3), (1, 2, 3)] + params["reduction_axes"] = [None, (0, 1, 3), (1, 2, 3)] params["quantile"] = [[0.01, 0.99], [0.001, 0.999]] elif reducer_name == "batch_mean": params["inplace"] = [False, True] elif reducer_name == "mean_per_ch": params["inplace"] = [False, True] - params["channel_dim"] = [1, 2] + params["channel_axis"] = [1, 2] else: raise RuntimeError( "test_min_max_mean_reducer_hash_equal configurated in a wrong way." diff --git a/tests/common/experimental/test_statistic_collector.py b/tests/common/experimental/test_statistic_collector.py index 9346b176bd6..ba99c65d328 100644 --- a/tests/common/experimental/test_statistic_collector.py +++ b/tests/common/experimental/test_statistic_collector.py @@ -10,20 +10,41 @@ # limitations under the License. from abc import abstractmethod -from typing import List, Optional +from typing import List, Optional, Type import numpy as np import pytest from nncf.common.tensor import NNCFTensor +from nncf.common.tensor_statistics.statistics import MeanTensorStatistic +from nncf.common.tensor_statistics.statistics import MedianMADTensorStatistic +from nncf.common.tensor_statistics.statistics import MinMaxTensorStatistic +from nncf.common.tensor_statistics.statistics import PercentileTensorStatistic +from nncf.common.tensor_statistics.statistics import RawTensorStatistic +from nncf.experimental.common.tensor_statistics.collectors import AggregatorBase from nncf.experimental.common.tensor_statistics.collectors import MergedTensorCollector -from nncf.experimental.common.tensor_statistics.collectors import TensorAggregatorBase from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.experimental.common.tensor_statistics.collectors import TensorReducerBase from nncf.experimental.common.tensor_statistics.collectors import TensorType - # pylint: disable=(protected-access) + + +class NumpyNNCFTensor(NNCFTensor): + def __init__(self, tensor: np.array, dummy_device: Optional[str] = None): + # In case somebody attempts to wrap + # tensor twice + if isinstance(tensor, self.__class__): + tensor = tensor.tensor + + super().__init__(tensor) + self.dummy_device = dummy_device + + @property + def device(self) -> Optional[str]: + return self.dummy_device + + class DummyTensorReducer(TensorReducerBase): def __init__(self, output_name: str, inplace: bool = False, inplace_mock=None): super().__init__(inplace=inplace) @@ -47,9 +68,9 @@ class DummyTensorReducerA(DummyTensorReducer): pass -class DummyTensorAggregator(TensorAggregatorBase): - def __init__(self, num_samples: Optional[int]): - super().__init__(None, num_samples) +class DummyTensorAggregator(AggregatorBase): + def __init__(self, num_samples: Optional[int] = None): + super().__init__(None, num_samples=num_samples) def _register_reduced_input_impl(self, x: TensorType): return self._container.append(x) @@ -69,7 +90,7 @@ def test_aggregator_enabled_and_reset(): collector.register_statistic_branch("A", reducer, aggregator) input_name = "input_name" inputs = TensorCollector.get_tensor_collector_inputs( - {input_name: NNCFTensor(np.array(100))}, [(hash(reducer), [input_name])] + {input_name: NumpyNNCFTensor(np.array(100))}, [(hash(reducer), [input_name])] ) for _ in range(3): @@ -125,7 +146,11 @@ def test_duplicated_statistics_are_merged(): [(hash(reducer_inplace), ["Dummy_inplace"]), (hash(reducer_a), ["A"]), (hash(reducer), ["Dummy"])] ) - outputs = {"Dummy": NNCFTensor(np.array(5)), "A": NNCFTensor(np.array(0)), "Dummy_inplace": NNCFTensor(np.array(6))} + outputs = { + "Dummy": NumpyNNCFTensor(np.array(5)), + "A": NumpyNNCFTensor(np.array(0)), + "Dummy_inplace": NumpyNNCFTensor(np.array(6)), + } target_inputs = TensorCollector.get_tensor_collector_inputs(outputs, output_info) collector.register_inputs(target_inputs) @@ -142,10 +167,10 @@ def test_duplicated_statistics_are_merged(): # Check aggregators recieved correct inputs assert len(statistics) == 6 for k in "ABC": - assert statistics[k] == NNCFTensor(np.array(5)) - assert statistics["D"] == NNCFTensor(np.array(5)) - assert statistics["E"] == NNCFTensor(np.array(0)) - assert statistics["F"] == NNCFTensor(np.array(6)) + assert statistics[k] == NumpyNNCFTensor(np.array(5)) + assert statistics["D"] == NumpyNNCFTensor(np.array(5)) + assert statistics["E"] == NumpyNNCFTensor(np.array(0)) + assert statistics["F"] == NumpyNNCFTensor(np.array(6)) def test_inplace_param(): @@ -192,8 +217,8 @@ def test_merged_tensor_collector(): assert collector.aggregators[common_branch_key] is common_aggregator output_info = merged_collector.get_output_info(None, None) - outputs = {"common_input": NNCFTensor(np.array(0))} - outputs.update({f"input_{idx + 1}": NNCFTensor(np.array(idx + 1)) for idx, _ in enumerate(collectors[:-1])}) + outputs = {"common_input": NumpyNNCFTensor(np.array(0))} + outputs.update({f"input_{idx + 1}": NumpyNNCFTensor(np.array(idx + 1)) for idx, _ in enumerate(collectors[:-1])}) target_inputs = TensorCollector.get_tensor_collector_inputs(outputs, output_info) merged_collector.register_inputs(target_inputs) @@ -204,8 +229,8 @@ def test_merged_tensor_collector(): statistic = collector.get_statistics() assert len(statistic) == 2 - assert statistic["common"] == NNCFTensor(np.array(0)) - assert statistic["unique"] == NNCFTensor(np.array(idx + 1)) + assert statistic["common"] == NumpyNNCFTensor(np.array(0)) + assert statistic["unique"] == NumpyNNCFTensor(np.array(idx + 1)) def test_ambigous_container_key(): @@ -257,7 +282,7 @@ def test_multiple_branch_reducer(): ], ) ] - inputs = {name: NNCFTensor(np.array(i)) for i, name in enumerate(ref_output_info[0][1])} + inputs = {name: NumpyNNCFTensor(np.array(i)) for i, name in enumerate(ref_output_info[0][1])} output_info = collector.get_output_info(target_node_name, 0) assert output_info == ref_output_info @@ -265,18 +290,75 @@ def test_multiple_branch_reducer(): target_inputs = collector.get_tensor_collector_inputs(inputs, output_info) collector.register_inputs(target_inputs) - ref_stats = {"0": NNCFTensor(np.array(0)), "1": NNCFTensor(np.array(1))} + ref_stats = {"0": NumpyNNCFTensor(np.array(0)), "1": NumpyNNCFTensor(np.array(1))} stats = collector.get_statistics() assert len(ref_stats) == len(stats) for key, value in ref_stats.items(): assert value == stats[key] +def test_register_unnamed_statistics(mocker): + tensor_collector = TensorCollector() + reducer_hashes = [] + for reducer_cls, key in zip([DummyTensorReducer, DummyTensorReducerA], "AB"): + reducer = reducer_cls(f"Dummy{key}") + tensor_collector.register_statistic_branch(key, reducer, DummyTensorAggregator(None)) + reducer_hashes.append(hash(reducer)) + + tensor_collector.register_inputs = mocker.MagicMock() + inputs_ = NumpyNNCFTensor(np.ones(5)) + tensor_collector.register_input_for_all_reducers(inputs_) + + tensor_collector.register_inputs.assert_called_once() + args = tensor_collector.register_inputs.call_args[0][0] + assert len(args) == 2 + for k, v in args.items(): + assert k in reducer_hashes + assert len(v) == 1 + assert all(v[0] == inputs_) + + +def test_wrong_statistic_container_class(): + class BadStatContainer: + pass + + tensor_collector = TensorCollector(BadStatContainer) + tensor_collector.register_statistic_branch("A", DummyTensorReducer("A"), DummyTensorAggregator()) + tensor_collector.register_input_for_all_reducers(NumpyNNCFTensor(1)) + with pytest.raises(RuntimeError): + tensor_collector.get_statistics() + + class TemplateTestStatisticCollector: @abstractmethod def get_nncf_tensor_cls(self): pass + @abstractmethod + @pytest.fixture + def min_max_statistic_cls(self) -> Type[MinMaxTensorStatistic]: + pass + + @abstractmethod + @pytest.fixture + def mean_statistic_cls(self) -> Type[MeanTensorStatistic]: + pass + + @abstractmethod + @pytest.fixture + def median_mad_statistic_cls(self) -> Type[MedianMADTensorStatistic]: + pass + + @abstractmethod + @pytest.fixture + def percentile_statistic_cls(self) -> Type[PercentileTensorStatistic]: + pass + + @abstractmethod + @pytest.fixture + def raw_statistic_cls(self) -> Type[RawTensorStatistic]: + pass + @pytest.mark.parametrize("inplace", [False, True]) @pytest.mark.parametrize("any_not_empty", [False, True]) def test_empty_tensors_register(self, inplace, any_not_empty): @@ -313,3 +395,89 @@ def test_empty_tensors_register(self, inplace, any_not_empty): stats = collector.get_statistics() assert len(stats) == 1 assert stats["A"] is None + + def test_min_max_stat_building(self, min_max_statistic_cls: MinMaxTensorStatistic): + tensor_collector = TensorCollector(min_max_statistic_cls) + tensor_collector.register_statistic_branch( + min_max_statistic_cls.MIN_STAT, DummyTensorReducer("A"), DummyTensorAggregator() + ) + tensor_collector.register_statistic_branch( + min_max_statistic_cls.MAX_STAT, DummyTensorReducer("B"), DummyTensorAggregator() + ) + tensor_collector.register_input_for_all_reducers(NumpyNNCFTensor(1)) + statistic = tensor_collector.get_statistics() + assert isinstance(statistic, MinMaxTensorStatistic) + assert statistic.min_values == statistic.max_values == NumpyNNCFTensor(1) + + def test_mean_max_stat_building(self, mean_statistic_cls: MeanTensorStatistic): + tensor_collector = TensorCollector(mean_statistic_cls) + tensor_collector.register_statistic_branch( + mean_statistic_cls.MEAN_STAT, DummyTensorReducer("A"), DummyTensorAggregator() + ) + tensor_collector.register_statistic_branch( + mean_statistic_cls.SHAPE_STAT, DummyTensorReducer("B"), DummyTensorAggregator() + ) + tensor_collector.register_input_for_all_reducers(NumpyNNCFTensor(1)) + statistic = tensor_collector.get_statistics() + assert isinstance(statistic, MeanTensorStatistic) + assert statistic.mean_values == statistic.shape == NumpyNNCFTensor(1) + + def test_median_mad_stat_building(self, median_mad_statistic_cls: MedianMADTensorStatistic): + class DummyMADPercentileAggregator(DummyTensorAggregator): + def _aggregate_impl(self): + return { + MedianMADTensorStatistic.MEDIAN_VALUES_STAT: self._container[0], + MedianMADTensorStatistic.MAD_VALUES_STAT: self._container[0], + } + + tensor_collector = TensorCollector(median_mad_statistic_cls) + tensor_collector.register_statistic_branch( + median_mad_statistic_cls.TENSOR_STATISTIC_OUTPUT_KEY, + DummyTensorReducer("A"), + DummyMADPercentileAggregator(), + ) + tensor_collector.register_input_for_all_reducers(NumpyNNCFTensor(1)) + statistic = tensor_collector.get_statistics() + assert isinstance(statistic, MedianMADTensorStatistic) + assert statistic.median_values == statistic.mad_values == NumpyNNCFTensor(1) + + def test_percentile_max_stat_building(self, percentile_statistic_cls: PercentileTensorStatistic): + class DummyPercentileTensorAggregator(DummyTensorAggregator): + def _aggregate_impl(self): + return {0.5: self._container[0]} + + tensor_collector = TensorCollector(percentile_statistic_cls) + tensor_collector.register_statistic_branch( + percentile_statistic_cls.TENSOR_STATISTIC_OUTPUT_KEY, + DummyTensorReducer("A"), + DummyPercentileTensorAggregator(), + ) + tensor_collector.register_input_for_all_reducers(NumpyNNCFTensor(1)) + statistic = tensor_collector.get_statistics() + assert isinstance(statistic, PercentileTensorStatistic) + assert statistic.percentile_vs_values_dict[0.5] == NumpyNNCFTensor(1) + + tensor_collector = TensorCollector(percentile_statistic_cls) + qs = [0.3, 0.5, 0.7] + for q in qs: + tensor_collector.register_statistic_branch( + (PercentileTensorStatistic.PERCENTILE_VS_VALUE_DICT, q), + DummyTensorReducer(f"A{q}"), + DummyTensorAggregator(), + ) + tensor_collector.register_input_for_all_reducers(NumpyNNCFTensor(1)) + statistic = tensor_collector.get_statistics() + assert isinstance(statistic, PercentileTensorStatistic) + assert len(statistic.percentile_vs_values_dict) == len(qs) + for q in qs: + assert statistic.percentile_vs_values_dict[q] == NumpyNNCFTensor(1) + + def test_raw_max_stat_building(self, raw_statistic_cls: RawTensorStatistic): + tensor_collector = TensorCollector(raw_statistic_cls) + tensor_collector.register_statistic_branch( + raw_statistic_cls.VALUES_STATS, DummyTensorReducer("A"), DummyTensorAggregator() + ) + tensor_collector.register_input_for_all_reducers(NumpyNNCFTensor(1)) + statistic = tensor_collector.get_statistics() + assert isinstance(statistic, RawTensorStatistic) + assert statistic.values == NNCFTensor(1) diff --git a/tests/common/test_statistics_aggregator.py b/tests/common/test_statistics_aggregator.py index cd0545a4580..6c5fc043904 100644 --- a/tests/common/test_statistics_aggregator.py +++ b/tests/common/test_statistics_aggregator.py @@ -98,7 +98,10 @@ def inplace_statistics(self) -> bool: @abstractmethod @pytest.fixture def is_backend_support_custom_estimators(self) -> bool: - pass + """ + False if backend can initialize only following tensor collectors: + MinMax, MeanMinMax. + """ @abstractmethod def reducers_map(self) -> List[TensorReducerBase]: @@ -375,6 +378,7 @@ def test_statistics_aggregator_min_max( inplace_statistics, is_backend_support_custom_estimators, ): + inplace_statistics = False model = self.get_backend_model(dataset_samples) quantizer_config = QuantizerConfig( mode=test_parameters.quantization_mode, per_channel=test_parameters.per_channel @@ -433,6 +437,10 @@ def filter_func(point): if isinstance(ref_min_val, np.ndarray): assert stat.min_values.shape == ref_min_val.shape assert stat.max_values.shape == ref_max_val.shape + else: + ref_shape = (1, 1, 1, 1) if is_stat_in_shape_of_scale else () + assert stat.min_values.shape == ref_shape + assert stat.max_values.shape == ref_shape class BiasCorrectionAlgos(Enum): BIAS_CORRECTION = "bias_correction" @@ -811,10 +819,10 @@ def test_same_collectors_different_attrs_dont_merge(self, statistics_type, test_ model = params["model"](dataset_samples) params = {} if statistics_type in [StatisticsType.MIN, StatisticsType.MAX, StatisticsType.ABS_MAX, StatisticsType.MEAN]: - params["reduction_shape"] = [None, (0, 1, 3), (1, 2, 3)] + params["reduction_axes"] = [None, (0, 1, 3), (1, 2, 3)] params["inplace"] = [False, True] elif statistics_type in [StatisticsType.QUANTILE, StatisticsType.ABS_QUANTILE]: - params["reduction_shape"] = [None, (0, 1, 3), (1, 2, 3)] + params["reduction_axes"] = [None, (0, 1, 3), (1, 2, 3)] params["quantile"] = [[0.01, 0.99], [0.001, 0.999]] elif statistics_type == "batch_mean": pytest.skip("Inplace statistic woun't work until openvino==2023.0.0 release") @@ -822,7 +830,7 @@ def test_same_collectors_different_attrs_dont_merge(self, statistics_type, test_ elif statistics_type == "mean_per_ch": # TODO(dlyakhov) uncoment when nncf will switch to openvino==2023.0.0 # params["inplace"] = [False, True] - params["channel_dim"] = [1, 2] + params["channel_axis"] = [1, 2] def product_dict(**kwargs): keys = kwargs.keys() @@ -844,7 +852,7 @@ def product_dict(**kwargs): dataset = self.get_dataset(dataset_samples) statistics_aggregator = self.get_statistics_aggregator(dataset) statistics_aggregator.register_statistic_points(statistics_points) - # Run statistic collection to check output names matches reduer names + # Run statistic collection to check output names matches reducer names graph = NNCFGraphFactory.create(model) statistics_aggregator.collect_statistics(model, graph) diff --git a/tests/onnx/quantization/test_quantizer_config.py b/tests/onnx/quantization/test_quantizer_config.py index 374ae440f13..8702f4a42f0 100644 --- a/tests/onnx/quantization/test_quantizer_config.py +++ b/tests/onnx/quantization/test_quantizer_config.py @@ -12,6 +12,8 @@ import pytest from nncf.common.graph.transformations.commands import TargetType +from nncf.common.tensor_statistics.collectors import ReductionAxes +from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXAddLayerMetatype from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXConvolutionMetatype from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXDepthwiseConvolutionMetatype @@ -27,6 +29,9 @@ ParamsCls = TemplateTestQuantizerConfig.TestGetStatisticsCollectorParameters +# pylint: disable=protected-access + + class TestQuantizerConfig(TemplateTestQuantizerConfig): def get_algo_backend(self): return ONNXMinMaxAlgoBackend() @@ -37,6 +42,9 @@ def check_is_min_max_statistic_collector(self, tensor_collector): def check_is_mean_min_max_statistic_collector(self, tensor_collector): assert isinstance(tensor_collector, ONNXMeanMinMaxStatisticCollector) + def get_reduction_axes(self, reducer: TensorStatisticCollectorBase) -> ReductionAxes: + return reducer._reduction_shape + @pytest.fixture( params=[ pytest.param( diff --git a/tests/openvino/native/quantization/test_fq_configurations.py b/tests/openvino/native/quantization/test_calculation_quantizer_params.py similarity index 100% rename from tests/openvino/native/quantization/test_fq_configurations.py rename to tests/openvino/native/quantization/test_calculation_quantizer_params.py diff --git a/tests/openvino/native/quantization/test_quantizer_config.py b/tests/openvino/native/quantization/test_quantizer_config.py index 45d41644ba4..773f5996604 100644 --- a/tests/openvino/native/quantization/test_quantizer_config.py +++ b/tests/openvino/native/quantization/test_quantizer_config.py @@ -12,10 +12,12 @@ import pytest from nncf.common.graph.transformations.commands import TargetType +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.experimental.common.tensor_statistics.collectors import MaxAggregator from nncf.experimental.common.tensor_statistics.collectors import MeanAggregator from nncf.experimental.common.tensor_statistics.collectors import MinAggregator from nncf.experimental.common.tensor_statistics.collectors import TensorCollector +from nncf.experimental.common.tensor_statistics.collectors import TensorReducerBase from nncf.openvino.graph.layer_attributes import OVLayerAttributes from nncf.openvino.graph.metatypes.openvino_metatypes import OVConvolutionMetatype from nncf.openvino.graph.metatypes.openvino_metatypes import OVDepthwiseConvolutionMetatype @@ -29,6 +31,9 @@ ParamsCls = TemplateTestQuantizerConfig.TestGetStatisticsCollectorParameters +# pylint: disable=protected-access + + class TestQuantizerConfig(TemplateTestQuantizerConfig): def get_algo_backend(self): return OVMinMaxAlgoBackend() @@ -45,6 +50,9 @@ def check_is_mean_min_max_statistic_collector(self, tensor_collector: TensorColl assert MeanAggregator in aggrs assert aggrs[0].__class__ == aggrs[1].