Merge remote-tracking branch 'origin/develop' into nl/compress_ingore…

…d_scope

README.md

@@ -375,7 +375,7 @@ You may also use one of the Dockerfiles in the [docker](./docker) directory to b
   - ONNX\* ~=1.13.1
   - OpenVINO\* >=2022.3.0
 
-This repository is tested on Python* 3.8.10, PyTorch* 2.0.1 (NVidia CUDA\* Toolkit 11.7) and TensorFlow* 2.12.1 (NVidia CUDA\* Toolkit 11.8).
+This repository is tested on Python* 3.8.10, PyTorch* 2.0.1 (NVidia CUDA\* Toolkit 11.8) and TensorFlow* 2.12.1 (NVidia CUDA\* Toolkit 11.8).
 
 ## NNCF Compressed Model Zoo
 

docs/Installation.md

@@ -34,7 +34,7 @@ Use the same `pip install` syntax as above to install NNCF along with the backen
 pip install .[<BACKEND>]
 ```
 
-List of  supported backends: `torch`, `tf`, `onnx` and `openvino`.
+List of supported backends: `torch`, `tf`, `onnx` and `openvino`.
 
 For development purposes install extra packages by
 
@@ -61,3 +61,14 @@ Note that in order for this to work for pip versions >= 21.3, your Git version m
 ## As a Docker image
 
 Use one of the Dockerfiles in the [docker](../docker) directory to build an image with an environment already set up and ready for running NNCF [sample scripts](../README.md#model-compression-samples).
+
+## Corresponding versions
+
+The following table lists the recommended corresponding versions of backend packages
+as well as the supported versions of Python:
+
+| NNCF    | OpenVINO   | PyTorch  | ONNX     | TensorFlow | Python  |
+| ------- | ---------- | -------- | -------- | ---------- | ------- |
+| `2.6.0` | `2023.1.0` | `2.0.1`  | `1.13.1` | `2.12.0`   | `3.8`   |
+| `2.5.0` | `2023.0.0` | `1.13.1` | `1.13.1` | `2.11.1`   | `3.8`   |
+| `2.4.0` | `2022.1.0` | `1.12.1` | `1.12.0` | `2.8.2`    | `3.8`   |

nncf/common/graph/patterns/patterns.py

@@ -287,6 +287,8 @@ class HWFusedPatternNames(Enum):
 
     # ATOMIC OPERATIONS
     L2_NORM = PatternDesc("l2_norm")
+    MVN = PatternDesc("mvn")
+    GELU = PatternDesc("gelu")
 
     # BLOCK PATTERNS
     ADD_SCALE_SHIFT_OUTPUT = PatternDesc("add_scale_shift_output")
@@ -338,6 +340,7 @@ class HWFusedPatternNames(Enum):
     LINEAR_ACTIVATIONS_BATCH_NORM = PatternDesc("linear_activations_batch_norm")
     LINEAR_ACTIVATIONS_SCALE_SHIFT = PatternDesc("linear_activations_scale_shift")
     LINEAR_ARITHMETIC = PatternDesc("linear_arithmetic")
+    LINEAR_SHIFT_SCALE = PatternDesc("linear_shift_scale")
     LINEAR_ARITHMETIC_ACTIVATIONS = PatternDesc("linear_arithmetic_activations")
     # Found in PicoDet models
     LINEAR_ARITHMETIC_ACTIVATIONS_ARITHMETIC = PatternDesc("linear_arithmetic_activations_arithmetic")

nncf/onnx/graph/metatypes/onnx_metatypes.py

@@ -617,6 +617,12 @@ class ONNXDeformableConvolutionMetatype(ONNXOpMetatype):
     op_names = ["DeformConv"]
 
 
+@ONNX_OPERATION_METATYPES.register()
+class ONNXErfMetatype(ONNXOpMetatype):
+    name = "ErfOp"
+    op_names = ["Erf"]
+
+
 def get_operator_metatypes() -> List[Type[OperatorMetatype]]:
     """
     Returns a list of the operator metatypes.

nncf/onnx/hardware/fused_patterns.py

@@ -24,6 +24,93 @@
 # BLOCK PATTERNS
 
 
+@ONNX_HW_FUSED_PATTERNS.register(HWFusedPatternNames.MVN)
+def create_mvn() -> GraphPattern:
+    pattern = GraphPattern()
+    pattern_input_node = pattern.add_node(
+        **{GraphPattern.LABEL_ATTR: "*INPUT_NODE*", GraphPattern.METATYPE_ATTR: GraphPattern.NON_PATTERN_NODE_TYPE}
+    )
+    reduce_mean_node_1 = pattern.add_node(
+        **{GraphPattern.LABEL_ATTR: "REDUCE_MEAN_1", GraphPattern.METATYPE_ATTR: om.ONNXReduceMeanMetatype}
+    )
+    sub_node = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "SUBTRACT",
+            GraphPattern.METATYPE_ATTR: [om.ONNXSubMetatype],
+        }
+    )
+    pow_node = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "POW",
+            GraphPattern.METATYPE_ATTR: [om.ONNXPowMetatype],
+        }
+    )
+    reduce_mean_node_2 = pattern.add_node(
+        **{GraphPattern.LABEL_ATTR: "REDUCE_MEAN_2", GraphPattern.METATYPE_ATTR: om.ONNXReduceMeanMetatype}
+    )
+    add_node = pattern.add_node(**{GraphPattern.LABEL_ATTR: "ADD", GraphPattern.METATYPE_ATTR: om.ONNXAddLayerMetatype})
+    sqrt_node = pattern.add_node(**{GraphPattern.LABEL_ATTR: "SQRT", GraphPattern.METATYPE_ATTR: om.ONNXSqrtMetatype})
+    div_node = pattern.add_node(**{GraphPattern.LABEL_ATTR: "DIV", GraphPattern.METATYPE_ATTR: om.ONNXDivLayerMetatype})
+
+    pattern.add_edge(pattern_input_node, reduce_mean_node_1)
+    pattern.add_edge(reduce_mean_node_1, sub_node)
+    pattern.add_edge(pattern_input_node, sub_node)
+    pattern.add_edge(sub_node, pow_node)
+    pattern.add_edge(pow_node, reduce_mean_node_2)
+    pattern.add_edge(reduce_mean_node_2, add_node)
+    pattern.add_edge(add_node, sqrt_node)
+    pattern.add_edge(sqrt_node, div_node)
+    pattern.add_edge(sub_node, div_node)
+    return pattern
+
+
+@ONNX_HW_FUSED_PATTERNS.register(HWFusedPatternNames.MVN_SCALE_SHIFT)
+def create_mvn_scale_shift() -> GraphPattern:
+    mvn = create_mvn()
+    scale_shift = create_scale_shift()
+
+    mvn.join_patterns(scale_shift)
+    return mvn
+
+
+@ONNX_HW_FUSED_PATTERNS.register(HWFusedPatternNames.GELU)
+def create_gelu() -> GraphPattern:
+    pattern = GraphPattern()
+    pattern_input_node = pattern.add_node(
+        **{GraphPattern.LABEL_ATTR: "*INPUT_NODE*", GraphPattern.METATYPE_ATTR: GraphPattern.NON_PATTERN_NODE_TYPE}
+    )
+    div_node = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "DIV",
+            GraphPattern.METATYPE_ATTR: [om.ONNXDivLayerMetatype, om.ONNXMulLayerMetatype],
+        }
+    )
+    erf_node = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "ERF",
+            GraphPattern.METATYPE_ATTR: om.ONNXErfMetatype,
+        }
+    )
+    add_node = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "ADD",
+            GraphPattern.METATYPE_ATTR: [om.ONNXAddLayerMetatype, om.ONNXSubMetatype],
+        }
+    )
+    mul_node = pattern.add_node(
+        **{
+            GraphPattern.LABEL_ATTR: "MUL",
+            GraphPattern.METATYPE_ATTR: [om.ONNXMulLayerMetatype, om.ONNXDivLayerMetatype],
+        }
+    )
+    pattern.add_edge(pattern_input_node, div_node)
+    pattern.add_edge(div_node, erf_node)
+    pattern.add_edge(erf_node, add_node)
+    pattern.add_edge(add_node, mul_node)
+    pattern.add_edge(pattern_input_node, mul_node)
+    return pattern
+
+
 @ONNX_HW_FUSED_PATTERNS.register(HWFusedPatternNames.SCALE_SHIFT)
 def create_scale_shift() -> GraphPattern:
     pattern = GraphPattern()
@@ -372,6 +459,15 @@ def create_linear_arithmetic_activations() -> GraphPattern:
     return linear
 
 
+@ONNX_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_SHIFT_SCALE)
+def create_linear_shift_scale() -> GraphPattern:
+    linear = linear_operations()
+    shift_scale = create_shift_scale()
+
+    linear.join_patterns(shift_scale)
+    return linear
+
+
 @ONNX_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_ARITHMETIC_ACTIVATIONS_ARITHMETIC)
 def create_linear_arithmetic_activations_arithmetic() -> GraphPattern:
     linear_arithmetic_activations = create_linear_arithmetic_activations()
@@ -423,6 +519,9 @@ def atomic_activations_operations() -> GraphPattern:
 
     hswish_without_denominator = create_hswish_without_denominator()
     pattern.add_pattern_alternative(hswish_without_denominator)
+
+    gelu = create_gelu()
+    pattern.add_pattern_alternative(gelu)
     return pattern
 
 

nncf/openvino/hardware/fused_patterns.py

@@ -575,6 +575,14 @@ def create_linear_arithmetic_activations() -> GraphPattern:
     return linear
 
 
+@OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_SHIFT_SCALE)
+def create_linear_shift_scale() -> GraphPattern:
+    linear = linear_operations()
+    shift_scale = create_shift_scale()
+    linear.join_patterns(shift_scale)
+    return linear
+
+
 @OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_ARITHMETIC_ACTIVATIONS_ARITHMETIC)
 def create_linear_arithmetic_activations_arithmetic() -> GraphPattern:
     linear_arithmetic_activations = create_linear_arithmetic_activations()

nncf/openvino/quantization/quantize_model.py

@@ -24,6 +24,7 @@
 from nncf.openvino.quantization.backend_parameters import BackendParameters
 from nncf.openvino.quantization.backend_parameters import is_weight_compression_needed
 from nncf.openvino.quantization.quantize_ifmodel import apply_algorithm_if_bodies
+from nncf.openvino.quantization.weights_compression import insert_pre_compression_operations
 from nncf.parameters import CompressWeightsMode
 from nncf.parameters import DropType
 from nncf.parameters import ModelType

nncf/torch/hardware/fused_patterns.py

@@ -76,6 +76,14 @@ def create_linear_arithmetic_operations() -> GraphPattern:
     return linear
 
 
+@PT_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_SHIFT_SCALE)
+def create_linear_shift_scale() -> GraphPattern:
+    linear = linear_operations()
+    shift_scale = create_shift_scale()
+    linear.join_patterns(shift_scale)
+    return linear
+
+
 @PT_HW_FUSED_PATTERNS.register(HWFusedPatternNames.BATCH_NORM_ACTIVATIONS)
 def create_batch_norm_activations_operations() -> GraphPattern:
     batch_norm = batch_norm_operations()

nncf/torch/tensor_statistics/statistics.py

@@ -9,8 +9,6 @@
 # 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
@@ -20,10 +18,6 @@
 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:

tests/onnx/test_pattern_manager.py

@@ -21,7 +21,6 @@
     HWFusedPatternNames.LINEAR_CONST_MULTIPLY: "Not relevant for ONNX.",
     HWFusedPatternNames.ADD_SCALE_SHIFT_OUTPUT: "Not relevant for ONNX.",
     HWFusedPatternNames.BATCH_INDEX: "Not relevant for ONNX.",
-    HWFusedPatternNames.MVN_SCALE_SHIFT: "Not relevant for ONNX.",
     HWFusedPatternNames.NORMALIZE_L2_MULTIPLY: "Not relevant for ONNX.",
     HWFusedPatternNames.LINEAR_WITH_BIAS: "Linear layers contains biases in ONNX.",
     HWFusedPatternNames.SE_BLOCK: "Not relevant for ONNX.",