mct-nightly 2.0.0.20240522.420__tar.gz → 2.0.0.20240523.418__tar.gz

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240522.420
+Version: 2.0.0.20240523.418
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN
@@ -12,7 +12,7 @@ Description: # Model Compression Toolkit (MCT)
         
         Specifically, this project aims to apply quantization to compress neural networks.
         
-        <img src="docsrc/images/mct_block_diagram.svg" width="10000">
+        <img src="https://github.com/sony/model_optimization/raw/main/docsrc/images/mct_block_diagram.svg" width="10000">
         
         MCT is developed by researchers and engineers working at Sony Semiconductor Israel.
         
@@ -20,12 +20,12 @@ Description: # Model Compression Toolkit (MCT)
         
         ## Table of Contents
         
-        - [Getting Started](#getting-started)
-        - [Supported features](#supported-features)
-        - [Results](#results)
-        - [Troubleshooting](#trouble-shooting)
-        - [Contributions](#contributions)
-        - [License](#license)
+        - [Getting Started](https://github.com/sony/model_optimization?tab=readme-ov-file#getting-started)
+        - [Supported features](https://github.com/sony/model_optimization?tab=readme-ov-file#supported-features)
+        - [Results](https://github.com/sony/model_optimization?tab=readme-ov-file#results)
+        - [Troubleshooting](https://github.com/sony/model_optimization?tab=readme-ov-file#trouble-shooting)
+        - [Contributions](https://github.com/sony/model_optimization?tab=readme-ov-file#contributions)
+        - [License](https://github.com/sony/model_optimization?tab=readme-ov-file#license)
         
         
         ## Getting Started
@@ -39,17 +39,17 @@ Description: # Model Compression Toolkit (MCT)
         pip install model-compression-toolkit
         ```
         
-        For installing the nightly version or installing from source, refer to the [installation guide](INSTALLATION.md).
+        For installing the nightly version or installing from source, refer to the [installation guide](https://github.com/sony/model_optimization/blob/main/INSTALLATION.md).
         
         
         ### Quick start & tutorials 
         
         Explore the Model Compression Toolkit (MCT) through our tutorials, 
-        covering compression techniques for Keras and PyTorch models. Access interactive [notebooks](tutorials/README.md) 
+        covering compression techniques for Keras and PyTorch models. Access interactive [notebooks](https://github.com/sony/model_optimization/blob/main/tutorials/README.md) 
         for hands-on learning. For example:
-        * [Keras MobileNetV2 post training quantization](tutorials/notebooks/imx500_notebooks/keras/example_keras_mobilenetv2_for_imx500.ipynb)
-        * [Post training quantization with PyTorch](tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_ptq_mnist.ipynb)
-        * [Data Generation for ResNet18 with PyTorch](tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_data_generation.ipynb).
+        * [Keras MobileNetV2 post training quantization](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/imx500_notebooks/keras/example_keras_mobilenetv2_for_imx500.ipynb)
+        * [Post training quantization with PyTorch](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_ptq_mnist.ipynb)
+        * [Data Generation for ResNet18 with PyTorch](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_data_generation.ipynb).
         
         
         ### Supported Versions
@@ -73,15 +73,15 @@ Description: # Model Compression Toolkit (MCT)
         ## Supported Features
         MCT offers a range of powerful features to optimize neural network models for efficient deployment. These supported features include:
         
-        ### Data Generation [*](#experimental-features)
+        ### Data Generation [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
         MCT provides tools for generating synthetic images based on the statistics stored in a model's batch normalization layers. These generated images are valuable for various compression tasks where image data is required, such as quantization and pruning. 
-        You can customize data generation configurations to suit your specific needs. [Go to the Data Generation page.](model_compression_toolkit/data_generation/README.md)
+        You can customize data generation configurations to suit your specific needs. [Go to the Data Generation page.](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/data_generation/README.md)
         