__class__ + def get_reduction_axes(self, reducer: TensorReducerBase) -> ReductionAxes: + return reducer._reduction_axes + @pytest.fixture( params=[ pytest.param( diff --git a/tests/openvino/native/quantization/test_reducers_and_aggregators.py b/tests/openvino/native/quantization/test_reducers_and_aggregators.py index b34fc27a7c4..7b13174e961 100644 --- a/tests/openvino/native/quantization/test_reducers_and_aggregators.py +++ b/tests/openvino/native/quantization/test_reducers_and_aggregators.py @@ -9,9 +9,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +from typing import Any, List, Optional, Tuple + import numpy as np import pytest +from nncf.common.graph.layer_attributes import Dtype from nncf.openvino.statistics.collectors import OVAbsMaxReducer from nncf.openvino.statistics.collectors import OVAbsQuantileReducer from nncf.openvino.statistics.collectors import OVBatchMeanReducer @@ -31,7 +34,7 @@ class TestReducersAggregators(TemplateTestReducersAggreagtors): def tensor_processor(self): return OVNNCFCollectorTensorProcessor - def get_nncf_tensor(self, x: np.array): + def get_nncf_tensor(self, x: np.array, dtype: Optional[Dtype] = None): return OVNNCFTensor(x) @pytest.fixture(scope="module") @@ -52,3 +55,12 @@ def all_close(self, val, ref) -> bool: val_ = np.array(val) ref_ = np.array(ref) return np.allclose(val_, ref_) and val_.shape == ref_.shape + + def squeeze_tensor(self, ref_tensor: List[Any], axes: Optional[Tuple[int]] = None): + return np.squeeze(np.array(ref_tensor), axes) + + def cast_tensor(self, tensor, dtype: Dtype): + return tensor + + def expand_dims(self, tensor, dims: Tuple[int, ...]): + return np.expand_dims(np.array(tensor), dims) diff --git a/tests/openvino/native/test_model_transformer.py b/tests/openvino/native/test_model_transformer.py index ccccfb38ae2..608e98a5e26 100644 --- a/tests/openvino/native/test_model_transformer.py +++ b/tests/openvino/native/test_model_transformer.py @@ -286,7 +286,7 @@ def test_split_inplace_fn_insertion(test_params: InplaceOpTestCase): ) def test_inplace_reduce_fn_dynamic_shapes(input_shape, raise_error): input_1 = opset.parameter(input_shape, name="Input") - fn = get_inplace_min_op("test", reduction_shape=None) + fn = get_inplace_min_op("test", reduction_axes=None) if raise_error: with pytest.raises(RuntimeError): fn(input_1, 0) @@ -297,8 +297,8 @@ def test_inplace_reduce_fn_dynamic_shapes(input_shape, raise_error): assert all(np.equal(get_prev_node(op, 1).get_data(), ref_const)) -@pytest.mark.parametrize("reduction_shape", [None, np.array([], dtype=np.int64)]) -def test_inplace_reduce_fn_zero_rank_output(reduction_shape): +@pytest.mark.parametrize("reduction_axes", [None, np.array([], dtype=np.int64)]) +def test_inplace_reduce_fn_zero_rank_output(reduction_axes): model = ZeroRankEltwiseModel().ov_model target_layer = "Add" port_id = 1 @@ -310,7 +310,7 @@ def test_inplace_reduce_fn_zero_rank_output(reduction_shape): OVInplaceFnInsertionCommand, port_id, { - "inplace_op_fn": get_inplace_min_op(name, reduction_shape=reduction_shape), + "inplace_op_fn": get_inplace_min_op(name, reduction_axes=reduction_axes), "fn_output_port_id": 0, }, ) diff --git a/tests/openvino/native/test_node_utils.py b/tests/openvino/native/test_node_utils.py index 4a2a24872d4..3fbd595af23 100644 --- a/tests/openvino/native/test_node_utils.py +++ b/tests/openvino/native/test_node_utils.py @@ -17,7 +17,7 @@ from nncf.openvino.graph.layer_attributes import OVLayerAttributes from nncf.openvino.graph.metatypes.openvino_metatypes import OVMatMulMetatype from nncf.openvino.graph.nncf_graph_builder import GraphConverter -from nncf.openvino.graph.node_utils import get_channel_agnostic_reduction_shape +from nncf.openvino.graph.node_utils import get_channel_agnostic_reduction_axes from nncf.openvino.graph.node_utils import get_weight_channel_axes from nncf.openvino.graph.node_utils import get_weight_value from nncf.openvino.graph.node_utils import is_node_with_bias @@ -91,7 +91,7 @@ def test_get_weight_channel_axes_for_matmul(weights_port_id, transpose, shape, e @pytest.mark.parametrize( - "shape, channel_axes, ref_reduction_shape", + "shape, channel_axes, ref_reduction_axes", [ ((1, 128), [-1], (0,)), ((1, 256, 1), [-2], (0, 2)), @@ -101,7 +101,7 @@ def test_get_weight_channel_axes_for_matmul(weights_port_id, transpose, shape, e ((1, 1, 12, 12), [1, 2], (0, 3)), ], ) -def test_get_channel_agnostic_reduction_shape(shape, channel_axes, ref_reduction_shape): - reduction_shape = get_channel_agnostic_reduction_shape(channel_axes=channel_axes, shape=shape) +def test_get_channel_agnostic_reduction_axes(shape, channel_axes, ref_reduction_axes): + reduction_axes = get_channel_agnostic_reduction_axes(channel_axes=channel_axes, shape=shape) - assert reduction_shape == ref_reduction_shape + assert reduction_axes == ref_reduction_axes diff --git a/tests/openvino/native/test_statistic_collector.py b/tests/openvino/native/test_statistic_collector.py index 2d52c0af4cd..32123644942 100644 --- a/tests/openvino/native/test_statistic_collector.py +++ b/tests/openvino/native/test_statistic_collector.py @@ -9,6 +9,18 @@ # See the License for the specific language governing permissions and # limitations under the License. +from typing import Type + +import pytest + +from nncf.common.tensor_statistics.statistics import MeanTensorStatistic +from nncf.common.tensor_statistics.statistics import MedianMADTensorStatistic +from nncf.common.tensor_statistics.statistics import MinMaxTensorStatistic +from nncf.common.tensor_statistics.statistics import PercentileTensorStatistic +from nncf.common.tensor_statistics.statistics import RawTensorStatistic +from nncf.openvino.statistics.statistics import OVMeanTensorStatistic +from nncf.openvino.statistics.statistics import OVMinMaxTensorStatistic +from nncf.openvino.statistics.statistics import OVRawTensorStatistic from nncf.openvino.tensor import OVNNCFTensor from tests.common.experimental.test_statistic_collector import TemplateTestStatisticCollector @@ -16,3 +28,31 @@ class TestOVStatisticCollector(TemplateTestStatisticCollector): def get_nncf_tensor_cls(self): return OVNNCFTensor + + @pytest.fixture + def min_max_statistic_cls(self) -> Type[MinMaxTensorStatistic]: + return OVMinMaxTensorStatistic + + @pytest.fixture + def mean_statistic_cls(self) -> Type[MeanTensorStatistic]: + return OVMeanTensorStatistic + + @pytest.fixture + def median_mad_statistic_cls(self) -> Type[MedianMADTensorStatistic]: + raise NotImplementedError() + + @pytest.mark.skip() + def test_median_mad_stat_building(self, median_mad_statistic_cls: MedianMADTensorStatistic): + pass + + @pytest.fixture + def percentile_statistic_cls(self) -> Type[PercentileTensorStatistic]: + raise NotImplementedError() + + @pytest.mark.skip + def test_percentile_max_stat_building(self, percentile_statistic_cls: PercentileTensorStatistic): + pass + + @pytest.fixture + def raw_statistic_cls(self) -> Type[RawTensorStatistic]: + return OVRawTensorStatistic diff --git a/tests/post_training/test_templates/test_calculate_quantizer_parameters.py b/tests/post_training/test_templates/test_calculate_quantizer_parameters.py index b9b93153cfc..bffe249d064 100644 --- a/tests/post_training/test_templates/test_calculate_quantizer_parameters.py +++ b/tests/post_training/test_templates/test_calculate_quantizer_parameters.py @@ -213,7 +213,7 @@ def test_calculate_quantizer_parameters(self, case_to_test): else: max_values = np.amax(data, axis=axes, keepdims=q_config.per_channel) - statistics = self.tensor_statistic(max_values=max_values, min_values=min_values) + statistics = self.tensor_statistic(min_values=min_values, max_values=max_values) if not case_to_test.should_fail: fq_params = calculate_quantizer_parameters(statistics, q_config, quant_group, narrow_range, half_range) diff --git a/tests/post_training/test_templates/test_channel_alignment.py b/tests/post_training/test_templates/test_channel_alignment.py index d3b6dd045e5..21f6cfdeaaf 100644 --- a/tests/post_training/test_templates/test_channel_alignment.py +++ b/tests/post_training/test_templates/test_channel_alignment.py @@ -480,20 +480,21 @@ class MockBackend(backend_cls): @pytest.mark.parametrize("inplace_ref", [False, True]) @pytest.mark.parametrize("q_ref", [1e-4, 0.3]) def test_statistic_collectors(self, inplace_ref, q_ref): - reduction_shape_ref = (0, 2, 3) + reduction_axes_ref = (0, 2, 3) num_samples_ref = 123 statistic_collector: TensorCollector = self.get_backend_cls().get_statistic_collector( - reduction_shape=reduction_shape_ref, q=q_ref, num_samples=num_samples_ref, inplace=inplace_ref + reduction_axes=reduction_axes_ref, q=q_ref, num_samples=num_samples_ref, inplace=inplace_ref ) assert len(statistic_collector.reducers) == 1 reducer = statistic_collector.reducers.pop() assert isinstance(reducer, QuantileReducer) - assert reducer._reduction_shape == reduction_shape_ref + assert reducer._reduction_axes == reduction_axes_ref assert np.allclose(reducer._quantile, (q_ref, 1 - q_ref)) assert len(statistic_collector.aggregators) == 2 for aggr in statistic_collector.aggregators.values(): assert isinstance(aggr, MedianAggregator) assert aggr.num_samples == num_samples_ref - assert not aggr._use_per_sample_stats + assert not aggr._keepdims + assert aggr._aggregation_axes == (0,) diff --git a/tests/post_training/test_templates/test_quantizer_config.py b/tests/post_training/test_templates/test_quantizer_config.py index e614138d0a9..72da2111a36 100644 --- a/tests/post_training/test_templates/test_quantizer_config.py +++ b/tests/post_training/test_templates/test_quantizer_config.py @@ -25,6 +25,7 @@ from nncf.common.quantization.structs import QuantizationPreset from nncf.common.quantization.structs import QuantizerConfig from nncf.common.quantization.structs import QuantizerGroup +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.experimental.common.tensor_statistics.collectors import AbsMaxReducer from nncf.experimental.common.tensor_statistics.collectors import MaxReducer