mct-nightly 1.8.0.27022023.post430__tar.gz → 1.8.0.27032023.post403__tar.gz

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 1.8.0.27022023.post430
+Version: 1.8.0.27032023.post403
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN
@@ -29,8 +29,8 @@ Description: # Model Compression Toolkit (MCT)
         ## Supported Features
         
         MCT supports different quantization methods:
-        * Post training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_post_training_quantization_experimental.html#ug-keras-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_post_training_quantization_experimental.html#ug-pytorch-post-training-quantization-experimental)
-        * Gradient-based post training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_gradient_post_training_quantization_experimental.html#ug-keras-gradient-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_gradient_post_training_quantization_experimental.html#ug-pytorch-gradient-post-training-quantization-experimental)
+        * Post-training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_post_training_quantization_experimental.html#ug-keras-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_post_training_quantization_experimental.html#ug-pytorch-post-training-quantization-experimental)
+        * Gradient-based post-training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_gradient_post_training_quantization_experimental.html#ug-keras-gradient-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_gradient_post_training_quantization_experimental.html#ug-pytorch-gradient-post-training-quantization-experimental)
         * Quantization aware training (QAT)[*](#experimental-features)
         
         
@@ -87,15 +87,15 @@ Description: # Model Compression Toolkit (MCT)
         pip install mct-nightly
         ```
         
-        ### Requierments
+        ### Requirements
         
-        To run MCT, one of the supported frameworks, Tenosflow/Pytorch, needs to be installed.
+        To run MCT, one of the supported frameworks, Tensorflow/Pytorch, needs to be installed.
         
-        For using with Tensorflow please install the packages: 
+        For use with Tensorflow please install the packages: 
         [tensorflow](https://www.tensorflow.org/install), 
         [tensorflow-model-optimization](https://www.tensorflow.org/model_optimization/guide/install)
         
-        For using with PyTorch please install the packages: 
+        For use with PyTorch please install the packages: 
         [torch](https://pytorch.org/)
         
         Also, a [requirements](requirements.txt) file can be used to set up your environment.

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/README.md

@@ -23,8 +23,8 @@ MCT is developed by researchers and engineers working at Sony Semiconductor Isra
 ## Supported Features
 
 MCT supports different quantization methods:
-* Post training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_post_training_quantization_experimental.html#ug-keras-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_post_training_quantization_experimental.html#ug-pytorch-post-training-quantization-experimental)
-* Gradient-based post training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_gradient_post_training_quantization_experimental.html#ug-keras-gradient-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_gradient_post_training_quantization_experimental.html#ug-pytorch-gradient-post-training-quantization-experimental)
+* Post-training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_post_training_quantization_experimental.html#ug-keras-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_post_training_quantization_experimental.html#ug-pytorch-post-training-quantization-experimental)
+* Gradient-based post-training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_gradient_post_training_quantization_experimental.html#ug-keras-gradient-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_gradient_post_training_quantization_experimental.html#ug-pytorch-gradient-post-training-quantization-experimental)
 * Quantization aware training (QAT)[*](#experimental-features)
 
 
@@ -81,15 +81,15 @@ A nightly package is also available (unstable):
 pip install mct-nightly
 ```
 
-### Requierments
+### Requirements
 
-To run MCT, one of the supported frameworks, Tenosflow/Pytorch, needs to be installed.
+To run MCT, one of the supported frameworks, Tensorflow/Pytorch, needs to be installed.
 
-For using with Tensorflow please install the packages: 
+For use with Tensorflow please install the packages: 
 [tensorflow](https://www.tensorflow.org/install), 
 [tensorflow-model-optimization](https://www.tensorflow.org/model_optimization/guide/install)
 
-For using with PyTorch please install the packages: 
+For use with PyTorch please install the packages: 
 [torch](https://pytorch.org/)
 
 Also, a [requirements](requirements.txt) file can be used to set up your environment.

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/mct_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 1.8.0.27022023.post430
+Version: 1.8.0.27032023.post403
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN
@@ -29,8 +29,8 @@ Description: # Model Compression Toolkit (MCT)
         ## Supported Features
         
         MCT supports different quantization methods:
-        * Post training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_post_training_quantization_experimental.html#ug-keras-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_post_training_quantization_experimental.html#ug-pytorch-post-training-quantization-experimental)
-        * Gradient-based post training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_gradient_post_training_quantization_experimental.html#ug-keras-gradient-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_gradient_post_training_quantization_experimental.html#ug-pytorch-gradient-post-training-quantization-experimental)
+        * Post-training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_post_training_quantization_experimental.html#ug-keras-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_post_training_quantization_experimental.html#ug-pytorch-post-training-quantization-experimental)
+        * Gradient-based post-training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/keras_gradient_post_training_quantization_experimental.html#ug-keras-gradient-post-training-quantization-experimental), [PyTorch API](https://sony.github.io/model_optimization/docs/api/experimental_api_docs/methods/pytorch_gradient_post_training_quantization_experimental.html#ug-pytorch-gradient-post-training-quantization-experimental)
         * Quantization aware training (QAT)[*](#experimental-features)
         
         
@@ -87,15 +87,15 @@ Description: # Model Compression Toolkit (MCT)
         pip install mct-nightly
         ```
         
-        ### Requierments
+        ### Requirements
         
-        To run MCT, one of the supported frameworks, Tenosflow/Pytorch, needs to be installed.
+        To run MCT, one of the supported frameworks, Tensorflow/Pytorch, needs to be installed.
         
-        For using with Tensorflow please install the packages: 
+        For use with Tensorflow please install the packages: 
         [tensorflow](https://www.tensorflow.org/install), 
         [tensorflow-model-optimization](https://www.tensorflow.org/model_optimization/guide/install)
         
-        For using with PyTorch please install the packages: 
+        For use with PyTorch please install the packages: 
         [torch](https://pytorch.org/)
         
         Also, a [requirements](requirements.txt) file can be used to set up your environment.

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/mct_nightly.egg-info/SOURCES.txt

@@ -345,7 +345,6 @@ model_compression_toolkit/gptq/common/__init__.py
 model_compression_toolkit/gptq/common/gptq_config.py
 model_compression_toolkit/gptq/common/gptq_constants.py
 model_compression_toolkit/gptq/common/gptq_graph.py
-model_compression_toolkit/gptq/common/gptq_quantizer_config.py
 model_compression_toolkit/gptq/common/gptq_training.py
 model_compression_toolkit/gptq/keras/__init__.py
 model_compression_toolkit/gptq/keras/gptq_loss.py
@@ -356,7 +355,9 @@ model_compression_toolkit/gptq/keras/quantizer/__init__.py
 model_compression_toolkit/gptq/keras/quantizer/base_keras_gptq_quantizer.py
 model_compression_toolkit/gptq/keras/quantizer/quant_utils.py
 model_compression_toolkit/gptq/keras/quantizer/quantization_builder.py
+model_compression_toolkit/gptq/keras/quantizer/regularization_factory.py
 model_compression_toolkit/gptq/keras/quantizer/soft_rounding/__init__.py
+model_compression_toolkit/gptq/keras/quantizer/soft_rounding/soft_quantizer_reg.py
 model_compression_toolkit/gptq/keras/quantizer/soft_rounding/symmetric_soft_quantizer.py
 model_compression_toolkit/gptq/keras/quantizer/ste_rounding/__init__.py
 model_compression_toolkit/gptq/keras/quantizer/ste_rounding/symmetric_ste.py
@@ -369,6 +370,11 @@ model_compression_toolkit/gptq/pytorch/quantizer/__init__.py
 model_compression_toolkit/gptq/pytorch/quantizer/base_pytorch_gptq_quantizer.py
 model_compression_toolkit/gptq/pytorch/quantizer/quant_utils.py
 model_compression_toolkit/gptq/pytorch/quantizer/quantization_builder.py
+model_compression_toolkit/gptq/pytorch/quantizer/regularization_factory.py