         ### Quantization
         MCT supports different quantization methods:
         * Post-training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/keras_post_training_quantization.html), [PyTorch API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/pytorch_post_training_quantization.html)
         * Gradient-based post-training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/keras_gradient_post_training_quantization.html), [PyTorch API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/pytorch_gradient_post_training_quantization.html)
-        * Quantization-aware training (QAT) [*](#experimental-features)
+        * Quantization-aware training (QAT) [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
         
         
         | Quantization Method                           | Complexity | Computational Cost          |
@@ -103,20 +103,20 @@ Description: # Model Compression Toolkit (MCT)
         * <ins>Advanced quantization algorithms:</ins> To prevent a performance degradation some algorithms are applied such as: 
           * <ins>Shift negative correction:</ins> Symmetric activation quantization can hurt the model's performance when some layers output both negative and positive activations, but their range is asymmetric. For more details please visit [1].
           * <ins>Outliers filtering:</ins> Computing z-score for activation statistics to detect and remove outliers.
-        * <ins>Clustering:</ins> Using non-uniform quantization grid to quantize the weights and activations to match their distributions.[*](#experimental-features)
+        * <ins>Clustering:</ins> Using non-uniform quantization grid to quantize the weights and activations to match their distributions.[*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
         * <ins>Mixed-precision search:</ins> Assigning quantization bit-width per layer (for weights/activations), based on the layer's sensitivity to different bit-widths.
         * <ins>Visualization:</ins> You can use TensorBoard to observe useful information for troubleshooting the quantized model's performance (for example, the model in different phases of the quantization, collected statistics, similarity between layers of the float and quantized model and bit-width configuration for mixed-precision quantization). For more details, please read the [visualization documentation](https://sony.github.io/model_optimization/docs/guidelines/visualization.html).   
-        * <ins>Target Platform Capabilities:</ins> The Target Platform Capabilities (TPC) describes the target platform (an edge device with dedicated hardware). For more details, please read the [TPC README](model_compression_toolkit/target_platform_capabilities/README.md).   
+        * <ins>Target Platform Capabilities:</ins> The Target Platform Capabilities (TPC) describes the target platform (an edge device with dedicated hardware). For more details, please read the [TPC README](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/target_platform_capabilities/README.md).   
         
         ### Enhanced Post-Training Quantization (EPTQ)
         As part of the GPTQ we provide an advanced optimization algorithm called EPTQ.
         
         The specifications of the algorithm are detailed in the paper: _"**EPTQ: Enhanced Post-Training Quantization via Label-Free Hessian**"_ [4].
         
-        More details on the how to use EPTQ via MCT can be found in the [EPTQ guidelines](model_compression_toolkit/gptq/README.md).
+        More details on the how to use EPTQ via MCT can be found in the [EPTQ guidelines](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/gptq/README.md).
         
         
-        ### Structured Pruning [*](#experimental-features)
+        ### Structured Pruning [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
         MCT introduces a structured and hardware-aware model pruning.
         This pruning technique is designed to compress models for specific hardware architectures, 
         taking into account the target platform's Single Instruction, Multiple Data (SIMD) capabilities. 
@@ -138,7 +138,7 @@ Description: # Model Compression Toolkit (MCT)
         Graph of [MobileNetV2](https://keras.io/api/applications/mobilenet/) accuracy on ImageNet vs average bit-width of weights, using 
         single-precision quantization, mixed-precision quantization, and mixed-precision quantization with GPTQ.
         
-        <img src="docsrc/images/mbv2_accuracy_graph.png">
+        <img src="https://github.com/sony/model_optimization/raw/main/docsrc/images/mbv2_accuracy_graph.png">
         
         For more results, please see [1]
         
@@ -174,11 +174,11 @@ Description: # Model Compression Toolkit (MCT)
         ## Contributions
         MCT aims at keeping a more up-to-date fork and welcomes contributions from anyone.
         
-        *You will find more information about contributions in the [Contribution guide](CONTRIBUTING.md).
+        *You will find more information about contributions in the [Contribution guide](https://github.com/sony/model_optimization/blob/main/CONTRIBUTING.md).
         
         
         ## License
-        [Apache License 2.0](LICENSE.md).
+        [Apache License 2.0](https://github.com/sony/model_optimization/blob/main/LICENSE.md).
         
         ## References 
         

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/README.md

@@ -6,7 +6,7 @@ This project provides researchers, developers, and engineers tools for optimizin
 
 Specifically, this project aims to apply quantization to compress neural networks.
 
-<img src="docsrc/images/mct_block_diagram.svg" width="10000">
+<img src="https://github.com/sony/model_optimization/raw/main/docsrc/images/mct_block_diagram.svg" width="10000">
 
 MCT is developed by researchers and engineers working at Sony Semiconductor Israel.
 