from nncf.experimental.common.tensor_statistics.collectors import MinReducer @@ -53,6 +54,10 @@ def check_is_min_max_statistic_collector(self, tensor_collector): def check_is_mean_min_max_statistic_collector(self, tensor_collector): pass + @abstractmethod + def get_reduction_axes(self, reducer) -> ReductionAxes: + pass + @abstractmethod @pytest.fixture def single_conv_nncf_graph(self) -> NNCFGraphToTest: @@ -72,8 +77,8 @@ def conv_sum_aggregation_nncf_graph(self) -> NNCFGraphToTestSumAggregation: class TestGetStatisticsCollectorParameters: target_type: TargetType target_node_name: str - ref_per_ch_reduction_shape: List[int] - ref_per_tensor_reduction_shape: List[int] + ref_per_ch_reduction_axes: List[int] + ref_per_tensor_reduction_axes: List[int] @abstractmethod @pytest.fixture @@ -278,8 +283,8 @@ def test_get_stat_collector( for reducer in reducers: if q_config_per_channel: - assert reducer._reduction_shape == params.ref_per_ch_reduction_shape + assert self.get_reduction_axes(reducer) == params.ref_per_ch_reduction_axes else: - assert reducer._reduction_shape == params.ref_per_tensor_reduction_shape + assert self.get_reduction_axes(reducer) == params.ref_per_tensor_reduction_axes assert tensor_collector.num_samples == num_samples diff --git a/tests/post_training/test_templates/test_smooth_quant.py b/tests/post_training/test_templates/test_smooth_quant.py index 42fe17e01b0..5ea93677f66 100644 --- a/tests/post_training/test_templates/test_smooth_quant.py +++ b/tests/post_training/test_templates/test_smooth_quant.py @@ -128,13 +128,13 @@ def test_smooth_quant_algo(self, model_cls, reference_values, tmpdir): # pylint:disable=protected-access def test_get_abs_max_channel_collector(self): backend = self.get_backend() - reduction_shape = (3, 2, 1) + reduction_axes = (3, 2, 1) samples = 1 for inplace_type in [False, True]: backend_tensor_collector = backend.get_abs_max_channel_collector( num_samples=samples, - stats_reduction_shape=reduction_shape, + stats_reduction_axes=reduction_axes, inplace=inplace_type, branch_key="test_branch", ) @@ -145,7 +145,7 @@ def test_get_abs_max_channel_collector(self): for reducer in backend_tensor_collector.reducers: assert isinstance(reducer, AbsMaxReducer) assert reducer.inplace == inplace_type - assert reducer._reduction_shape == reduction_shape + assert reducer._reduction_axes == reduction_axes @pytest.mark.parametrize( "model_cls, references", diff --git a/tests/tensorflow/tensor_statistics/test_tensor_statistics.py b/tests/tensorflow/tensor_statistics/test_tensor_statistics.py index e99487c69d9..30a675f7308 100644 --- a/tests/tensorflow/tensor_statistics/test_tensor_statistics.py +++ b/tests/tensorflow/tensor_statistics/test_tensor_statistics.py @@ -16,7 +16,7 @@ import tensorflow as tf from nncf.common.tensor_statistics.collectors import OfflineTensorStatisticCollector -from nncf.common.tensor_statistics.collectors import ReductionShape +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.common.tensor_statistics.collectors import StatisticsNotCollectedError from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase from nncf.common.tensor_statistics.statistics import TensorStatistic @@ -101,7 +101,7 @@ class TestCollectedStatistics: def test_collected_statistics_with_shape_convert( self, collector: Type[TensorStatisticCollectorBase], - reduction_shapes_vs_ref_statistic: Dict[Tuple[ReductionShape, ReductionShape], TensorStatistic], + reduction_shapes_vs_ref_statistic: Dict[Tuple[ReductionAxes, ReductionAxes], TensorStatistic], ): for reduction_shape in reduction_shapes_vs_ref_statistic.keys(): collector_obj = collector(use_abs_max=True, reduction_shape=reduction_shape) @@ -179,7 +179,7 @@ def test_collected_statistics_with_shape_convert( def test_collected_statistics( self, collector: Type[TensorStatisticCollectorBase], - reduction_shapes_vs_ref_statistic: Dict[ReductionShape, TensorStatistic], + reduction_shapes_vs_ref_statistic: Dict[ReductionAxes, TensorStatistic], ): for reduction_shape in reduction_shapes_vs_ref_statistic.keys(): collector_obj = collector(reduction_shape=reduction_shape) @@ -263,7 +263,7 @@ def test_num_samples(self, collector_for_num_samples_test: OfflineTensorStatisti class TestCollectorTensorProcessor: - tensor_processor = TFNNCFCollectorTensorProcessor() + tensor_processor = TFNNCFCollectorTensorProcessor def test_unstack(self): # Unstack tensor with dimensions diff --git a/tests/torch/ptq/helpers.py b/tests/torch/ptq/helpers.py index e019eb00f02..5e7b672aae1 100644 --- a/tests/torch/ptq/helpers.py +++ b/tests/torch/ptq/helpers.py @@ -94,6 +94,6 @@ def mock_collect_statistics(mocker): min_, max_ = 0.0, 1.0 min_, max_ = torch.tensor(min_), torch.tensor(max_) _ = mocker.patch( - "nncf.common.tensor_statistics.collectors.TensorStatisticCollectorBase.get_statistics", - return_value=PTMinMaxTensorStatistic(min_, max_), + "nncf.experimental.common.tensor_statistics.collectors.TensorCollector.get_statistics", + return_value=PTMinMaxTensorStatistic(min_values=min_, max_values=max_), ) diff --git a/tests/torch/ptq/test_calculation_quantizer_params.py b/tests/torch/ptq/test_calculation_quantizer_params.py index 15e9c1ea9c6..3c0cf83a64e 100644 --- a/tests/torch/ptq/test_calculation_quantizer_params.py +++ b/tests/torch/ptq/test_calculation_quantizer_params.py @@ -271,7 +271,8 @@ def calculate_statistics(data, mode, qgroup, half_range=False): max_values = np.amax(data, axes) statistics = PTMinMaxTensorStatistic( - min_values=torch.from_numpy(np.array(min_values)), max_values=torch.from_numpy(np.array(max_values)) + min_values=torch.from_numpy(np.array(min_values)), + max_values=torch.from_numpy(np.array(max_values)), ) signedness_to_force = True if qgroup == QuantizerGroup.WEIGHTS else None qconfig = QuantizerConfig(num_bits=8, mode=mode, per_channel=per_ch, signedness_to_force=signedness_to_force) diff --git a/tests/torch/ptq/test_ptq_params.py b/tests/torch/ptq/test_ptq_params.py index c174ec8b322..35ddfe3128e 100644 --- a/tests/torch/ptq/test_ptq_params.py +++ b/tests/torch/ptq/test_ptq_params.py @@ -18,6 +18,10 @@ from nncf.common.graph.transformations.commands import TargetType from nncf.common.quantization.structs import QuantizationPreset from nncf.common.utils.backend import BackendType +from nncf.experimental.common.tensor_statistics.collectors import MaxAggregator +from nncf.experimental.common.tensor_statistics.collectors import MeanAggregator +from nncf.experimental.common.tensor_statistics.collectors import MinAggregator +from nncf.experimental.common.tensor_statistics.collectors import TensorCollector from nncf.parameters import ModelType from nncf.parameters import TargetDevice from nncf.quantization.advanced_parameters import AdvancedQuantizationParameters @@ -33,8 +37,6 @@ from nncf.torch.graph.operator_metatypes import PTModuleLinearMetatype from nncf.torch.graph.operator_metatypes import PTSoftmaxMetatype from nncf.torch.quantization.quantize_model import _create_nncf_config -from nncf.torch.tensor_statistics.collectors import PTMeanMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMinMaxStatisticCollector from tests.common.quantization.metatypes import Conv2dTestMetatype from tests.common.quantization.metatypes import LinearTestMetatype from tests.common.quantization.metatypes import SoftmaxTestMetatype @@ -104,11 +106,17 @@ class TestPTQParams(TemplateTestPTQParams): def get_algo_backend(self): return PTMinMaxAlgoBackend() - def check_is_min_max_statistic_collector(self, tensor_collector): - assert isinstance(tensor_collector, PTMinMaxStatisticCollector) - - def check_is_mean_min_max_statistic_collector(self, tensor_collector): - assert isinstance(tensor_collector, PTMeanMinMaxStatisticCollector) + def check_is_min_max_statistic_collector(self, tensor_collector: TensorCollector): + aggrs = [aggr.__class__ for aggr in tensor_collector.aggregators.values()] + assert len(aggrs) == 2 + assert MinAggregator in aggrs + assert MaxAggregator in aggrs + + def check_is_mean_min_max_statistic_collector(self, tensor_collector: TensorCollector): + aggrs = [aggr.__class__ for aggr in tensor_collector.aggregators.values()] + assert len(aggrs) == 2 + assert MeanAggregator in aggrs + assert aggrs[0].__class__ == aggrs[1].__class__ def check_quantize_outputs_fq_num(self, quantize_outputs, act_num_q, weight_num_q): if quantize_outputs: diff --git a/tests/torch/ptq/test_quantizer_config.py b/tests/torch/ptq/test_quantizer_config.py index 41cab6438b5..98e7de76ca2 100644 --- a/tests/torch/ptq/test_quantizer_config.py +++ b/tests/torch/ptq/test_quantizer_config.py @@ -9,12 +9,18 @@ # See the License for the specific language governing permissions and # limitations under the License. +from typing import Tuple + import pytest from nncf.common.graph.transformations.commands import TargetType +from nncf.common.tensor_statistics.collectors import ReductionAxes +from nncf.experimental.common.tensor_statistics.collectors import MaxAggregator +from nncf.experimental.common.tensor_statistics.collectors import MeanAggregator +from nncf.experimental.common.tensor_statistics.collectors import MinAggregator +from nncf.experimental.common.tensor_statistics.collectors import TensorCollector +from nncf.experimental.common.tensor_statistics.collectors import TensorReducerBase from nncf.quantization.algorithms.min_max.torch_backend import PTMinMaxAlgoBackend -from nncf.torch.tensor_statistics.collectors import PTMeanMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMinMaxStatisticCollector from tests.post_training.test_templates.models import NNCFGraphToTest from tests.post_training.test_templates.models import NNCFGraphToTestDepthwiseConv from tests.post_training.test_templates.models import NNCFGraphToTestSumAggregation @@ -26,15 +32,27 @@ ParamsCls = TemplateTestQuantizerConfig.TestGetStatisticsCollectorParameters +# pylint: disable=protected-access + + class TestQuantizerConfig(TemplateTestQuantizerConfig): def get_algo_backend(self): return PTMinMaxAlgoBackend() - def check_is_min_max_statistic_collector(self, tensor_collector): - assert isinstance(tensor_collector, PTMinMaxStatisticCollector) + def check_is_min_max_statistic_collector(self, tensor_collector: TensorCollector): + aggrs = [aggr.