+model_compression_toolkit/gptq/pytorch/quantizer/soft_rounding/__init__.py
+model_compression_toolkit/gptq/pytorch/quantizer/soft_rounding/soft_quantizer_reg.py
+model_compression_toolkit/gptq/pytorch/quantizer/soft_rounding/symmetric_soft_quantizer.py
+model_compression_toolkit/gptq/pytorch/quantizer/soft_rounding/uniform_soft_quantizer.py
 model_compression_toolkit/gptq/pytorch/quantizer/ste_rounding/__init__.py
 model_compression_toolkit/gptq/pytorch/quantizer/ste_rounding/symmetric_ste.py
 model_compression_toolkit/ptq/__init__.py
@@ -402,7 +408,6 @@ model_compression_toolkit/qat/pytorch/quantizer/ste_rounding/uniform_ste.py
 model_compression_toolkit/quantizers_infrastructure/__init__.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/__init__.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/common/__init__.py
-model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/common/activation_lut_pot_inferable_quantizer.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/common/base_inferable_quantizer.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/common/constants.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/common/get_all_subclasses.py
@@ -437,6 +442,7 @@ model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/pyt
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/pytorch/quantizers/base_uniform_inferable_quantizer.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/pytorch/quantizers/constants.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/pytorch/quantizers/activation_inferable_quantizers/__init__.py
+model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/pytorch/quantizers/activation_inferable_quantizers/activation_lut_pot_inferable_quantizer.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/pytorch/quantizers/activation_inferable_quantizers/activation_pot_inferable_quantizer.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/pytorch/quantizers/activation_inferable_quantizers/activation_symmetric_inferable_quantizer.py
 model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/pytorch/quantizers/activation_inferable_quantizers/activation_uniform_inferable_quantizer.py

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/__init__.py

@@ -14,8 +14,6 @@
 # ==============================================================================
 
 from model_compression_toolkit.core.common.quantization.debug_config import DebugConfig
-from model_compression_toolkit.gptq.common.gptq_config import GradientPTQConfig, RoundingType, GradientPTQConfigV2
-from model_compression_toolkit.gptq.common.gptq_quantizer_config import GPTQQuantizerConfig, SoftQuantizerConfig
 from model_compression_toolkit.core.common.quantization import quantization_config
 from model_compression_toolkit.core.common.mixed_precision import mixed_precision_quantization_config
 from model_compression_toolkit.core.common.quantization.quantization_config import QuantizationConfig, \
@@ -36,25 +34,21 @@ from model_compression_toolkit.core.common import network_editors as network_edi
 from model_compression_toolkit.core.keras.quantization_facade import keras_post_training_quantization, \
     keras_post_training_quantization_mixed_precision
 from model_compression_toolkit.ptq.keras.quantization_facade import keras_post_training_quantization_experimental
-from model_compression_toolkit.gptq.keras.quantization_facade import \
-    keras_gradient_post_training_quantization_experimental
-from model_compression_toolkit.gptq.keras.quantization_facade import get_keras_gptq_config
-from model_compression_toolkit.qat.keras.quantization_facade import keras_quantization_aware_training_init, \
-    keras_quantization_aware_training_finalize
-from model_compression_toolkit.qat.pytorch.quantization_facade import pytorch_quantization_aware_training_init, \
-    pytorch_quantization_aware_training_finalize
-from model_compression_toolkit.core.pytorch.quantization_facade import pytorch_post_training_quantization, \
-    pytorch_post_training_quantization_mixed_precision
+from model_compression_toolkit.qat.keras.quantization_facade import keras_quantization_aware_training_init, keras_quantization_aware_training_finalize
+from model_compression_toolkit.qat.pytorch.quantization_facade import pytorch_quantization_aware_training_init, pytorch_quantization_aware_training_finalize