@@ -14,12 +14,12 @@ MCT is developed by researchers and engineers working at Sony Semiconductor Isra
 
 ## Table of Contents
 
-- [Getting Started](#getting-started)
-- [Supported features](#supported-features)
-- [Results](#results)
-- [Troubleshooting](#trouble-shooting)
-- [Contributions](#contributions)
-- [License](#license)
+- [Getting Started](https://github.com/sony/model_optimization?tab=readme-ov-file#getting-started)
+- [Supported features](https://github.com/sony/model_optimization?tab=readme-ov-file#supported-features)
+- [Results](https://github.com/sony/model_optimization?tab=readme-ov-file#results)
+- [Troubleshooting](https://github.com/sony/model_optimization?tab=readme-ov-file#trouble-shooting)
+- [Contributions](https://github.com/sony/model_optimization?tab=readme-ov-file#contributions)
+- [License](https://github.com/sony/model_optimization?tab=readme-ov-file#license)
 
 
 ## Getting Started
@@ -33,17 +33,17 @@ To install the latest stable release of MCT, run the following command:
 pip install model-compression-toolkit
 ```
 
-For installing the nightly version or installing from source, refer to the [installation guide](INSTALLATION.md).
+For installing the nightly version or installing from source, refer to the [installation guide](https://github.com/sony/model_optimization/blob/main/INSTALLATION.md).
 
 
 ### Quick start & tutorials 
 
 Explore the Model Compression Toolkit (MCT) through our tutorials, 
-covering compression techniques for Keras and PyTorch models. Access interactive [notebooks](tutorials/README.md) 
+covering compression techniques for Keras and PyTorch models. Access interactive [notebooks](https://github.com/sony/model_optimization/blob/main/tutorials/README.md) 
 for hands-on learning. For example:
-* [Keras MobileNetV2 post training quantization](tutorials/notebooks/imx500_notebooks/keras/example_keras_mobilenetv2_for_imx500.ipynb)
-* [Post training quantization with PyTorch](tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_ptq_mnist.ipynb)
-* [Data Generation for ResNet18 with PyTorch](tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_data_generation.ipynb).
+* [Keras MobileNetV2 post training quantization](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/imx500_notebooks/keras/example_keras_mobilenetv2_for_imx500.ipynb)
+* [Post training quantization with PyTorch](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_ptq_mnist.ipynb)
+* [Data Generation for ResNet18 with PyTorch](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_data_generation.ipynb).
 
 
 ### Supported Versions
@@ -67,15 +67,15 @@ Currently, MCT is being tested on various Python, Pytorch and TensorFlow version
 ## Supported Features
 MCT offers a range of powerful features to optimize neural network models for efficient deployment. These supported features include:
 
-### Data Generation [*](#experimental-features)
+### Data Generation [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
 MCT provides tools for generating synthetic images based on the statistics stored in a model's batch normalization layers. These generated images are valuable for various compression tasks where image data is required, such as quantization and pruning. 
-You can customize data generation configurations to suit your specific needs. [Go to the Data Generation page.](model_compression_toolkit/data_generation/README.md)
+You can customize data generation configurations to suit your specific needs. [Go to the Data Generation page.](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/data_generation/README.md)
 
 ### Quantization
 MCT supports different quantization methods:
 * Post-training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/keras_post_training_quantization.html), [PyTorch API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/pytorch_post_training_quantization.html)
 * Gradient-based post-training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/keras_gradient_post_training_quantization.html), [PyTorch API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/pytorch_gradient_post_training_quantization.html)
-* Quantization-aware training (QAT) [*](#experimental-features)
+* Quantization-aware training (QAT) [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
 
 
 | Quantization Method                           | Complexity | Computational Cost          |
@@ -97,20 +97,20 @@ Main features:
 * <ins>Advanced quantization algorithms:</ins> To prevent a performance degradation some algorithms are applied such as: 
   * <ins>Shift negative correction:</ins> Symmetric activation quantization can hurt the model's performance when some layers output both negative and positive activations, but their range is asymmetric. For more details please visit [1].
   * <ins>Outliers filtering:</ins> Computing z-score for activation statistics to detect and remove outliers.
-* <ins>Clustering:</ins> Using non-uniform quantization grid to quantize the weights and activations to match their distributions.[*](#experimental-features)
+* <ins>Clustering:</ins> Using non-uniform quantization grid to quantize the weights and activations to match their distributions.[*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
 * <ins>Mixed-precision search:</ins> Assigning quantization bit-width per layer (for weights/activations), based on the layer's sensitivity to different bit-widths.
 * <ins>Visualization:</ins> You can use TensorBoard to observe useful information for troubleshooting the quantized model's performance (for example, the model in different phases of the quantization, collected statistics, similarity between layers of the float and quantized model and bit-width configuration for mixed-precision quantization). For more details, please read the [visualization documentation](https://sony.github.io/model_optimization/docs/guidelines/visualization.html).   
-* <ins>Target Platform Capabilities:</ins> The Target Platform Capabilities (TPC) describes the target platform (an edge device with dedicated hardware). For more details, please read the [TPC README](model_compression_toolkit/target_platform_capabilities/README.md).   
+* <ins>Target Platform Capabilities:</ins> The Target Platform Capabilities (TPC) describes the target platform (an edge device with dedicated hardware). For more details, please read the [TPC README](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/target_platform_capabilities/README.md).   
 
 ### Enhanced Post-Training Quantization (EPTQ)
 As part of the GPTQ we provide an advanced optimization algorithm called EPTQ.
 
 The specifications of the algorithm are detailed in the paper: _"**EPTQ: Enhanced Post-Training Quantization via Label-Free Hessian**"_ [4].
 
-More details on the how to use EPTQ via MCT can be found in the [EPTQ guidelines](model_compression_toolkit/gptq/README.md).
+More details on the how to use EPTQ via MCT can be found in the [EPTQ guidelines](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/gptq/README.md).
 
 
-### Structured Pruning [*](#experimental-features)
+### Structured Pruning [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
 MCT introduces a structured and hardware-aware model pruning.
 This pruning technique is designed to compress models for specific hardware architectures, 
 taking into account the target platform's Single Instruction, Multiple Data (SIMD) capabilities. 
@@ -132,7 +132,7 @@ For more details, we highly recommend visiting our project website where experim
 Graph of [MobileNetV2](https://keras.io/api/applications/mobilenet/) accuracy on ImageNet vs average bit-width of weights, using 
 single-precision quantization, mixed-precision quantization, and mixed-precision quantization with GPTQ.
 
-<img src="docsrc/images/mbv2_accuracy_graph.png">
+<img src="https://github.com/sony/model_optimization/raw/main/docsrc/images/mbv2_accuracy_graph.png">
 
 For more results, please see [1]
 
@@ -168,11 +168,11 @@ Check out the [FAQ](https://github.com/sony/model_optimization/tree/main/FAQ.md)
 ## Contributions
 MCT aims at keeping a more up-to-date fork and welcomes contributions from anyone.
 
-*You will find more information about contributions in the [Contribution guide](CONTRIBUTING.md).
+*You will find more information about contributions in the [Contribution guide](https://github.com/sony/model_optimization/blob/main/CONTRIBUTING.md).
 
 
 ## License
-[Apache License 2.0](LICENSE.md).
+[Apache License 2.0](https://github.com/sony/model_optimization/blob/main/LICENSE.md).
 
 ## References 
 

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/mct_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240522.420
+Version: 2.0.0.20240523.418
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN
@@ -12,7 +12,7 @@ Description: # Model Compression Toolkit (MCT)
         
         Specifically, this project aims to apply quantization to compress neural networks.
         
-        <img src="docsrc/images/mct_block_diagram.svg" width="10000">
+        <img src="https://github.com/sony/model_optimization/raw/main/docsrc/images/mct_block_diagram.svg" width="10000">
         
         MCT is developed by researchers and engineers working at Sony Semiconductor Israel.
         
@@ -20,12 +20,12 @@ Description: # Model Compression Toolkit (MCT)
         
         ## Table of Contents
         
-        - [Getting Started](#getting-started)
-        - [Supported features](#supported-features)
-        - [Results](#results)
-        - [Troubleshooting](#trouble-shooting)
-        - [Contributions](#contributions)
-        - [License](#license)
+        - [Getting Started](https://github.com/sony/model_optimization?tab=readme-ov-file#getting-started)
+        - [Supported features](https://github.com/sony/model_optimization?tab=readme-ov-file#supported-features)
+        - [Results](https://github.com/sony/model_optimization?tab=readme-ov-file#results)
+        - [Troubleshooting](https://github.com/sony/model_optimization?tab=readme-ov-file#trouble-shooting)
+        - [Contributions](https://github.com/sony/model_optimization?tab=readme-ov-file#contributions)
+        - [License](https://github.com/sony/model_optimization?tab=readme-ov-file#license)
         
         
         ## Getting Started
@@ -39,17 +39,17 @@ Description: # Model Compression Toolkit (MCT)
         pip install model-compression-toolkit
         ```
         
-        For installing the nightly version or installing from source, refer to the [installation guide](INSTALLATION.md).
+        For installing the nightly version or installing from source, refer to the [installation guide](https://github.com/sony/model_optimization/blob/main/INSTALLATION.md).
         
         
         ### Quick start & tutorials 
         
         Explore the Model Compression Toolkit (MCT) through our tutorials, 
-        covering compression techniques for Keras and PyTorch models. Access interactive [notebooks](tutorials/README.md) 
+        covering compression techniques for Keras and PyTorch models. Access interactive [notebooks](https://github.com/sony/model_optimization/blob/main/tutorials/README.md) 
         for hands-on learning. For example:
-        * [Keras MobileNetV2 post training quantization](tutorials/notebooks/imx500_notebooks/keras/example_keras_mobilenetv2_for_imx500.ipynb)
-        * [Post training quantization with PyTorch](tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_ptq_mnist.ipynb)
-        * [Data Generation for ResNet18 with PyTorch](tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_data_generation.ipynb).
+        * [Keras MobileNetV2 post training quantization](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/imx500_notebooks/keras/example_keras_mobilenetv2_for_imx500.ipynb)
+        * [Post training quantization with PyTorch](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_ptq_mnist.ipynb)
+        * [Data Generation for ResNet18 with PyTorch](https://github.com/sony/model_optimization/blob/main/tutorials/notebooks/mct_features_notebooks/pytorch/example_pytorch_data_generation.ipynb).
         
         
         ### Supported Versions
@@ -73,15 +73,15 @@ Description: # Model Compression Toolkit (MCT)
         ## Supported Features
         MCT offers a range of powerful features to optimize neural network models for efficient deployment. These supported features include:
         
-        ### Data Generation [*](#experimental-features)
+        ### Data Generation [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
         MCT provides tools for generating synthetic images based on the statistics stored in a model's batch normalization layers. These generated images are valuable for various compression tasks where image data is required, such as quantization and pruning. 
-        You can customize data generation configurations to suit your specific needs. [Go to the Data Generation page.](model_compression_toolkit/data_generation/README.md)
+        You can customize data generation configurations to suit your specific needs. [Go to the Data Generation page.](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/data_generation/README.md)
         
         ### Quantization
         MCT supports different quantization methods:
         * Post-training quantization (PTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/keras_post_training_quantization.html), [PyTorch API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/pytorch_post_training_quantization.html)
         * Gradient-based post-training quantization (GPTQ): [Keras API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/keras_gradient_post_training_quantization.html), [PyTorch API](https://sony.github.io/model_optimization/docs/api/api_docs/methods/pytorch_gradient_post_training_quantization.html)
-        * Quantization-aware training (QAT) [*](#experimental-features)
+        * Quantization-aware training (QAT) [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
         
         
         | Quantization Method                           | Complexity | Computational Cost          |
@@ -103,20 +103,20 @@ Description: # Model Compression Toolkit (MCT)
         * <ins>Advanced quantization algorithms:</ins> To prevent a performance degradation some algorithms are applied such as: 
           * <ins>Shift negative correction:</ins> Symmetric activation quantization can hurt the model's performance when some layers output both negative and positive activations, but their range is asymmetric. For more details please visit [1].
           * <ins>Outliers filtering:</ins> Computing z-score for activation statistics to detect and remove outliers.
-        * <ins>Clustering:</ins> Using non-uniform quantization grid to quantize the weights and activations to match their distributions.[*](#experimental-features)
+        * <ins>Clustering:</ins> Using non-uniform quantization grid to quantize the weights and activations to match their distributions.[*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
         * <ins>Mixed-precision search:</ins> Assigning quantization bit-width per layer (for weights/activations), based on the layer's sensitivity to different bit-widths.
         * <ins>Visualization:</ins> You can use TensorBoard to observe useful information for troubleshooting the quantized model's performance (for example, the model in different phases of the quantization, collected statistics, similarity between layers of the float and quantized model and bit-width configuration for mixed-precision quantization). For more details, please read the [visualization documentation](https://sony.github.io/model_optimization/docs/guidelines/visualization.html).   
-        * <ins>Target Platform Capabilities:</ins> The Target Platform Capabilities (TPC) describes the target platform (an edge device with dedicated hardware). For more details, please read the [TPC README](model_compression_toolkit/target_platform_capabilities/README.md).   
+        * <ins>Target Platform Capabilities:</ins> The Target Platform Capabilities (TPC) describes the target platform (an edge device with dedicated hardware). For more details, please read the [TPC README](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/target_platform_capabilities/README.md).   
         
         ### Enhanced Post-Training Quantization (EPTQ)
         As part of the GPTQ we provide an advanced optimization algorithm called EPTQ.
         
         The specifications of the algorithm are detailed in the paper: _"**EPTQ: Enhanced Post-Training Quantization via Label-Free Hessian**"_ [4].
         
-        More details on the how to use EPTQ via MCT can be found in the [EPTQ guidelines](model_compression_toolkit/gptq/README.md).
+        More details on the how to use EPTQ via MCT can be found in the [EPTQ guidelines](https://github.com/sony/model_optimization/blob/main/model_compression_toolkit/gptq/README.md).
         
         
-        ### Structured Pruning [*](#experimental-features)
+        ### Structured Pruning [*](https://github.com/sony/model_optimization?tab=readme-ov-file#experimental-features)
         MCT introduces a structured and hardware-aware model pruning.
         This pruning technique is designed to compress models for specific hardware architectures, 
         taking into account the target platform's Single Instruction, Multiple Data (SIMD) capabilities. 
@@ -138,7 +138,7 @@ Description: # Model Compression Toolkit (MCT)
         Graph of [MobileNetV2](https://keras.io/api/applications/mobilenet/) accuracy on ImageNet vs average bit-width of weights, using 
         single-precision quantization, mixed-precision quantization, and mixed-precision quantization with GPTQ.
         
-        <img src="docsrc/images/mbv2_accuracy_graph.png">
+        <img src="https://github.com/sony/model_optimization/raw/main/docsrc/images/mbv2_accuracy_graph.png">
         
         For more results, please see [1]
         
@@ -174,11 +174,11 @@ Description: # Model Compression Toolkit (MCT)
         ## Contributions
         MCT aims at keeping a more up-to-date fork and welcomes contributions from anyone.
         
-        *You will find more information about contributions in the [Contribution guide](CONTRIBUTING.md).
+        *You will find more information about contributions in the [Contribution guide](https://github.com/sony/model_optimization/blob/main/CONTRIBUTING.md).
         
         
         ## License
-        [Apache License 2.0](LICENSE.md).
+        [Apache License 2.0](https://github.com/sony/model_optimization/blob/main/LICENSE.md).
         
         ## References 
         

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/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__ = "2.0.0.20240522.000420"
+__version__ = "2.0.0.20240523.000418"

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/model_compression_toolkit/core/common/framework_implementation.py

@@ -348,13 +348,14 @@ class FrameworkImplementation(ABC):
         raise NotImplemented(f'{self.__class__.__name__} have to implement the '
                              f'framework\'s count_node_for_mixed_precision_interest_points method.')  # pragma: no cover
 
-    def get_node_distance_fn(self, layer_class: type,
+    def get_mp_node_distance_fn(self, layer_class: type,
                              framework_attrs: Dict[str, Any],
                              compute_distance_fn: Callable = None,
-                             axis: int = None) -> Callable:
+                             axis: int = None,
+                             norm_mse: bool = False) -> Callable:
         """
         A mapping between layers' types and a distance function for computing the distance between
-        two tensors (for loss computation purposes). Returns a specific function if node of specific types is
+        two tensors in mixed precision (for loss computation purposes). Returns a specific function if node of specific types is
         given, or a default (normalized MSE) function otherwise.
 
         Args:
@@ -362,12 +363,13 @@ class FrameworkImplementation(ABC):
             framework_attrs: Framework attributes the layer had which the graph node holds.
             compute_distance_fn: An optional distance function to use globally for all nodes.
             axis: The axis on which the operation is preformed (if specified).
+            norm_mse: whether to normalize mse distance function.
 
         Returns: A distance function between two tensors.
         """
 
         raise NotImplemented(f'{self.__class__.__name__} have to implement the '
-                             f'framework\'s get_node_distance_fn method.')  # pragma: no cover
+                             f'framework\'s get_mp_node_distance_fn method.')  # pragma: no cover
 
 
     @abstractmethod

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/model_compression_toolkit/core/common/graph/base_node.py

@@ -238,8 +238,12 @@ class BaseNode:
         """
         for pos, weight in sorted((pos, weight) for pos, weight in self.weights.items()
                                   if isinstance(pos, int)):
-            assert pos <= len(input_tensors), 'Positional weight index mismatch'
-            input_tensors.insert(pos, weight)
+            if pos > len(input_tensors):
+                Logger.critical("The positional weight index cannot exceed the number of input tensors to the node.")  # pragma: no cover
+            # Insert only positional weights that are not subject to quantization. If the positional weight is
+            # subject to quantization, the quantization wrapper inserts the positional weight into the node.
+            if not self.is_weights_quantization_enabled(pos):
+                input_tensors.insert(pos, weight)
 
         return input_tensors
 

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/model_compression_toolkit/core/common/mixed_precision/sensitivity_evaluation.py

@@ -89,10 +89,13 @@ class SensitivityEvaluation:
                                                       fw_impl.count_node_for_mixed_precision_interest_points,
                                                       quant_config.num_interest_points_factor)
 
-        self.ips_distance_fns, self.ips_axis = self._init_metric_points_lists(self.interest_points)
+        # We use normalized MSE when not running hessian-based. For Hessian-based normalized MSE is not needed
+        # beacause hessian weights already do normalization.
+        use_normalized_mse = self.quant_config.use_hessian_based_scores is False
+        self.ips_distance_fns, self.ips_axis = self._init_metric_points_lists(self.interest_points, use_normalized_mse)
 
         self.output_points = get_output_nodes_for_metric(graph)
-        self.out_ps_distance_fns, self.out_ps_axis = self._init_metric_points_lists(self.output_points)
+        self.out_ps_distance_fns, self.out_ps_axis = self._init_metric_points_lists(self.output_points, use_normalized_mse)
 
         # Setting lists with relative position of the interest points
         # and output points in the list of all mp model activation tensors
@@ -128,7 +131,7 @@ class SensitivityEvaluation:
             self.interest_points_hessians = self._compute_hessian_based_scores()
             self.quant_config.distance_weighting_method = lambda d: self.interest_points_hessians
 
-    def _init_metric_points_lists(self, points: List[BaseNode]) -> Tuple[List[Callable], List[int]]:
+    def _init_metric_points_lists(self, points: List[BaseNode], norm_mse: bool = False) -> Tuple[List[Callable], List[int]]:
         """
         Initiates required lists for future use when computing the sensitivity metric.
         Each point on which the metric is computed uses a dedicated distance function based on its type.
@@ -136,6 +139,7 @@ class SensitivityEvaluation:
 
         Args:
             points: The set of nodes in the graph for which we need to initiate the lists.
+            norm_mse: whether to normalize mse distance function.
 
         Returns: A lists with distance functions and an axis list for each node.
 
@@ -144,11 +148,12 @@ class SensitivityEvaluation:
         axis_list = []
         for n in points:
             axis = n.framework_attr.get(AXIS) if not isinstance(n, FunctionalNode) else n.op_call_kwargs.get(AXIS)
-            distance_fn = self.fw_impl.get_node_distance_fn(
+            distance_fn = self.fw_impl.get_mp_node_distance_fn(
                 layer_class=n.layer_class,
                 framework_attrs=n.framework_attr,
                 compute_distance_fn=self.quant_config.compute_distance_fn,
-                axis=axis)
+                axis=axis,
+                norm_mse=norm_mse)
             distance_fns_list.append(distance_fn)
             # Axis is needed only for KL Divergence calculation, otherwise we use per-tensor computation
             axis_list.append(axis if distance_fn==compute_kl_divergence else None)

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/model_compression_toolkit/core/keras/keras_implementation.py

@@ -421,13 +421,14 @@ class KerasImplementation(FrameworkImplementation):
 
         return False
 
-    def get_node_distance_fn(self, layer_class: type,
+    def get_mp_node_distance_fn(self, layer_class: type,
                              framework_attrs: Dict[str, Any],
                              compute_distance_fn: Callable = None,
-                             axis: int = None) -> Callable:
+                             axis: int = None,
+                             norm_mse: bool = False) -> Callable:
         """
         A mapping between layers' types and a distance function for computing the distance between
-        two tensors (for loss computation purposes). Returns a specific function if node of specific types is
+        two tensors in mixed precision (for loss computation purposes). Returns a specific function if node of specific types is
         given, or a default (normalized MSE) function otherwise.
 
         Args:
@@ -435,6 +436,7 @@ class KerasImplementation(FrameworkImplementation):
             framework_attrs: Framework attributes the layer had which the graph node holds.
             compute_distance_fn: An optional distance function to use globally for all nodes.
             axis: The axis on which the operation is preformed (if specified).
+            norm_mse: whether to normalize mse distance function.
 
         Returns: A distance function between two tensors.
         """
@@ -456,7 +458,7 @@ class KerasImplementation(FrameworkImplementation):
             return compute_cs
         elif layer_class == Dense:
             return compute_cs
-        return compute_mse
+        return partial(compute_mse, norm=norm_mse)
 
     def get_trace_hessian_calculator(self,
                                      graph: Graph,

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/model_compression_toolkit/core/pytorch/back2framework/pytorch_model_builder.py

@@ -67,8 +67,7 @@ def _build_input_tensors_list(node: BaseNode,
             _input_tensors = node_to_output_tensors_dict[ie.source_node]
             input_tensors.append(_input_tensors)
         input_tensors = [tensor for tensor_list in input_tensors for tensor in tensor_list]  # flat list of lists
-        if not is_op_quantize_wrapper:
-            input_tensors = node.insert_positional_weights_to_input_list(input_tensors)
+        input_tensors = node.insert_positional_weights_to_input_list(input_tensors)
         # convert inputs from positional weights (numpy arrays) to tensors. Must handle each element in the
         # list separately, because in FX the tensors are FX objects and fail to_torch_tensor
         input_tensors = [to_torch_tensor(t, numpy_type=t.dtype) if isinstance(t, np.ndarray) else t

{mct-nightly-2.0.0.20240522.420 → mct-nightly-2.0.0.20240523.418}/model_compression_toolkit/core/pytorch/pytorch_implementation.py

@@ -403,13 +403,14 @@ class PytorchImplementation(FrameworkImplementation):
             return True
         return False
 
-    def get_node_distance_fn(self, layer_class: type,
+    def get_mp_node_distance_fn(self, layer_class: type,
                              framework_attrs: Dict[str, Any],
                              compute_distance_fn: Callable = None,
-                             axis: int = None) -> Callable:
+                             axis: int = None,
+                             norm_mse: bool = False) -> Callable:
         """
         A mapping between layers' types and a distance function for computing the distance between
-        two tensors (for loss computation purposes). Returns a specific function if node of specific types is
+        two tensors in mixed precision (for loss computation purposes). Returns a specific function if node of specific types is
         given, or a default (normalized MSE) function otherwise.
 
         Args:
@@ -417,6 +418,7 @@ class PytorchImplementation(FrameworkImplementation):
             framework_attrs: Framework attributes the layer had which the graph node holds.
             compute_distance_fn: An optional distance function to use globally for all nodes.
             axis: The axis on which the operation is preformed (if specified).
+            norm_mse: whether to normalize mse distance function.
 
         Returns: A distance function between two tensors.
         """
@@ -430,7 +432,7 @@ class PytorchImplementation(FrameworkImplementation):
             return compute_cs
         elif layer_class == Linear:
             return compute_cs
-        return compute_mse
+        return partial(compute_mse, norm=norm_mse)
 
     def is_output_node_compatible_for_hessian_score_computation(self,
                                                                 node: BaseNode) -> bool: