Comments

nncf/experimental/common/quantization/algorithms/post_training/algorithm.py → nncf/experimental/quantization/algorithms/post_training/algorithm.py

@@ -16,8 +16,8 @@
 from nncf.common.graph.graph import NNCFGraph
 from nncf.common.tensor_statistics.statistic_point import StatisticPointsContainer
 from nncf.common.utils.backend import BackendType
-from nncf.experimental.common.quantization.algorithms.post_training.pipeline import experimental_create_ptq_pipeline
-from nncf.experimental.common.quantization.algorithms.quantizer.base_quantizer import NNCFQuantizer
+from nncf.experimental.quantization.algorithms.post_training.pipeline import experimental_create_ptq_pipeline
+from nncf.experimental.quantization.algorithms.quantizer.base_quantizer import Quantizer as NNCFQuantizer
 from nncf.quantization.advanced_parameters import AdvancedBiasCorrectionParameters
 from nncf.quantization.advanced_parameters import AdvancedSmoothQuantParameters
 from nncf.quantization.advanced_parameters import RangeEstimatorParameters

nncf/experimental/common/quantization/algorithms/post_training/pipeline.py → nncf/experimental/quantization/algorithms/post_training/pipeline.py

@@ -11,8 +11,8 @@
 
 from typing import Optional, TypeVar
 
-from nncf.experimental.common.quantization.algorithms.quantizer.base_quantizer import NNCFQuantizer
-from nncf.experimental.common.quantization.algorithms.range_estimator.range_estimator import MinMaxRangeEstimator
+from nncf.experimental.quantization.algorithms.quantizer.base_quantizer import Quantizer as NNCFQuantizer
+from nncf.experimental.quantization.algorithms.range_estimator.algorithm import MinMaxRangeEstimator
 from nncf.quantization.advanced_parameters import AdvancedBiasCorrectionParameters
 from nncf.quantization.advanced_parameters import AdvancedSmoothQuantParameters
 from nncf.quantization.advanced_parameters import RangeEstimatorParameters
@@ -35,7 +35,7 @@ def experimental_create_ptq_pipeline(
     smooth_quant_params: Optional[AdvancedSmoothQuantParameters] = None,
     activations_range_estimator_params: Optional[RangeEstimatorParameters] = None,
     weights_range_estimator_params: Optional[RangeEstimatorParameters] = None,
-    batchwise_statistics: bool = None,
+    batchwise_statistics: bool = False,
 ) -> Pipeline:
     """
     Creates an experimental post-training quantization pipeline.

nncf/experimental/quantization/algorithms/quantizer/__init__.py

@@ -0,0 +1,10 @@
+# Copyright (c) 2024 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.

nncf/experimental/common/quantization/algorithms/quantizer/base_quantizer.py → nncf/experimental/quantization/algorithms/quantizer/base_quantizer.py

@@ -18,7 +18,7 @@
 TModel = TypeVar("TModel")
 
 
-class NNCFQuantizer:
+class Quantizer:
     @abstractmethod
     def get_quantization_setup(self, model: TModel, nncf_graph: NNCFGraph) -> SingleConfigQuantizerSetup:
         """

nncf/experimental/common/quantization/algorithms/quantizer/fx_quantizer.py → nncf/experimental/quantization/algorithms/quantizer/fx_quantizer.py

@@ -29,7 +29,7 @@
 from nncf.common.quantization.quantizer_setup import WeightQuantizationInsertionPoint
 from nncf.common.quantization.structs import QuantizationScheme as QuantizationMode
 from nncf.common.quantization.structs import QuantizerConfig
-from nncf.experimental.common.quantization.algorithms.quantizer.base_quantizer import NNCFQuantizer
+from nncf.experimental.quantization.algorithms.quantizer.base_quantizer import Quantizer as NNCFQuantizer
 
 EdgeOrNode = Union[Tuple[torch.fx.Node, torch.fx.Node]]
 

nncf/experimental/common/quantization/algorithms/range_estimator/range_estimator.py → nncf/experimental/quantization/algorithms/range_estimator/algorithm.py

@@ -15,7 +15,7 @@
 from nncf.common.graph.graph import NNCFGraph
 from nncf.common.tensor_statistics.statistic_point import StatisticPointsContainer
 from nncf.common.utils.backend import BackendType
-from nncf.experimental.common.quantization.algorithms.quantizer.base_quantizer import NNCFQuantizer
+from nncf.experimental.quantization.algorithms.quantizer.base_quantizer import Quantizer as NNCFQuantizer
 from nncf.quantization.algorithms.algorithm import Algorithm
 from nncf.quantization.algorithms.min_max.algorithm import MinMaxQuantization
 from nncf.quantization.range_estimator import RangeEstimatorParameters

nncf/experimental/torch/fx/quantization/quantize_pt2e.py

@@ -16,7 +16,6 @@
 import torch.fx
 from torch.ao.quantization.pt2e.duplicate_dq_pass import DuplicateDQPass
 from torch.ao.quantization.pt2e.port_metadata_pass import PortNodeMetaForQDQ
-from torch.ao.quantization.pt2e.qat_utils import _fold_conv_bn_qat
 from torch.ao.quantization.pt2e.utils import _disallow_eval_train
 from torch.ao.quantization.quantizer import Quantizer
 from torch.fx import GraphModule
@@ -26,11 +25,9 @@
 from nncf.common.factory import NNCFGraphFactory
 from nncf.common.logging import nncf_logger
 from nncf.data import Dataset
-from nncf.experimental.common.quantization.algorithms.post_training.algorithm import (
-    ExperimentalPostTrainingQuantization,
-)
-from nncf.experimental.common.quantization.algorithms.quantizer.base_quantizer import NNCFQuantizer
-from nncf.experimental.common.quantization.algorithms.quantizer.fx_quantizer import NNCFFXQuantizer
+from nncf.experimental.quantization.algorithms.post_training.algorithm import ExperimentalPostTrainingQuantization
+from nncf.experimental.quantization.algorithms.quantizer.base_quantizer import Quantizer as NNCFQuantizer
+from nncf.experimental.quantization.algorithms.quantizer.fx_quantizer import NNCFFXQuantizer
 from nncf.experimental.torch.fx.constant_folding import constant_fold
 from nncf.experimental.torch.fx.transformations import QUANTIZE_NODE_TARGETS
 from nncf.experimental.torch.fx.transformations import fuse_conv_bn
@@ -50,7 +47,7 @@ def quantize_pt2e(
     smooth_quant_params: Optional[AdvancedSmoothQuantParameters] = None,
     activations_range_estimator_params: Optional[RangeEstimatorParameters] = None,
     weights_range_estimator_params: Optional[RangeEstimatorParameters] = None,
-    batchwise_statistics: bool = False,
+    batchwise_statistics: Optional[bool] = None,
     fold_quantize: bool = False,
 ) -> torch.fx.GraphModule:
     """
@@ -72,21 +69,17 @@ def quantize_pt2e(
     :param weights_range_estimator_params: Contains parameters for estimating the range
         of weights of the model.
     :param batchwise_statistics: Determines whether quantizer statistics should be calculated
-        for each item of the batch or for the entire batch, default is False.
+        for each item of the batch or for the entire batch, default is None, which means
+        it set True if batch_size > 1 otherwise False.
     :param fold_quantize: Boolean flag for whether fold the quantize op or not.
     """
-    nncf_logger.warning(
-        "Experimental Torch FX quantization backend is being used for the given torch.fx.GraphModule model."
-        " Torch FX PTQ is an experimental feature, consider using Torch or OpenVino PTQ backends"
-        " in case of errors or a poor model performance."
-    )
+    nncf_logger.warning("This is an experimental feature and may change in the future without notice.")
 
     if subset_size < 1:
         raise nncf.ValidationError("Subset size must be positive.")
 
     batch_size = calibration_dataset.get_batch_size()
-    if batch_size is not None and batch_size > 1 and batchwise_statistics is not None:
-        batchwise_statistics = True
+    batchwise_statistics = batchwise_statistics is None and batch_size is not None and batch_size > 1
 
     original_graph_meta = model.meta
 
@@ -121,7 +114,6 @@ def quantize_pt2e(
     # is not preformant
     quantized_model = GraphModule(quantized_model, quantized_model.graph)
 
-    quantized_model = _fold_conv_bn_qat(quantized_model)
     if fold_quantize:
         constant_fold(quantized_model, _quant_node_constraint)
 

nncf/quantization/algorithms/min_max/algorithm.py

@@ -759,7 +759,7 @@ def fill_quantization_target_points(
         """
         Initializes a cache, finds quantization target points and them puts in the cache.
 
-        :param model: Backend-specific model, for which Quantization Target Points are being seek.
+        :param quantizer_setup: Quantization Target Points in format of SingleConfigQuantizerSetup.
         :param nncf_graph: NNCFGraph instance.
         :return: Mapping of quantization target points with associated quantization configuration,
         along with target points for scale unification.