+from model_compression_toolkit.core.pytorch.quantization_facade import pytorch_post_training_quantization, pytorch_post_training_quantization_mixed_precision
 from model_compression_toolkit.ptq.pytorch.quantization_facade import pytorch_post_training_quantization_experimental
-from model_compression_toolkit.gptq.pytorch.quantization_facade import \
-    pytorch_gradient_post_training_quantization_experimental
-from model_compression_toolkit.gptq.pytorch.quantization_facade import get_pytorch_gptq_config
 
 from model_compression_toolkit.core.keras.kpi_data_facade import keras_kpi_data, keras_kpi_data_experimental
 from model_compression_toolkit.core.pytorch.kpi_data_facade import pytorch_kpi_data, pytorch_kpi_data_experimental
 
 from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.load_model import keras_load_quantized_model
 
-from model_compression_toolkit.exporter.model_exporter import tflite_export_model, TFLiteExportMode, keras_export_model, KerasExportMode, pytorch_export_model, PyTorchExportMode
+
+from model_compression_toolkit import exporter
+
+from model_compression_toolkit import gptq
+from model_compression_toolkit.gptq import GradientPTQConfig
+
 
 __version__ = "1.8.0"

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/core/common/logger.py

@@ -17,7 +17,6 @@
 import logging
 import os
 from datetime import datetime
-from os import path
 from pathlib import Path
 
 LOGGER_NAME = 'Constrained Model Optimization'
@@ -43,7 +42,7 @@ class Logger:
 
         """
 
-        if not path.exists(log_path):
+        if not os.path.exists(log_path):
             Path(log_path).mkdir(parents=True, exist_ok=True)
 
     @staticmethod
@@ -93,6 +92,15 @@ class Logger:
 
         print(f'log file is in {log_name}')
 
+    @staticmethod
+    def shutdown():
+        """
+        An orderly command to shutdown by flushing and closing all logging handlers.
+
+        """
+        Logger.LOG_PATH = None
+        logging.shutdown()
+
     ########################################
     # Delegating methods to wrapped logger
     ########################################

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py

@@ -22,6 +22,8 @@ from model_compression_toolkit.core.common import Logger
 from model_compression_toolkit.core.common.mixed_precision.kpi_tools.kpi import KPI, KPITarget
 from model_compression_toolkit.core.common.mixed_precision.mixed_precision_search_manager import MixedPrecisionSearchManager
 
+# Limit ILP solver runtime in seconds
+SOLVER_TIME_LIMIT = 60
 
 def mp_integer_programming_search(search_manager: MixedPrecisionSearchManager,
                                   target_kpi: KPI = None) -> List[int]:
@@ -64,7 +66,10 @@ def mp_integer_programming_search(search_manager: MixedPrecisionSearchManager,
                                     target_kpi,
                                     search_manager)
 
-    lp_problem.solve()  # Try to solve the problem.
+    # Use default PULP solver. Limit runtime in seconds
+    solver = PULP_CBC_CMD(timeLimit=SOLVER_TIME_LIMIT)
+    lp_problem.solve(solver=solver)  # Try to solve the problem.
+
     assert lp_problem.status == LpStatusOptimal, Logger.critical(
         "No solution was found during solving the LP problem")
     Logger.info(LpStatus[lp_problem.status])

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/core/keras/quantization_facade.py

@@ -19,7 +19,7 @@ from model_compression_toolkit.core import common
 from model_compression_toolkit.core.common import Logger
 from model_compression_toolkit.core.common.constants import TENSORFLOW
 from model_compression_toolkit.core.common.user_info import UserInformation
-from model_compression_toolkit.gptq.common.gptq_config import GradientPTQConfig, GradientPTQConfigV2
+from model_compression_toolkit.gptq import GradientPTQConfig, GradientPTQConfigV2
 from model_compression_toolkit.core.common.mixed_precision.kpi_tools.kpi import KPI
 from model_compression_toolkit.core.common.framework_info import FrameworkInfo
 from model_compression_toolkit.core.common.network_editors.actions import EditRule

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/core/pytorch/constants.py

@@ -92,3 +92,7 @@ IN_PROJ_WEIGHT = 'in_proj_weight'
 IN_PROJ_BIAS = 'in_proj_bias'
 BIAS_K = 'bias_k'
 BIAS_V = 'bias_v'
+
+# # Batch size value for 'reshape' and 'view' operators,
+# # the value is -1 so the batch size is inferred from the length of the array and remaining dimensions.