__class__ for aggr in tensor_collector.aggregators.values()] + assert len(aggrs) == 2 + assert MinAggregator in aggrs + assert MaxAggregator in aggrs + + def check_is_mean_min_max_statistic_collector(self, tensor_collector: TensorCollector): + aggrs = [aggr.__class__ for aggr in tensor_collector.aggregators.values()] + assert len(aggrs) == 2 + assert MeanAggregator in aggrs + assert aggrs[0].__class__ == aggrs[1].__class__ - def check_is_mean_min_max_statistic_collector(self, tensor_collector): - assert isinstance(tensor_collector, PTMeanMinMaxStatisticCollector) + def get_reduction_axes(self, reducer: TensorReducerBase) -> ReductionAxes: + return reducer._reduction_axes @pytest.fixture( params=[ diff --git a/tests/torch/ptq/test_reducers_and_aggregators.py b/tests/torch/ptq/test_reducers_and_aggregators.py new file mode 100644 index 00000000000..5f63a9da861 --- /dev/null +++ b/tests/torch/ptq/test_reducers_and_aggregators.py @@ -0,0 +1,103 @@ +# Copyright (c) 2023 Intel Corporation +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# http://www.apache.org/licenses/LICENSE-2.0 +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from abc import ABC +from typing import Any, List, Optional, Tuple + +import numpy as np +import pytest +import torch + +from nncf.common.graph.layer_attributes import Dtype +from nncf.torch.tensor import PTNNCFTensor +from nncf.torch.tensor_statistics.collectors import PTAbsMaxReducer +from nncf.torch.tensor_statistics.collectors import PTAbsQuantileReducer +from nncf.torch.tensor_statistics.collectors import PTBatchMeanReducer +from nncf.torch.tensor_statistics.collectors import PTMaxReducer +from nncf.torch.tensor_statistics.collectors import PTMeanPerChanelReducer +from nncf.torch.tensor_statistics.collectors import PTMeanReducer +from nncf.torch.tensor_statistics.collectors import PTMinReducer +from nncf.torch.tensor_statistics.collectors import PTNNCFCollectorTensorProcessor +from nncf.torch.tensor_statistics.collectors import PTNoopReducer +from nncf.torch.tensor_statistics.collectors import PTQuantileReducer +from tests.common.experimental.test_reducers_and_aggregators import TemplateTestReducersAggreagtors + + +class BaseTestReducersAggregators(TemplateTestReducersAggreagtors, ABC): + @pytest.fixture + def tensor_processor(self): + return PTNNCFCollectorTensorProcessor + + def _get_torch_tensor(self, x: np.ndarray, dtype: Optional[Dtype] = None): + torch_tensor = torch.tensor(x) + if dtype == Dtype.FLOAT: + torch_tensor = torch_tensor.float() + elif dtype == Dtype.INTEGER: + torch_tensor = torch_tensor.int() + return torch_tensor + + @pytest.fixture(scope="module") + def reducers(self): + return { + "noop": PTNoopReducer, + "min": PTMinReducer, + "max": PTMaxReducer, + "abs_max": PTAbsMaxReducer, + "mean": PTMeanReducer, + "quantile": PTQuantileReducer, + "abs_quantile": PTAbsQuantileReducer, + "batch_mean": PTBatchMeanReducer, + "mean_per_ch": PTMeanPerChanelReducer, + } + + def all_close(self, val, ref) -> bool: + val_ = val + ref_ = torch.tensor(ref).to(val_.device) + return torch.allclose(val_, ref_) and val_.shape == ref_.shape + + def squeeze_tensor(self, ref_tensor: List[Any], axes: Optional[Tuple[int]] = None): + if axes is None: + return torch.tensor(ref_tensor).squeeze() + return torch.tensor(ref_tensor).squeeze(axes) + + def cast_tensor(self, tensor, dtype: Dtype): + tensor = torch.tensor(tensor) + if dtype == Dtype.FLOAT: + return tensor.float() + if dtype == Dtype.INTEGER: + return tensor.int() + raise RuntimeError() + + def expand_dims(self, tensor, dims: Tuple[int, ...]): + tensor_ = torch.tensor(tensor) + shape = list(tensor_.shape) + for dim in dims: + shape.insert(dim, 1) + return tensor_.view(shape) + + +class TestCPUReducersAggregators(BaseTestReducersAggregators): + def get_nncf_tensor(self, x: np.array, dtype: Optional[Dtype] = None): + return PTNNCFTensor(self._get_torch_tensor(x, dtype=dtype).cpu()) + + def all_close(self, val: torch.Tensor, ref) -> bool: + assert not val.is_cuda + return super().all_close(val, ref) + + +@pytest.mark.skipif(not torch.cuda.is_available(), reason="Cuda is not available in current environment") +class TestCudaReducersAggregators(BaseTestReducersAggregators): + def get_nncf_tensor(self, x: np.array, dtype: Optional[Dtype] = None): + return PTNNCFTensor(self._get_torch_tensor(x, dtype=dtype).cuda()) + + def all_close(self, val: torch.Tensor, ref) -> bool: + assert val.is_cuda + return super().all_close(val, ref) diff --git a/tests/torch/ptq/test_statistic_collector.py b/tests/torch/ptq/test_statistic_collector.py new file mode 100644 index 00000000000..0ab1ef2bb55 --- /dev/null +++ b/tests/torch/ptq/test_statistic_collector.py @@ -0,0 +1,55 @@ +# Copyright (c) 2023 Intel Corporation +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# http://www.apache.org/licenses/LICENSE-2.0 +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Type + +import pytest + +from nncf.common.tensor_statistics.statistics import MeanTensorStatistic +from nncf.common.tensor_statistics.statistics import MedianMADTensorStatistic +from nncf.common.tensor_statistics.statistics import MinMaxTensorStatistic +from nncf.common.tensor_statistics.statistics import PercentileTensorStatistic +from nncf.common.tensor_statistics.statistics import RawTensorStatistic +from nncf.torch.tensor import PTNNCFTensor +from nncf.torch.tensor_statistics.statistics import PTMeanTensorStatistic +from nncf.torch.tensor_statistics.statistics import PTMedianMADTensorStatistic +from nncf.torch.tensor_statistics.statistics import PTMinMaxTensorStatistic +from nncf.torch.tensor_statistics.statistics import PTPercentileTensorStatistic +from tests.common.experimental.test_statistic_collector import TemplateTestStatisticCollector + + +class TestOVStatisticCollector(TemplateTestStatisticCollector): + def get_nncf_tensor_cls(self): + return PTNNCFTensor + + @pytest.fixture + def min_max_statistic_cls(self) -> Type[MinMaxTensorStatistic]: + return PTMinMaxTensorStatistic + + @pytest.fixture + def mean_statistic_cls(self) -> Type[MeanTensorStatistic]: + return PTMeanTensorStatistic + + @pytest.fixture + def median_mad_statistic_cls(self) -> Type[MedianMADTensorStatistic]: + return PTMedianMADTensorStatistic + + @pytest.fixture + def percentile_statistic_cls(self) -> Type[PercentileTensorStatistic]: + return PTPercentileTensorStatistic + + @pytest.fixture + def raw_statistic_cls(self) -> Type[RawTensorStatistic]: + raise NotImplementedError() + + @pytest.mark.skip + def test_raw_max_stat_building(self, raw_statistic_cls: RawTensorStatistic): + pass diff --git a/tests/torch/quantization/test_range_init.py b/tests/torch/quantization/test_range_init.py index 84975de7bd4..6ef968a8bf9 100644 --- a/tests/torch/quantization/test_range_init.py +++ b/tests/torch/quantization/test_range_init.py @@ -11,8 +11,9 @@ import itertools import re from collections import namedtuple +from dataclasses import dataclass from functools import partial -from typing import List, Tuple +from typing import List, Tuple, Union import pytest import torch @@ -22,6 +23,7 @@ from torch.utils.data import DataLoader from torchvision.models import squeezenet1_1 +import nncf.torch.tensor_statistics.collectors as pt_collectors from nncf.common.graph import NNCFNodeName from nncf.common.quantization.initialization.range import PerLayerRangeInitConfig from nncf.common.quantization.initialization.range import RangeInitConfig @@ -46,9 +48,7 @@ from nncf.torch.quantization.layers import BaseQuantizer from nncf.torch.quantization.layers import PTQuantizerSpec from nncf.torch.quantization.layers import SymmetricQuantizer -from nncf.torch.tensor_statistics.collectors import PTMeanMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMedianMADStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMinMaxStatisticCollector +from nncf.torch.tensor import PTNNCFTensor from nncf.torch.tensor_statistics.statistics import pt_convert_stat_to_min_max_tensor_stat from nncf.torch.utils import get_all_modules_by_type from nncf.torch.utils import safe_thread_call @@ -64,6 +64,8 @@ from tests.torch.quantization.quantization_helpers import post_compression_test_distr_init # pylint:disable=unused-import +# pylint:disable=protected-access +# pylint:disable=too-many-lines def scale_signed_dumping_worker(gpu, ngpus_per_node, config, tmp_path): @@ -493,21 +495,23 @@ def forward(self, input_): return self.conv2d(input_) +def _get_init_tensor_for_range_init_test() -> torch.Tensor: + test_input_sample = torch.empty([3, 100, 100]) + test_input_sample[0] = torch.range(1, 10_000).view((100, 100)) + test_input_sample[1] = test_input_sample[0] * -2 + test_input_sample[2] = test_input_sample[0] * 3 + return test_input_sample + + class SingleConv2dSyntheticWeightModel(torch.nn.Module): def __init__(self): super().__init__() self.conv2d = nn.Conv2d(3, 3, 100) with torch.no_grad(): - for i in range(0, 100): - for j in range(0, 100): - self.conv2d.weight[0][0][i][j] = i * 100 + j - + value = _get_init_tensor_for_range_init_test() for i in range(0, 3): - for j in range(0, 3): - if not (i == 0 and j == 0): - self.conv2d.weight[i][j] = self.conv2d.weight[0][0] - self.conv2d.weight[i][j] = self.conv2d.weight[0][0] + self.conv2d.weight[:, i] = value def forward(self, input_): return self.conv2d(input_) @@ -519,36 +523,228 @@ def init_idfn(val): return val +@dataclass +class SymQuantizerScaleRef: + scale: Tuple[float, ...] + + +@dataclass +class AsymQuantizerScaleRef: + input_low: Tuple[float, ...] + input_range: Tuple[float, ...] + + +@dataclass +class GranularityQuantizerRefs: + per_channel: Union[SymQuantizerScaleRef, AsymQuantizerScaleRef] + per_tensor: Union[SymQuantizerScaleRef, AsymQuantizerScaleRef] + + +@dataclass +class RangeInitTestCase: + collector_name: str + weights_refs_symmetric: GranularityQuantizerRefs + weights_refs_assymetric: GranularityQuantizerRefs + activations_refs_symmetric: GranularityQuantizerRefs + activations_refs_assymetric: GranularityQuantizerRefs + + @pytest.mark.parametrize( - "range_init_type_vs_ref_vals", + "range_init_test_case", ( [ - ("min_max", 9999, 0, 9999), - ("mixed_min_max", 9999, 0, 9999), - ("mean_min_max", 9999, 0, 9999), - ("threesigma", 16119.5, -6119.5, 22239), - ("percentile", 6789, 3210, 3578), + RangeInitTestCase( + collector_name="min_max", + weights_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((10000.0, 20000.0, 30000.0)).view(((3, 1, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=30000.0), + ), + weights_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((1.0, -20000.0, 3.0)).view(((3, 1, 1, 1))), + input_range=torch.tensor((9999.0, 19998.0, 29997.0)).view(((3, 1, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-20000.0, input_range=50000.0), + ), + activations_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((20000.0, 40000.0, 60000.0)).view(((1, 3, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=60000.0), + ), + activations_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((1.0, -40000.0, 3.0)).view(((1, 3, 1, 1))), + input_range=torch.tensor((19999.0, 39998.0, 59997.0)).view(((1, 3, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-40000.0, input_range=100000.0), + ), + ), + RangeInitTestCase( + collector_name="mixed_min_max", + weights_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((10000.0, 20000.0, 30000.0)).view(((3, 1, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=30000.0), + ), + weights_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((1.0, -20000.0, 3.0)).view(((3, 1, 1, 1))), + input_range=torch.tensor((9999.0, 19998.0, 29997.0)).view(((3, 1, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-20000.0, input_range=50000.0), + ), + activations_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((20000.0, 40000.0, 60000.0)).view(((1, 3, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=45000.0), + ), + activations_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((1.0, -40000.0, 3.0)).view(((1, 3, 1, 1))), + input_range=torch.tensor((19999.0, 39998.0, 59997.0)).view(((1, 3, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-30000.0, input_range=75000.0), + ), + ), + RangeInitTestCase( + collector_name="mean_min_max", + weights_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((10000.0, 20000.0, 30000.0)).view(((3, 1, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=30000.0), + ), + weights_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((1.0, -20000.0, 3.0)).view(((3, 1, 1, 1))), + input_range=torch.tensor((9999.0, 19998.0, 29997.0)).view(((3, 1, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-20000.0, input_range=50000.0), + ), + activations_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((15000.0, 30000.0, 45000.0)).view(((1, 3, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=45000.0), + ), + activations_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((1.5, -30000.0, 4.5)).view(((1, 3, 1, 1))), + input_range=torch.tensor((14998.5000, 29997.0000, 44995.5000)).view(((1, 3, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-30000.0, input_range=75000.0), + ), + ), + RangeInitTestCase( + collector_name="threesigma", + weights_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((16120.1719, 32240.3438, 48360.5156)).view(((3, 1, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=33780.2891), + ), + weights_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((-6119.1719, -32240.3438, -18357.5156)).view(((3, 1, 1, 1))), + input_range=torch.tensor((22239.3438, 44478.6875, 66718.0312)).view(((3, 1, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-26279.2871, input_range=60059.5781), + ), + activations_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((21494.4707, 42988.9414, 64483.4141)).view(((1, 3, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=52662.1367), + ), + activations_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((-8159.4707, -42988.9414, -24478.4141)).view(((1, 3, 1, 1))), + input_range=torch.tensor((29653.9414, 59307.8828, 88961.8281)).view(((1, 3, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-42660.1367, input_range=95322.2734), + ), + ), + RangeInitTestCase( + collector_name="percentile", + weights_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((6789.3213, 13580.6416, 20367.9629)).view(((3, 1, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=7776.0), + ), + weights_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((3210.6790, -13580.6416, 9632.0371)).view(((3, 1, 1, 1))), + input_range=torch.tensor((3578.6423, 7157.2837, 10735.9258)).view(((3, 1, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-740.6420, input_range=8516.6416), + ), + activations_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((9052.3213, 18108.0000, 27156.9629)).view(((1, 3, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=10734.6426), + ), + activations_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((4280.6792, -18108.0000, 12842.0371)).view(((1, 3, 1, 1))), + input_range=torch.tensor((4771.6421, 9544.0000, 14314.9258)).view(((1, 3, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-988.0, input_range=11722.6426), + ), + ), + RangeInitTestCase( + collector_name="mean_percentile", + weights_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((9990.0010, 19980.0020, 29970.0039)).view(((3, 1, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=29910.0039), + ), + weights_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((10.999, -19980.0, 32.997)).view(((3, 1, 1, 1))), + input_range=torch.tensor((9979.0020, 19958.0039, 29937.0078)).view(((3, 1, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-19940.0020, input_range=49850.0078), + ), + activations_refs_symmetric=GranularityQuantizerRefs( + per_channel=SymQuantizerScaleRef( + scale=torch.tensor((14985.0020, 29970.0039, 44955.0078)).view(((1, 3, 1, 1))) + ), + per_tensor=SymQuantizerScaleRef(scale=44865.0078), + ), + activations_refs_assymetric=GranularityQuantizerRefs( + per_channel=AsymQuantizerScaleRef( + input_low=torch.tensor((16.498, -2.9970e04, 49.496)).view(((1, 3, 1, 1))), + input_range=torch.tensor((14968.5039, 29937.0078, 44905.5117)).view(((1, 3, 1, 1))), + ), + per_tensor=AsymQuantizerScaleRef(input_low=-29910.0039, input_range=74775.0156), + ), + ), ] ), ids=init_idfn, ) def test_init_ranges_are_set( - quantization_mode: str, is_per_channel: bool, range_init_type_vs_ref_vals: Tuple[str, float, float, float] + quantization_mode: str, + is_per_channel: bool, + range_init_test_case: RangeInitTestCase, ): class SyntheticDataset(torch.utils.data.Dataset): def __init__(self): super().__init__() - self._length = 1 + self._length = 2 def __getitem__(self, idx): if idx >= self._length: raise StopIteration - test_input_sample = torch.zeros([3, 100, 100]) - for i in range(0, 100): - for j in range(0, 100): - test_input_sample[0][i][j] = i * 100 + j - test_input_sample[1] = test_input_sample[0] - test_input_sample[2] = test_input_sample[0] + test_input_sample = _get_init_tensor_for_range_init_test() * (idx + 1) return test_input_sample, test_input_sample def __len__(self): @@ -556,7 +752,7 @@ def __len__(self): data_loader = torch.utils.data.DataLoader(SyntheticDataset(), batch_size=1, drop_last=True) - range_init_type = range_init_type_vs_ref_vals[0] + range_init_type = range_init_test_case.collector_name config_with_init = NNCFConfig() config_with_init.update( { @@ -566,7 +762,7 @@ def __len__(self): "algorithm": "quantization", "activations": {"mode": quantization_mode, "per_channel": is_per_channel}, "weights": {"mode": quantization_mode, "per_channel": is_per_channel}, - "initializer": {"range": {"num_init_samples": 1, "type": range_init_type}}, + "initializer": {"range": {"num_init_samples": 2, "type": range_init_type}}, }, } ) @@ -585,22 +781,30 @@ def __len__(self): act_quantizer_info = next(iter(compression_ctrl.non_weight_quantizers.values())) - ref_scale = range_init_type_vs_ref_vals[1] - ref_input_low = range_init_type_vs_ref_vals[2] - ref_input_high = range_init_type_vs_ref_vals[3] + if is_per_channel: + ref_scale = range_init_test_case.activations_refs_symmetric.per_channel.scale + ref_input_low = range_init_test_case.activations_refs_assymetric.per_channel.input_low + ref_input_range = range_init_test_case.activations_refs_assymetric.per_channel.input_range + else: + ref_scale = range_init_test_case.activations_refs_symmetric.per_tensor.scale + ref_input_low = range_init_test_case.activations_refs_assymetric.per_tensor.input_low + ref_input_range = range_init_test_case.activations_refs_assymetric.per_tensor.input_range def check_scales(quantizer: BaseQuantizer, per_channel: bool): # Absolute tolerance is 1.0 due to percentile value interpolation if quantization_mode == "symmetric": - assert torch.allclose(quantizer.scale, torch.ones_like(quantizer.scale) * ref_scale, atol=1.0) + assert torch.allclose(quantizer.scale, torch.tensor(ref_scale), atol=1.0) if per_channel: assert quantizer.scale.numel() == 3 else: assert quantizer.scale.numel() == 1 else: - assert torch.allclose(quantizer.input_low, torch.ones_like(quantizer.input_low) * ref_input_low, atol=1.0) + assert torch.allclose(quantizer.input_low, torch.tensor(ref_input_low), atol=1.0) + assert torch.allclose( - quantizer.input_range, torch.ones_like(quantizer.input_low) * ref_input_high, atol=1.0 + quantizer.input_range, + torch.tensor(ref_input_range), + atol=1.0, ) if per_channel: assert quantizer.input_low.numel() == 3 @@ -612,9 +816,19 @@ def check_scales(quantizer: BaseQuantizer, per_channel: bool): check_scales(act_quantizer_info.quantizer_module_ref, is_per_channel) # Weight init check synth_weight_model = SingleConv2dSyntheticWeightModel() + config_with_init["compression"]["initializer"]["range"]["num_init_samples"] = 1 _, compression_ctrl = create_compressed_model_and_algo_for_test(synth_weight_model, config_with_init) weight_quantizer_info = next(iter(compression_ctrl.weight_quantizers.values())) + if is_per_channel: + ref_scale = range_init_test_case.weights_refs_symmetric.per_channel.scale + ref_input_low = range_init_test_case.weights_refs_assymetric.per_channel.input_low + ref_input_range = range_init_test_case.weights_refs_assymetric.per_channel.input_range + else: + ref_scale = range_init_test_case.weights_refs_symmetric.per_tensor.scale + ref_input_low = range_init_test_case.weights_refs_assymetric.per_tensor.input_low + ref_input_range = range_init_test_case.weights_refs_assymetric.per_tensor.input_range + check_scales(weight_quantizer_info.quantizer_module_ref, is_per_channel) @@ -671,53 +885,66 @@ def range_init_call_count_test_struct(request): return request.param +class CustomSpy: + def __init__(self, fn) -> None: + self._fn = fn + self.call_count = 0 + self.return_values_list = [] + + def __call__(self, *args, **kwargs): + self.call_count += 1 + retval = self._fn(*args, **kwargs) + self.return_values_list.append(retval) + return retval + + # pylint:disable=redefined-outer-name def test_per_layer_range_init_collectors_are_called_the_required_number_of_times( range_init_call_count_test_struct, mocker ): + range_minmax_init_create_spy = CustomSpy(pt_collectors.get_min_max_statistic_collector) + mocker.patch("nncf.torch.quantization.init_range.get_min_max_statistic_collector", new=range_minmax_init_create_spy) + range_meanminmax_init_create_spy = CustomSpy(pt_collectors.get_mixed_min_max_statistic_collector) + mocker.patch( + "nncf.torch.quantization.init_range.get_mixed_min_max_statistic_collector", new=range_meanminmax_init_create_spy + ) + range_threesigma_init_create_spy = CustomSpy(pt_collectors.get_median_mad_statistic_collector) + mocker.patch( + "nncf.torch.quantization.init_range.get_median_mad_statistic_collector", new=range_threesigma_init_create_spy + ) + config = create_config() config["compression"]["initializer"]["range"] = range_init_call_count_test_struct.range_init_config data_loader = TestRangeInit.create_dataloader(True, config, 10) config.register_extra_structs([QuantizationRangeInitArgs(data_loader)]) - range_minmax_init_create_spy = mocker.spy(PTMinMaxStatisticCollector, "__init__") - range_meanminmax_init_create_spy = mocker.spy(PTMeanMinMaxStatisticCollector, "__init__") - range_threesigma_init_create_spy = mocker.spy(PTMedianMADStatisticCollector, "__init__") - - range_minmax_init_register_input_spy = mocker.spy(PTMinMaxStatisticCollector, "_register_input") - range_meanminmax_init_register_input_spy = mocker.spy(PTMeanMinMaxStatisticCollector, "_register_input") - range_threesigma_init_register_input_spy = mocker.spy(PTMedianMADStatisticCollector, "_register_input") - TestRangeInit.create_algo_and_compressed_model(config) - assert ( - range_minmax_init_create_spy.call_count - == range_init_call_count_test_struct.expected_call_count_initializer_create["min_max"] - ) - assert ( - range_meanminmax_init_create_spy.call_count - == range_init_call_count_test_struct.expected_call_count_initializer_create["mean_min_max"] - ) - assert ( - range_threesigma_init_create_spy.call_count - == range_init_call_count_test_struct.expected_call_count_initializer_create["three_sigma"] - ) - - assert ( - range_minmax_init_register_input_spy.call_count - == range_init_call_count_test_struct.expected_call_count_register_input["min_max"] - ) - assert ( - range_meanminmax_init_register_input_spy.call_count - == range_init_call_count_test_struct.expected_call_count_register_input["mean_min_max"] - ) - assert ( - range_threesigma_init_register_input_spy.call_count - == range_init_call_count_test_struct.expected_call_count_register_input["three_sigma"] - ) - - -QUANTIZER_RANGE_INITIALIZERS = ["min_max", "threesigma", "mean_min_max", "percentile", "mixed_min_max"] + for stat_type, spy in [ + ("min_max", range_minmax_init_create_spy), + ("mean_min_max", range_meanminmax_init_create_spy), + ("three_sigma", range_threesigma_init_create_spy), + ]: + assert spy.call_count == range_init_call_count_test_struct.expected_call_count_initializer_create[stat_type] + collected_samples = 0 + for tensor_collector in spy.return_values_list: + cur_values = set() + for aggr in tensor_collector.aggregators.values(): + cur_values.add(aggr._collected_samples) + assert len(cur_values) == 1 + collected_samples += cur_values.pop() + + assert collected_samples == range_init_call_count_test_struct.expected_call_count_register_input[stat_type] + + +QUANTIZER_RANGE_INITIALIZERS = [ + "min_max", + "threesigma", + "mean_min_max", + "percentile", + "mixed_min_max", + "mean_percentile", +] class QuantizeRangeInitScaleShapeTestStruct: @@ -794,7 +1021,7 @@ def test_quantize_range_init_sets_correct_scale_shapes(quantizer_range_init_test collector = StatCollectorGenerator.generate_stat_collector_for_range_init_config( range_init_config, tuple(quantizer.scale_shape), collector_params ) - collector.register_input(torch.ones(test_struct.input_shape)) + collector.register_input_for_all_reducers(PTNNCFTensor(torch.ones(test_struct.input_shape))) stat = collector.get_statistics() minmax_values = pt_convert_stat_to_min_max_tensor_stat(stat) quantizer.apply_minmax_init(min_values=minmax_values.min_values, max_values=minmax_values.max_values) diff --git a/tests/torch/tensor_statistics/test_tensor_statistics.py b/tests/torch/tensor_statistics/test_tensor_statistics.py index 5c5cccc0220..1d04f2a706b 100644 --- a/tests/torch/tensor_statistics/test_tensor_statistics.py +++ b/tests/torch/tensor_statistics/test_tensor_statistics.py @@ -15,19 +15,16 @@ import pytest import torch -from nncf.common.tensor_statistics.collectors import OfflineTensorStatisticCollector -from nncf.common.tensor_statistics.collectors import ReductionShape -from nncf.common.tensor_statistics.collectors import StatisticsNotCollectedError +from nncf.common.tensor_statistics.collectors import ReductionAxes from nncf.common.tensor_statistics.collectors import TensorStatisticCollectorBase from nncf.common.tensor_statistics.statistics import TensorStatistic from nncf.torch.tensor import PTNNCFTensor -from nncf.torch.tensor_statistics.collectors import PTMeanMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMeanPercentileStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMedianMADStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMinMaxStatisticCollector -from nncf.torch.tensor_statistics.collectors import PTMixedMinMaxStatisticCollector from nncf.torch.tensor_statistics.collectors import PTNNCFCollectorTensorProcessor -from nncf.torch.tensor_statistics.collectors import PTPercentileStatisticCollector +from nncf.torch.tensor_statistics.collectors import get_mean_percentile_statistic_collector +from nncf.torch.tensor_statistics.collectors import get_median_mad_statistic_collector +from nncf.torch.tensor_statistics.collectors import get_min_max_statistic_collector +from nncf.torch.tensor_statistics.collectors import get_mixed_min_max_statistic_collector +from nncf.torch.tensor_statistics.collectors import get_percentile_tensor_collector from nncf.torch.tensor_statistics.statistics import PTMedianMADTensorStatistic from nncf.torch.tensor_statistics.statistics import PTMinMaxTensorStatistic from nncf.torch.tensor_statistics.statistics import PTPercentileTensorStatistic @@ -40,19 +37,21 @@ class TestCollectedStatistics: ] @pytest.mark.parametrize( - ("collector", "reduction_shapes_vs_ref_statistic"), + ("collector", "reduction_axes_vs_ref_statistic"), [ ( - PTMinMaxStatisticCollector, + get_min_max_statistic_collector, { ((1,), (0, 