mct-nightly 2.0.0.20240508.145608__tar.gz → 2.0.0.20240509.406__tar.gz

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240508.145608
+Version: 2.0.0.20240509.406
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN
@@ -51,9 +51,6 @@ Description: # Model Compression Toolkit (MCT)
         * [Post training quantization with PyTorch](tutorials/notebooks/pytorch/ptq/example_pytorch_quantization_mnist.ipynb)
         * [Data Generation for ResNet18 with PyTorch](tutorials/notebooks/pytorch/data_generation/example_pytorch_data_generation.ipynb).
         
-        Additionally, for quick quantization of a variety of models from well-known collections,
-        visit the [quick-start page](tutorials/quick_start/README.md) and the
-        [results CSV](tutorials/quick_start/results/model_quantization_results.csv).
         
         ### Supported Versions
         

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/README.md

@@ -45,9 +45,6 @@ for hands-on learning. For example:
 * [Post training quantization with PyTorch](tutorials/notebooks/pytorch/ptq/example_pytorch_quantization_mnist.ipynb)
 * [Data Generation for ResNet18 with PyTorch](tutorials/notebooks/pytorch/data_generation/example_pytorch_data_generation.ipynb).
 
-Additionally, for quick quantization of a variety of models from well-known collections,
-visit the [quick-start page](tutorials/quick_start/README.md) and the
-[results CSV](tutorials/quick_start/results/model_quantization_results.csv).
 
 ### Supported Versions
 

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/mct_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240508.145608
+Version: 2.0.0.20240509.406
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN
@@ -51,9 +51,6 @@ Description: # Model Compression Toolkit (MCT)
         * [Post training quantization with PyTorch](tutorials/notebooks/pytorch/ptq/example_pytorch_quantization_mnist.ipynb)
         * [Data Generation for ResNet18 with PyTorch](tutorials/notebooks/pytorch/data_generation/example_pytorch_data_generation.ipynb).
         
-        Additionally, for quick quantization of a variety of models from well-known collections,
-        visit the [quick-start page](tutorials/quick_start/README.md) and the
-        [results CSV](tutorials/quick_start/results/model_quantization_results.csv).
         
         ### Supported Versions
         

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/model_compression_toolkit/__init__.py

@@ -27,4 +27,4 @@ from model_compression_toolkit import data_generation
 from model_compression_toolkit import pruning
 from model_compression_toolkit.trainable_infrastructure.keras.load_model import keras_load_quantized_model
 
-__version__ = "2.0.0.20240508.145608"
+__version__ = "2.0.0.20240509.000406"

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/model_compression_toolkit/constants.py

@@ -20,7 +20,6 @@ TENSORFLOW = 'tensorflow'
 PYTORCH = 'pytorch'
 FOUND_TF = importlib.util.find_spec(TENSORFLOW) is not None
 FOUND_TORCH = importlib.util.find_spec("torch") is not None
-FOUND_TORCHVISION = importlib.util.find_spec("torchvision") is not None
 FOUND_ONNX = importlib.util.find_spec("onnx") is not None
 FOUND_ONNXRUNTIME = importlib.util.find_spec("onnxruntime") is not None
 FOUND_SONY_CUSTOM_LAYERS = importlib.util.find_spec('sony_custom_layers') is not None

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/model_compression_toolkit/core/common/visualization/tensorboard_writer.py

@@ -72,19 +72,10 @@ def get_node_properties(node_dict_to_log: dict,
     # Create protobuf for the node's output shapes
     if output_shapes is not None:
         tshape_protos = []
-        is_tf_combined_non_max_suppression = len(output_shapes) == 1 and 'function' in node_dict_to_log and node_dict_to_log['function'] == 'image.combined_non_max_suppression'
-
-        if is_tf_combined_non_max_suppression:
-            combined_nms_output = output_shapes[0]
-            output_shapes = [combined_nms_output.nmsed_boxes,
-                             combined_nms_output.nmsed_scores,
-                             combined_nms_output.nmsed_classes,
-                             combined_nms_output.valid_detections]
-
         for output_shape in output_shapes:  # create protobuf for each output shape
             proto_dims_list = []
             for dim in output_shape:
-                proto_dims_list.append(TensorShapeProto.Dim(size=dim)) # dim shold ne an integer
+                proto_dims_list.append(TensorShapeProto.Dim(size=dim))
             tshape_proto = TensorShapeProto(dim=proto_dims_list)
             tshape_protos.append(tshape_proto)
         node_properties['_output_shapes'] = AttrValue(list=AttrValue.ListValue(shape=tshape_protos))

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/model_compression_toolkit/data_generation/__init__.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 # ==============================================================================
 
-from model_compression_toolkit.constants import FOUND_TORCH, FOUND_TF, FOUND_TORCHVISION
+from model_compression_toolkit.constants import FOUND_TORCH, FOUND_TF
 from model_compression_toolkit.data_generation.common.data_generation_config import DataGenerationConfig
 from model_compression_toolkit.data_generation.common.enums import ImageGranularity, DataInitType, SchedulerType, BNLayerWeightingType, OutputLossType, BatchNormAlignemntLossType, ImagePipelineType, ImageNormalizationType
 
@@ -21,6 +21,6 @@ if FOUND_TF:
     from model_compression_toolkit.data_generation.keras.keras_data_generation import (
         keras_data_generation_experimental, get_keras_data_generation_config)
 
-if FOUND_TORCH and FOUND_TORCHVISION:
+if FOUND_TORCH:
     from model_compression_toolkit.data_generation.pytorch.pytorch_data_generation import (
         pytorch_data_generation_experimental, get_pytorch_data_generation_config)

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/model_compression_toolkit/data_generation/pytorch/pytorch_data_generation.py

@@ -17,7 +17,7 @@ from typing import Callable, Any, Tuple, List
 
 from tqdm import tqdm
 
-from model_compression_toolkit.constants import FOUND_TORCH, FOUND_TORCHVISION
+from model_compression_toolkit.constants import FOUND_TORCH
 from model_compression_toolkit.core.pytorch.utils import set_model
 from model_compression_toolkit.data_generation.common.constants import DEFAULT_N_ITER, DEFAULT_DATA_GEN_BS
 from model_compression_toolkit.data_generation.common.data_generation import get_data_generation_classes
@@ -44,7 +44,7 @@ from model_compression_toolkit.data_generation.pytorch.optimization_functions.sc
 from model_compression_toolkit.data_generation.pytorch.optimization_utils import PytorchImagesOptimizationHandler
 from model_compression_toolkit.logger import Logger
 
-if FOUND_TORCH and FOUND_TORCHVISION:
+if FOUND_TORCH:
     # Importing necessary libraries
     import torch
     from torch import Tensor
@@ -354,9 +354,10 @@ else:
     # If torch is not installed,
     # we raise an exception when trying to use these functions.
     def get_pytorch_data_generation_config(*args, **kwargs):
-        msg = f"torch and torchvision must be installed to use get_pytorch_data_generation_config. " + ("" if FOUND_TORCH else "'torch' package is missing. ") + ("" if FOUND_TORCHVISION else "'torchvision' package is missing. ") # pragma: no cover
-        Logger.critical(msg)  # pragma: no cover
+        Logger.critical('PyTorch must be installed to use get_pytorch_data_generation_config. '
+                        "The 'torch' package is missing.")  # pragma: no cover
+
 
     def pytorch_data_generation_experimental(*args, **kwargs):
-        msg = f"torch and torchvision must be installed to use pytorch_data_generation_experimental. " + ("" if FOUND_TORCH else "'torch' package is missing. ") + ("" if FOUND_TORCHVISION else "'torchvision' package is missing. ") # pragma: no cover
-        Logger.critical(msg)  # pragma: no cover
+        Logger.critical("PyTorch must be installed to use 'pytorch_data_generation_experimental'. "
+                        "The 'torch' package is missing.")  # pragma: no cover

mct-nightly-2.0.0.20240509.406/model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py

@@ -0,0 +1,132 @@
+# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# 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 Callable
+from io import BytesIO
+
+import torch.nn
+import onnx
+
+from mct_quantizers import PytorchActivationQuantizationHolder, PytorchQuantizationWrapper
+from model_compression_toolkit.logger import Logger
+from model_compression_toolkit.core.pytorch.utils import to_torch_tensor
+from model_compression_toolkit.exporter.model_exporter.pytorch.base_pytorch_exporter import BasePyTorchExporter
+from mct_quantizers import pytorch_quantizers
+from mct_quantizers.pytorch.metadata import add_onnx_metadata
+
+DEFAULT_ONNX_OPSET_VERSION=15
+
+
+class FakelyQuantONNXPyTorchExporter(BasePyTorchExporter):
+    """
+    Exporter for fakely-quant PyTorch models.
+    The exporter expects to receive an exportable model (where each layer's full quantization parameters
+    can be retrieved), and convert it into a fakely-quant model (namely, weights that are in fake-quant
+    format) and fake-quant layers for the activations.
+    """
+
+    def __init__(self,
+                 model: torch.nn.Module,
+                 is_layer_exportable_fn: Callable,
+                 save_model_path: str,
+                 repr_dataset: Callable,
+                 use_onnx_custom_quantizer_ops: bool = False,
+                 onnx_opset_version=DEFAULT_ONNX_OPSET_VERSION):
+        """
+
+        Args:
+            model: Model to export.
+            is_layer_exportable_fn: Callable to check whether a layer can be exported or not.
+            save_model_path: Path to save the exported model.
+            repr_dataset: Representative dataset (needed for creating torch script).
+            use_onnx_custom_quantizer_ops: Whether to export quantizers custom ops in ONNX or not.
+            onnx_opset_version: ONNX opset version to use for exported ONNX model.
+        """
+
+        super().__init__(model,
+                         is_layer_exportable_fn,
+                         save_model_path,
+                         repr_dataset)
+
+        self._use_onnx_custom_quantizer_ops = use_onnx_custom_quantizer_ops
+        self._onnx_opset_version = onnx_opset_version
+
+    def export(self) -> None:
+        """
+        Convert an exportable (fully-quantized) PyTorch model to a fakely-quant model
+        (namely, weights that are in fake-quant format) and fake-quant layers for the activations.
+
+        Returns:
+            Fake-quant PyTorch model.
+        """
+        for layer in self.model.children():
+            self.is_layer_exportable_fn(layer)
+
+        # Set forward that is used during onnx export.
+        # If _use_onnx_custom_quantizer_ops is set to True, the quantizer forward function will use
+        # the custom implementation when exporting the operator into onnx model. If not, it removes the
+        # wraps and quantizes the ops in place (for weights, for activation torch quantization function is
+        # exported since it's used during forward).
+        if self._use_onnx_custom_quantizer_ops:
+            self._enable_onnx_custom_ops_export()
+        else:
+            self._substitute_fully_quantized_model()
+
+        if self._use_onnx_custom_quantizer_ops:
+            Logger.info(f"Exporting onnx model with MCTQ quantizers: {self.save_model_path}")
+        else:
+            Logger.info(f"Exporting fake-quant onnx model: {self.save_model_path}")
+
+        model_input = to_torch_tensor(next(self.repr_dataset())[0])
+
+        if hasattr(self.model, 'metadata'):
+            onnx_bytes = BytesIO()
+            torch.onnx.export(self.model,
+                              model_input,
+                              onnx_bytes,
+                              opset_version=self._onnx_opset_version,
+                              verbose=False,
+                              input_names=['input'],
+                              output_names=['output'],
+                              dynamic_axes={'input': {0: 'batch_size'},
+                                            'output': {0: 'batch_size'}})
+            onnx_model = onnx.load_from_string(onnx_bytes.getvalue())
+            onnx_model = add_onnx_metadata(onnx_model, self.model.metadata)
+            onnx.save_model(onnx_model, self.save_model_path)
+        else:
+            torch.onnx.export(self.model,
+                              model_input,
+                              self.save_model_path,
+                              opset_version=self._onnx_opset_version,
+                              verbose=False,
+                              input_names=['input'],
+                              output_names=['output'],
+                              dynamic_axes={'input': {0: 'batch_size'},
+                                            'output': {0: 'batch_size'}})
+
+    def _enable_onnx_custom_ops_export(self):
+        """
+        Enable the custom implementation forward in quantizers, so it is exported
+        with custom quantizers.
+        """
+
+        for n, m in self.model.named_modules():
+            if isinstance(m, PytorchActivationQuantizationHolder):
+                assert isinstance(m.activation_holder_quantizer, pytorch_quantizers.BasePyTorchInferableQuantizer)
+                m.activation_holder_quantizer.enable_custom_impl()
+
+            if isinstance(m, PytorchQuantizationWrapper):
+                for wq in m.weights_quantizers.values():
+                    assert isinstance(wq, pytorch_quantizers.BasePyTorchInferableQuantizer)
+                    wq.enable_custom_impl()

{mct-nightly-2.0.0.20240508.145608 → mct-nightly-2.0.0.20240509.406}/model_compression_toolkit/exporter/model_exporter/pytorch/pytorch_export_facade.py

@@ -14,20 +14,20 @@
 # ==============================================================================
 from typing import Callable
 
-from model_compression_toolkit.constants import FOUND_TORCH, FOUND_ONNX
+from model_compression_toolkit.constants import FOUND_TORCH
 from model_compression_toolkit.exporter.model_exporter.fw_agonstic.quantization_format import QuantizationFormat
 from model_compression_toolkit.exporter.model_exporter.pytorch.export_serialization_format import \
     PytorchExportSerializationFormat
 from model_compression_toolkit.logger import Logger
-
-
-DEFAULT_ONNX_OPSET_VERSION = 15
+from model_compression_toolkit.target_platform_capabilities.target_platform import TargetPlatformCapabilities
 
 
 if FOUND_TORCH:
     import torch.nn
-    from model_compression_toolkit.exporter.model_exporter.pytorch.fakely_quant_onnx_pytorch_exporter import FakelyQuantONNXPyTorchExporter
-    from model_compression_toolkit.exporter.model_exporter.pytorch.fakely_quant_torchscript_pytorch_exporter import FakelyQuantTorchScriptPyTorchExporter
+    from model_compression_toolkit.exporter.model_exporter.pytorch.fakely_quant_onnx_pytorch_exporter import \
+    FakelyQuantONNXPyTorchExporter, DEFAULT_ONNX_OPSET_VERSION
+    from model_compression_toolkit.exporter.model_exporter.pytorch.fakely_quant_torchscript_pytorch_exporter import \
+        FakelyQuantTorchScriptPyTorchExporter
     from model_compression_toolkit.exporter.model_wrapper.pytorch.validate_layer import is_pytorch_layer_exportable
 
     supported_serialization_quantization_export_dict = {

mct-nightly-2.0.0.20240508.145608/model_compression_toolkit/exporter/model_exporter/pytorch/fakely_quant_onnx_pytorch_exporter.py

@@ -1,139 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# 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 Callable
-from io import BytesIO
-
-import torch.nn
-
-from mct_quantizers import PytorchActivationQuantizationHolder, PytorchQuantizationWrapper
-from model_compression_toolkit.constants import FOUND_ONNX
-from model_compression_toolkit.logger import Logger
-from model_compression_toolkit.core.pytorch.utils import to_torch_tensor
-from model_compression_toolkit.exporter.model_exporter.pytorch.base_pytorch_exporter import BasePyTorchExporter
-from mct_quantizers import pytorch_quantizers
-
-
-if FOUND_ONNX:
-    import onnx
-    from mct_quantizers.pytorch.metadata import add_onnx_metadata
-
-    class FakelyQuantONNXPyTorchExporter(BasePyTorchExporter):
-        """
-        Exporter for fakely-quant PyTorch models.
-        The exporter expects to receive an exportable model (where each layer's full quantization parameters
-        can be retrieved), and convert it into a fakely-quant model (namely, weights that are in fake-quant
-        format) and fake-quant layers for the activations.
-        """
-
-        def __init__(self,
-                     model: torch.nn.Module,
-                     is_layer_exportable_fn: Callable,
-                     save_model_path: str,
-                     repr_dataset: Callable,
-                     onnx_opset_version: int,
-                     use_onnx_custom_quantizer_ops: bool = False,):
-            """
-
-            Args:
-                model: Model to export.
-                is_layer_exportable_fn: Callable to check whether a layer can be exported or not.
-                save_model_path: Path to save the exported model.
-                repr_dataset: Representative dataset (needed for creating torch script).
-                onnx_opset_version: ONNX opset version to use for exported ONNX model.
-                use_onnx_custom_quantizer_ops: Whether to export quantizers custom ops in ONNX or not.
-            """
-
-            super().__init__(model,
-                             is_layer_exportable_fn,
-                             save_model_path,
-                             repr_dataset)
-
-            self._use_onnx_custom_quantizer_ops = use_onnx_custom_quantizer_ops
-            self._onnx_opset_version = onnx_opset_version
-
-        def export(self) -> None:
-            """
-            Convert an exportable (fully-quantized) PyTorch model to a fakely-quant model
-            (namely, weights that are in fake-quant format) and fake-quant layers for the activations.
-
-            Returns:
-                Fake-quant PyTorch model.
-            """
-            for layer in self.model.children():
-                self.is_layer_exportable_fn(layer)
-
-            # Set forward that is used during onnx export.
-            # If _use_onnx_custom_quantizer_ops is set to True, the quantizer forward function will use
-            # the custom implementation when exporting the operator into onnx model. If not, it removes the
-            # wraps and quantizes the ops in place (for weights, for activation torch quantization function is
-            # exported since it's used during forward).
-            if self._use_onnx_custom_quantizer_ops:
-                self._enable_onnx_custom_ops_export()
-            else:
-                self._substitute_fully_quantized_model()
-
-            if self._use_onnx_custom_quantizer_ops:
-                Logger.info(f"Exporting onnx model with MCTQ quantizers: {self.save_model_path}")
-            else:
-                Logger.info(f"Exporting fake-quant onnx model: {self.save_model_path}")
-
-            model_input = to_torch_tensor(next(self.repr_dataset())[0])
-
-            if hasattr(self.model, 'metadata'):
-                onnx_bytes = BytesIO()
-                torch.onnx.export(self.model,
-                                  model_input,
-                                  onnx_bytes,
-                                  opset_version=self._onnx_opset_version,
-                                  verbose=False,
-                                  input_names=['input'],
-                                  output_names=['output'],
-                                  dynamic_axes={'input': {0: 'batch_size'},
-                                                'output': {0: 'batch_size'}})
-                onnx_model = onnx.load_from_string(onnx_bytes.getvalue())
-                onnx_model = add_onnx_metadata(onnx_model, self.model.metadata)
-                onnx.save_model(onnx_model, self.save_model_path)
-            else:
-                torch.onnx.export(self.model,
-                                  model_input,
-                                  self.save_model_path,
-                                  opset_version=self._onnx_opset_version,
-                                  verbose=False,
-                                  input_names=['input'],
-                                  output_names=['output'],
-                                  dynamic_axes={'input': {0: 'batch_size'},
-                                                'output': {0: 'batch_size'}})
-
-        def _enable_onnx_custom_ops_export(self):
-            """
-            Enable the custom implementation forward in quantizers, so it is exported
-            with custom quantizers.
-            """
-
-            for n, m in self.model.named_modules():
-                if isinstance(m, PytorchActivationQuantizationHolder):
-                    assert isinstance(m.activation_holder_quantizer, pytorch_quantizers.BasePyTorchInferableQuantizer)
-                    m.activation_holder_quantizer.enable_custom_impl()
-
-                if isinstance(m, PytorchQuantizationWrapper):
-                    for wq in m.weights_quantizers.values():
-                        assert isinstance(wq, pytorch_quantizers.BasePyTorchInferableQuantizer)
-                        wq.enable_custom_impl()
-
-else:
-    def FakelyQuantONNXPyTorchExporter(*args, **kwargs):
-            Logger.critical('Installing onnx is mandatory '
-                            'when using FakelyQuantONNXPyTorchExporter. '
-                            'Could not find onnx package.')  # pragma: no cover