+BATCH_DIM_VALUE = -1

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/multi_head_attention_decomposition.py

@@ -58,21 +58,15 @@ class MHAParams:
 
         # Check if Add Bias KV feature is Active
         if BIAS_K and BIAS_V in mha_node.weights.keys():
-            if mha_node.weights[BIAS_K] and mha_node.weights[BIAS_V] is not None:
+            if mha_node.weights[BIAS_K] is not None and mha_node.weights[BIAS_V] is not None:
                 Logger.error('Add BIAS_KV feature is Not Implemented')  # pragma: no cover
 
         self.embed_dim = mha_node.framework_attr[EMBED_DIM]
         self.num_heads = mha_node.framework_attr[NUM_HEADS]
 
-        if KEY_DIM in mha_node.framework_attr:
-            self.kdim = mha_node.framework_attr[KEY_DIM]
-        else:
-            self.kdim = False
+        self.kdim = mha_node.framework_attr[KEY_DIM]
 
-        if VALUE_DIM in mha_node.framework_attr:
-            self.vdim = mha_node.framework_attr[VALUE_DIM]
-        else:
-            self.vdim = False
+        self.vdim = mha_node.framework_attr[VALUE_DIM]
 
         self.qdim = int(self.embed_dim / self.num_heads)
 
@@ -708,7 +702,7 @@ class MultiHeadAttentionDecomposition(common.BaseSubstitution):
         """
 
         if mha_node.reuse:
-            raise Exception("MCT doesn't support reuse of MultiHeadAttention layer")
+            raise Exception("MCT doesn't support reuse of MultiHeadAttention layer")  # pragma: no cover
         params = MHAParams(mha_node)
 
         # project

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/reshape_with_static_shapes.py

@@ -14,10 +14,13 @@
 # ==============================================================================
 from torch import reshape
 import torch
+
+from model_compression_toolkit.core.common import Logger
 from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
 from model_compression_toolkit.core import common
 from model_compression_toolkit.core.common.graph.base_graph import Graph
 from model_compression_toolkit.core.common.graph.base_node import BaseNode
+from model_compression_toolkit.core.pytorch.constants import BATCH_DIM_VALUE
 
 
 class ReshapeWithStaticShapes(common.BaseSubstitution):
@@ -47,14 +50,25 @@ class ReshapeWithStaticShapes(common.BaseSubstitution):
         Returns:
             Graph after applying the substitution.
         """
+        # we want the batch size value to infer from the length of the array and remaining dimensions
+        if len(node.output_shape) == 1:
+            node.output_shape[0][0] = BATCH_DIM_VALUE
+        else:
+            Logger.error('Reshape or view nodes should have a single output shape')  # pragma: no cover
+
         # configure the new static output shape attribute
         node.op_call_args = node.output_shape
 
         # modify the node input info
         node.input_shape = [node.input_shape[0]]
+
+        # the first input is the tensor to be reshaped, we want his batch size value to infer
+        # from the length of the array and remaining dimensions
+        node.input_shape[0][0] = BATCH_DIM_VALUE
+
         nodes_to_check = []
         for in_edge in graph.incoming_edges(node):
-            if in_edge.sink_index > 0: # the first input is the tensor to be reshaped
+            if in_edge.sink_index > 0:  # the first input is the tensor to be reshaped
                 nodes_to_check.append(in_edge.source_node)
                 graph.remove_edge(in_edge.source_node, node)
         for n in nodes_to_check:
@@ -80,4 +94,4 @@ def clean_graph_from_nodes_without_out_edges(graph: Graph,
             graph.remove_edge(in_edge.source_node, node)
         graph.remove_node(node)
         for n in nodes_to_check:
-            clean_graph_from_nodes_without_out_edges(graph, n)
+            clean_graph_from_nodes_without_out_edges(graph, n)

{mct-nightly-1.8.0.27022023.post430/model_compression_toolkit/exporter/model_exporter → mct-nightly-1.8.0.27032023.post403/model_compression_toolkit/exporter}/__init__.py

@@ -16,3 +16,4 @@
 from model_compression_toolkit.exporter.model_exporter.keras.keras_export_facade import keras_export_model, KerasExportMode
 from model_compression_toolkit.exporter.model_exporter.pytorch.pytorch_export_facade import PyTorchExportMode, pytorch_export_model
 from model_compression_toolkit.exporter.model_exporter.tflite.tflite_export_facade import tflite_export_model, TFLiteExportMode
+

mct-nightly-1.8.0.27032023.post403/model_compression_toolkit/exporter/model_exporter/__init__.py

@@ -0,0 +1,15 @@
+# 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.