1)): PTMinMaxTensorStatistic( min_values=torch.tensor([-4.0]), max_values=torch.tensor([6.1]) ), ((3, 1), (1,)): PTMinMaxTensorStatistic( - min_values=torch.tensor([[1.0], [-4.0], [4.0]]), max_values=torch.tensor([[4.5], [4.0], [6.1]]) + min_values=torch.tensor([[1.0], [-4.0], [4.0]]), + max_values=torch.tensor([[4.5], [4.0], [6.1]]), ), ((1, 3), (0,)): PTMinMaxTensorStatistic( - min_values=torch.tensor([[-1.3, -4.0, -3.5]]), max_values=torch.tensor([[4.5, 5.8, 6.1]]) + min_values=torch.tensor([[-1.3, -4.0, -3.5]]), + max_values=torch.tensor([[4.5, 5.8, 6.1]]), ), # Not supported for now: # ((3, 3), ): PTMinMaxTensorStatistic( @@ -70,7 +69,11 @@ class TestCollectedStatistics: }, ), ( - partial(PTMeanMinMaxStatisticCollector, use_per_sample_stats=False), + partial( + get_mixed_min_max_statistic_collector, + use_means_of_mins=True, + use_means_of_maxs=True, + ), { ((1,), (0, 1)): PTMinMaxTensorStatistic( min_values=torch.tensor([-3.5]), max_values=torch.tensor([6.05]) @@ -80,14 +83,14 @@ class TestCollectedStatistics: max_values=torch.tensor([[3.75], [3.5], [6.05]]), ), ((1, 3), (0,)): PTMinMaxTensorStatistic( - min_values=torch.tensor([[-1.15, -3, -3.25]]), max_values=torch.tensor([[4.25, 5.4, 6.05]]) + min_values=torch.tensor([[-1.15, -3, -3.25]]), + max_values=torch.tensor([[4.25, 5.4, 6.05]]), ), }, ), ( partial( - PTMixedMinMaxStatisticCollector, - use_per_sample_stats=False, + get_mixed_min_max_statistic_collector, use_means_of_mins=False, use_means_of_maxs=True, ), @@ -100,7 +103,8 @@ class TestCollectedStatistics: max_values=torch.tensor([[3.75], [3.5], [6.05]]), ), ((1, 3), (0,)): PTMinMaxTensorStatistic( - min_values=torch.tensor([[-1.3, -4.0, -3.5]]), max_values=torch.tensor([[4.25, 5.4, 6.05]]) + min_values=torch.tensor([[-1.3, -4.0, -3.5]]), + max_values=torch.tensor([[4.25, 5.4, 6.05]]), ), }, ), @@ -109,29 +113,40 @@ class TestCollectedStatistics: def test_collected_statistics_with_shape_convert( self, collector: Type[TensorStatisticCollectorBase], - reduction_shapes_vs_ref_statistic: Dict[Tuple[ReductionShape, ReductionShape], TensorStatistic], + reduction_axes_vs_ref_statistic: Dict[Tuple[ReductionAxes, ReductionAxes], TensorStatistic], ): - for shapes in reduction_shapes_vs_ref_statistic.keys(): - output_shape, reduction_shape = shapes - collector_obj = collector(use_abs_max=True, reduction_shape=reduction_shape, output_shape=output_shape) + for shapes in reduction_axes_vs_ref_statistic.keys(): + scale_shape, reducer_axes = shapes + collector_obj = collector( + scale_shape=scale_shape, + use_abs_max=True, + reduction_axes=reducer_axes, + aggregation_axes=(0,), + num_samples=None, + ) for input_ in TestCollectedStatistics.REF_INPUTS: - collector_obj.register_input(input_) + collector_obj.register_input_for_all_reducers(PTNNCFTensor(input_)) test_stats = collector_obj.get_statistics() - assert reduction_shapes_vs_ref_statistic[shapes] == test_stats + assert reduction_axes_vs_ref_statistic[shapes] == test_stats @pytest.mark.parametrize( - ("collector", "reduction_shapes_vs_ref_statistic"), + ("collector", "reduction_axes_vs_ref_statistic"), [ ( - PTMedianMADStatisticCollector, + get_median_mad_statistic_collector, + # PTMedianMADStatisticCollector, { - (1,): PTMedianMADTensorStatistic(median_values=torch.tensor([2.8]), mad_values=torch.tensor([2.6])), + (1,): PTMedianMADTensorStatistic( + median_values=torch.tensor([2.8]), + mad_values=torch.tensor([2.6]), + ), (3, 1): PTMedianMADTensorStatistic( median_values=torch.tensor([[2.8], [-2.5], [5.4]]), mad_values=torch.tensor([[0.85], [1.1], [0.65]]), ), (1, 3): PTMedianMADTensorStatistic( - median_values=torch.tensor([[2.5, 2.3, 3.35]]), mad_values=torch.tensor([[1.9, 3.1, 2.7]]) + median_values=torch.tensor([[2.5, 2.3, 3.35]]), + mad_values=torch.tensor([[1.9, 3.1, 2.7]]), ), # Not supported for now: # (3, 3): PTMedianMADTensorStatistic( @@ -149,7 +164,7 @@ def test_collected_statistics_with_shape_convert( }, ), ( - partial(PTPercentileStatisticCollector, percentiles_to_collect=[10.0]), + partial(get_percentile_tensor_collector, percentiles_to_collect=[10.0]), { (1,): PTPercentileTensorStatistic({10.0: torch.tensor([-3.15])}), (3, 1): PTPercentileTensorStatistic({10.0: torch.tensor([[1.5], [-3.75], [4.15]])}), @@ -167,7 +182,7 @@ def test_collected_statistics_with_shape_convert( }, ), ( - partial(PTMeanPercentileStatisticCollector, percentiles_to_collect=[10.0]), + partial(get_mean_percentile_statistic_collector, percentiles_to_collect=[10.0]), { (1,): PTPercentileTensorStatistic({10.0: torch.tensor([-2.9])}), (3, 1): PTPercentileTensorStatistic({10.0: torch.tensor([[2.0100], [-3.3500], [4.4000]])}), @@ -189,90 +204,24 @@ def test_collected_statistics_with_shape_convert( def test_collected_statistics( self, collector: Type[TensorStatisticCollectorBase], - reduction_shapes_vs_ref_statistic: Dict[ReductionShape, TensorStatistic], + reduction_axes_vs_ref_statistic: Dict[ReductionAxes, TensorStatistic], ): - for shapes in reduction_shapes_vs_ref_statistic.keys(): - reduction_shape = shapes - collector_obj = collector(reduction_shape=reduction_shape) + for reduction_axes in reduction_axes_vs_ref_statistic: + if len(reduction_axes) > 1: + reducer_axes = ([dim for dim, val in enumerate(reduction_axes) if val == 1][0],) + else: + reducer_axes = (0, 1) + + collector_obj = collector( + scale_shape=reduction_axes, + reduction_axes=reducer_axes, + aggregation_axes=(0,), + num_samples=None, + ) for input_ in TestCollectedStatistics.REF_INPUTS: - collector_obj.register_input(input_) + collector_obj.register_input_for_all_reducers(PTNNCFTensor(input_)) test_stats = collector_obj.get_statistics() - assert reduction_shapes_vs_ref_statistic[shapes] == test_stats - - COLLECTORS = [ - partial(PTMinMaxStatisticCollector, use_abs_max=False, output_shape=(1,)), - partial( - PTMixedMinMaxStatisticCollector, - use_per_sample_stats=False, - use_abs_max=False, - use_means_of_mins=False, - use_means_of_maxs=False, - output_shape=(1,), - ), - partial(PTMeanMinMaxStatisticCollector, use_per_sample_stats=False, use_abs_max=False, output_shape=(1,)), - PTMedianMADStatisticCollector, - partial(PTPercentileStatisticCollector, percentiles_to_collect=[10.0]), - partial(PTMeanPercentileStatisticCollector, percentiles_to_collect=[10.0]), - ] - - @pytest.fixture(params=COLLECTORS) - def collector_for_interface_test(self, request): - collector_type = request.param - return collector_type(reduction_shape=(1,)) - - def test_collected_samples(self, collector_for_interface_test: TensorStatisticCollectorBase): - for input_ in TestCollectedStatistics.REF_INPUTS: - collector_for_interface_test.register_input(input_) - assert collector_for_interface_test.collected_samples() == len(TestCollectedStatistics.REF_INPUTS) - - def test_reset(self, collector_for_interface_test: TensorStatisticCollectorBase): - for input_ in TestCollectedStatistics.REF_INPUTS: - collector_for_interface_test.register_input(input_) - collector_for_interface_test.reset() - assert collector_for_interface_test.collected_samples() == 0 - with pytest.raises(StatisticsNotCollectedError): - collector_for_interface_test.get_statistics() - - def test_enable_disable(self, collector_for_interface_test: TensorStatisticCollectorBase): - for input_ in TestCollectedStatistics.REF_INPUTS: - collector_for_interface_test.register_input(input_) - - collector_for_interface_test.disable() - for input_ in TestCollectedStatistics.REF_INPUTS: - collector_for_interface_test.register_input(input_) - assert collector_for_interface_test.collected_samples() == len(TestCollectedStatistics.REF_INPUTS) - - collector_for_interface_test.enable() - for input_ in TestCollectedStatistics.REF_INPUTS: - collector_for_interface_test.register_input(input_) - assert collector_for_interface_test.collected_samples() == 2 * len(TestCollectedStatistics.REF_INPUTS) - - OFFLINE_COLLECTORS = [ - partial( - PTMixedMinMaxStatisticCollector, - use_per_sample_stats=False, - use_abs_max=False, - use_means_of_mins=False, - use_means_of_maxs=False, - output_shape=(1,), - ), - partial(PTMeanMinMaxStatisticCollector, use_per_sample_stats=False, use_abs_max=False, output_shape=(1,)), - PTMedianMADStatisticCollector, - partial(PTPercentileStatisticCollector, percentiles_to_collect=[10.0]), - partial(PTMeanPercentileStatisticCollector, percentiles_to_collect=[10.0]), - ] - - REF_NUM_SAMPLES = 3 - - @pytest.fixture(params=OFFLINE_COLLECTORS) - def collector_for_num_samples_test(self, request): - collector_type = request.param - return collector_type(reduction_shape=(1,), num_samples=TestCollectedStatistics.REF_NUM_SAMPLES) - - def test_num_samples(self, collector_for_num_samples_test: OfflineTensorStatisticCollector): - for input_ in TestCollectedStatistics.REF_INPUTS * 10: - collector_for_num_samples_test.register_input(input_) - assert collector_for_num_samples_test.collected_samples() == TestCollectedStatistics.REF_NUM_SAMPLES + assert reduction_axes_vs_ref_statistic[reduction_axes] == test_stats class TestCollectorTensorProcessor: diff --git a/tests/torch/test_statistics_aggregator.py b/tests/torch/test_statistics_aggregator.py index 4a4b8f48914..60bfb99015a 100644 --- a/tests/torch/test_statistics_aggregator.py +++ b/tests/torch/test_statistics_aggregator.py @@ -62,7 +62,7 @@ def get_backend_model(self, dataset_samples): @pytest.fixture def is_backend_support_custom_estimators(self) -> bool: - return False + return True @pytest.fixture(scope="session") def test_params(self):