+# ==============================================================================
+

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/exporter/model_exporter/tflite/fakely_quant_tflite_exporter.py

@@ -19,7 +19,7 @@ from typing import Callable
 import keras.models
 import tensorflow as tf
 
-from model_compression_toolkit import keras_load_quantized_model
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.load_model import keras_load_quantized_model
 from model_compression_toolkit.core.common import Logger
 from model_compression_toolkit.exporter.model_exporter.keras.fakely_quant_keras_exporter import FakelyQuantKerasExporter
 

mct-nightly-1.8.0.27032023.post403/model_compression_toolkit/exporter/model_wrapper/__init__.py

@@ -0,0 +1,20 @@
+# 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 model_compression_toolkit.exporter.model_wrapper.keras.validate_layer import is_keras_layer_exportable
+from model_compression_toolkit.exporter.model_wrapper.keras.builder.fully_quantized_model_builder import get_exportable_keras_model
+
+from model_compression_toolkit.exporter.model_wrapper.pytorch.validate_layer import is_pytorch_layer_exportable
+from model_compression_toolkit.exporter.model_wrapper.pytorch.builder.fully_quantized_model_builder import get_exportable_pytorch_model

mct-nightly-1.8.0.27032023.post403/model_compression_toolkit/exporter/model_wrapper/keras/builder/fully_quantized_model_builder.py

@@ -0,0 +1,65 @@
+# 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 Tuple
+
+
+from model_compression_toolkit import quantizers_infrastructure as qi
+from model_compression_toolkit.core import common
+from model_compression_toolkit.core.common import Graph, Logger
+from model_compression_toolkit.core.common.constants import FOUND_TF
+from model_compression_toolkit.core.common.user_info import UserInformation
+
+if FOUND_TF:
+    import tensorflow as tf
+    from tensorflow.keras.layers import Layer
+    from model_compression_toolkit.core.keras.back2framework.keras_model_builder import KerasModelBuilder
+    from model_compression_toolkit.exporter.model_wrapper.keras.builder.node_to_quantizers import get_quantization_quantizers
+
+    def _get_wrapper(node: common.BaseNode,
+                     layer: Layer) -> qi.KerasQuantizationWrapper:
+        """
+        A function which takes a computational graph node and a keras layer and perform the quantization wrapping
+        Args:
+            n: A node of mct graph.
+            layer: A keras layer
+
+        Returns: Wrapped layer with weights quantizers and activation quantizers
+
+        """
+        weights_quantizers, activation_quantizers = get_quantization_quantizers(node)
+        return qi.KerasQuantizationWrapper(layer, weights_quantizers, activation_quantizers)
+
+
+    def get_exportable_keras_model(graph: Graph) -> Tuple[tf.keras.models.Model, UserInformation]:
+        """
+        Convert graph to an exportable Keras model (model with all quantization parameters).
+        An exportable model can then be exported using model_exporter, to retrieve the
+        final exported model.
+
+        Args:
+            graph: Graph to convert to an exportable Keras model.
+
+        Returns:
+            Exportable Keras model and user information.
+        """
+        exportable_model, user_info = KerasModelBuilder(graph=graph,
+                                             wrapper=_get_wrapper).build_model()
+        exportable_model.trainable = False
+        return exportable_model, user_info
+else:
+    def get_exportable_keras_model(*args, **kwargs):  # pragma: no cover
+        Logger.error('Installing tensorflow and tensorflow_model_optimization is mandatory '
+                     'when using get_exportable_keras_model. '
+                     'Could not find Tensorflow package.')

{mct-nightly-1.8.0.27022023.post430 → mct-nightly-1.8.0.27032023.post403}/model_compression_toolkit/exporter/model_wrapper/keras/builder/node_to_quantizer.py

@@ -15,15 +15,12 @@
 from typing import Dict, Any
 
 from model_compression_toolkit.core.common import BaseNode, Logger
-from model_compression_toolkit.core.common.constants import THRESHOLD, RANGE_MIN, RANGE_MAX, SIGNED
+from model_compression_toolkit.core.common.constants import THRESHOLD, RANGE_MIN, RANGE_MAX, SIGNED, CLUSTER_CENTERS, SCALE_PER_CHANNEL
 from model_compression_toolkit.core.common.target_platform import QuantizationMethod
 from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.base_inferable_quantizer import QuantizationTarget
-from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.get_quantizers import \
-    get_inferable_quantizer_class
-from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.quantizers.base_keras_inferable_quantizer \
-    import \
-    BaseKerasInferableQuantizer
-
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.get_quantizers import get_inferable_quantizer_class
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.quantizers.base_keras_inferable_quantizer import BaseKerasInferableQuantizer
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.quantizers import constants as qi_keras_consts
 
 def get_inferable_quantizer_kwargs(node: BaseNode,
                                    quantization_target: QuantizationTarget) -> Dict[str, Any]:
@@ -44,19 +41,29 @@ def get_inferable_quantizer_kwargs(node: BaseNode,
         # Return the appropriate quantization parameters based on the quantization method
         if quantization_method in [QuantizationMethod.POWER_OF_TWO,
                                    QuantizationMethod.SYMMETRIC]:
-            return {'num_bits': node_w_qc.weights_n_bits,
-                    'threshold': list(node_w_qc.weights_quantization_params[THRESHOLD].flatten()),
-                    'per_channel': node_w_qc.weights_per_channel_threshold,
-                    'channel_axis': node_w_qc.weights_channels_axis,
-                    'input_rank': len(node_w_qc.weights_quantization_params[THRESHOLD].shape)}
+            return {qi_keras_consts.NUM_BITS: node_w_qc.weights_n_bits,
+                    qi_keras_consts.THRESHOLD: list(node_w_qc.weights_quantization_params[THRESHOLD].flatten()),
+                    qi_keras_consts.PER_CHANNEL: node_w_qc.weights_per_channel_threshold,
+                    qi_keras_consts.CHANNEL_AXIS: node_w_qc.weights_channels_axis,
+                    qi_keras_consts.INPUT_RANK: len(node_w_qc.weights_quantization_params[THRESHOLD].shape)}
 
         elif quantization_method in [QuantizationMethod.UNIFORM]:
-            return {'num_bits': node_w_qc.weights_n_bits,
-                    'per_channel': node_w_qc.weights_per_channel_threshold,
-                    'min_range': list(node_w_qc.weights_quantization_params[RANGE_MIN].flatten()),
-                    'max_range': list(node_w_qc.weights_quantization_params[RANGE_MAX].flatten()),
-                    'channel_axis': node_w_qc.weights_channels_axis,
-                    'input_rank': len(node_w_qc.weights_quantization_params[THRESHOLD].shape)}
+            return {qi_keras_consts.NUM_BITS: node_w_qc.weights_n_bits,
+                    qi_keras_consts.PER_CHANNEL: node_w_qc.weights_per_channel_threshold,
+                    qi_keras_consts.MIN_RANGE: list(node_w_qc.weights_quantization_params[RANGE_MIN].flatten()),
+                    qi_keras_consts.MAX_RANGE: list(node_w_qc.weights_quantization_params[RANGE_MAX].flatten()),
+                    qi_keras_consts.CHANNEL_AXIS: node_w_qc.weights_channels_axis,
+                    qi_keras_consts.INPUT_RANK: len(node_w_qc.weights_quantization_params[RANGE_MIN].shape)}
+
+        elif quantization_method in [QuantizationMethod.LUT_SYM_QUANTIZER, QuantizationMethod.LUT_POT_QUANTIZER]:
+            return {qi_keras_consts.NUM_BITS: node_w_qc.weights_n_bits,
+                    qi_keras_consts.PER_CHANNEL: node_w_qc.weights_per_channel_threshold,
+                    qi_keras_consts.CLUSTER_CENTERS: node_w_qc.weights_quantization_params[CLUSTER_CENTERS],
+                    qi_keras_consts.THRESHOLD: list(node_w_qc.weights_quantization_params[SCALE_PER_CHANNEL].flatten()),
+                    qi_keras_consts.CHANNEL_AXIS: node_w_qc.weights_channels_axis,
+                    # TODO: how to pass multiplier nbits and eps for a specific node?
+                    qi_keras_consts.INPUT_RANK: len(node_w_qc.weights_quantization_params[SCALE_PER_CHANNEL].shape)}
+
         else:
             Logger.critical(f'Not supported quantization method for inferable quantizers.')  # pragma: no cover
 
@@ -68,16 +75,24 @@ def get_inferable_quantizer_kwargs(node: BaseNode,
         # Return the appropriate quantization parameters based on the quantization method
         if quantization_method in [QuantizationMethod.POWER_OF_TWO,
                                    QuantizationMethod.SYMMETRIC]:
-            return {'num_bits': node_qc.activation_n_bits,
+            return {qi_keras_consts.NUM_BITS: node_qc.activation_n_bits,
                     # In activation quantization is per-tensor only - thus we hold the threshold as a list with a len of 1
-                    'threshold': [node_qc.activation_quantization_params[THRESHOLD]],
-                    'signed': node_qc.activation_quantization_params[SIGNED]}
+                    qi_keras_consts.THRESHOLD: [node_qc.activation_quantization_params[THRESHOLD]],
+                    qi_keras_consts.SIGNED: node_qc.activation_quantization_params[SIGNED]}
 
         elif quantization_method in [QuantizationMethod.UNIFORM]:
-            return {'num_bits': node_qc.activation_n_bits,
+            return {qi_keras_consts.NUM_BITS: node_qc.activation_n_bits,
                     # In activation quantization is per-tensor only - thus we hold the min/max as a list with a len of 1
-                    'min_range': [node_qc.activation_quantization_params[RANGE_MIN]],
-                    'max_range': [node_qc.activation_quantization_params[RANGE_MAX]]}
+                    qi_keras_consts.MIN_RANGE: [node_qc.activation_quantization_params[RANGE_MIN]],
+                    qi_keras_consts.MAX_RANGE: [node_qc.activation_quantization_params[RANGE_MAX]]}
+
+        elif quantization_method in [QuantizationMethod.LUT_POT_QUANTIZER]:
+            return {qi_keras_consts.NUM_BITS: node_qc.activation_n_bits,
+                    qi_keras_consts.SIGNED: node_qc.activation_quantization_params[SIGNED],
+                    qi_keras_consts.CLUSTER_CENTERS: node_qc.activation_quantization_params[CLUSTER_CENTERS],
+                    qi_keras_consts.THRESHOLD: [node_qc.activation_quantization_params[THRESHOLD]]
+                    # TODO: how to pass multiplier nbits and eps for a specific node?
+                    }
         else:
             Logger.critical(f'Not supported quantization method for inferable quantizers.')  # pragma: no cover
     else: