[SSS Review] Bug fixes during verification & Exclude Python 3.9 from support (#140)

Author: yt0705

.github/workflows/publish_release.yml

@@ -10,7 +10,7 @@ jobs:
       - name: Install Python 3
         uses: actions/setup-python@v5
         with:
-          python-version: 3.8
+          python-version: 3.11
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip

.github/workflows/run_tests_suite_pip.yml

@@ -2,10 +2,6 @@ name: Test pip install
 on:
   workflow_dispatch: # Allow manual triggers
     inputs:
-      mct_quantizers_version:
-        description: 'MCT Quantizers version'
-        required: true
-        default: 'v1.6.0'
       python_version:
         description: 'Python version'
         required: false
@@ -24,7 +20,6 @@ jobs:
           python-version: ${{  inputs.python_version  }}
       - name: Install dependencies
         run: |
-          git checkout tags/${{  inputs.mct_quantizers_version  }}
           python -m pip install --upgrade pip
           pip install -r requirements.txt
           pip install twine wheel
@@ -58,5 +53,4 @@ jobs:
           pip install torch==2.6.* torchvision "onnx<1.18" "onnxruntime<1.22" "onnxruntime-extensions<0.14"
       - name: Run Torch Tests
         run: |
-          python -m unittest discover tests/pytorch_tests --verbose
-          
+          python -m unittest discover tests/pytorch_tests --verbose

.github/workflows/run_various_python_versions_tf.yml

@@ -16,7 +16,7 @@ jobs:
     timeout-minutes: 10
     strategy:
       matrix:
-        python-version: ["3.9", "3.10", "3.11"]
+        python-version: ["3.10", "3.11"]
     steps:
       - uses: actions/checkout@v4
       - name: Install Python ${{ matrix.python-version }}

.github/workflows/run_various_python_versions_torch.yml

@@ -16,7 +16,7 @@ jobs:
     timeout-minutes: 10
     strategy:
       matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.10", "3.11", "3.12"]
     steps:
       - uses: actions/checkout@v4
       - name: Install Python ${{ matrix.python-version }}

README.md

@@ -15,22 +15,22 @@ Users can set the quantizers and all the quantization information for each layer
 
 Please note that the quantization wrapper and the quantizers are framework-specific.
 
-<img src="https://github.com/sony/mct_quantizers/raw/main/quantization_infra.png" width="700">
+<img src="https://github.com/SonySemiconductorSolutions/mct-quantization-layers/raw/main/quantization_infra.png" width="700">
 
 ## Quantizers:
 
 The library provides the "Inferable Quantizer" interface for implementing new quantizers. 
-This interface is based on the [`BaseInferableQuantizer`](https://github.com/sony/mct_quantizers/blob/main/mct_quantizers/common/base_inferable_quantizer.py) class, which allows the definition of quantizers used for emulating inference-time quantization.
+This interface is based on the [`BaseInferableQuantizer`](https://github.com/SonySemiconductorSolutions/mct-quantization-layers/blob/main/mct_quantizers/common/base_inferable_quantizer.py) class, which allows the definition of quantizers used for emulating inference-time quantization.
 
 On top of `BaseInferableQuantizer` the library defines a set of framework-specific quantizers for both weights and activations:
-1. [Keras Quantizers](https://github.com/sony/mct_quantizers/tree/main/mct_quantizers/keras/quantizers)
-2. [Pytorch Quantizers](https://github.com/sony/mct_quantizers/tree/main/mct_quantizers/pytorch/quantizers)
+1. [Keras Quantizers](https://github.com/SonySemiconductorSolutions/mct-quantization-layers/tree/main/mct_quantizers/keras/quantizers)
+2. [Pytorch Quantizers](https://github.com/SonySemiconductorSolutions/mct-quantization-layers/tree/main/mct_quantizers/pytorch/quantizers)
 
 ### The mark_quantizer Decorator
 
-The [`@mark_quantizer`](https://github.com/sony/mct_quantizers/blob/main/mct_quantizers/common/base_inferable_quantizer.py) decorator is used to assign each quantizer with static properties that define its task compatibility. Each quantizer class should be decorated with this decorator, which defines the following properties:
- - [`QuantizationTarget`](https://github.com/sony/mct_quantizers/blob/main/mct_quantizers/common/base_inferable_quantizer.py): An Enum that indicates whether the quantizer is intended for weights or activations quantization.
- - [`QuantizationMethod`](https://github.com/sony/mct_quantizers/blob/main/mct_quantizers/common/quant_info.py): A list of quantization methods (Uniform, Symmetric, etc.).
+The [`@mark_quantizer`](https://github.com/SonySemiconductorSolutions/mct-quantization-layers/blob/main/mct_quantizers/common/base_inferable_quantizer.py) decorator is used to assign each quantizer with static properties that define its task compatibility. Each quantizer class should be decorated with this decorator, which defines the following properties:
+ - [`QuantizationTarget`](https://github.com/SonySemiconductorSolutions/mct-quantization-layers/blob/main/mct_quantizers/common/base_inferable_quantizer.py): An Enum that indicates whether the quantizer is intended for weights or activations quantization.
+ - [`QuantizationMethod`](https://github.com/SonySemiconductorSolutions/mct-quantization-layers/blob/main/mct_quantizers/common/quant_info.py): A list of quantization methods (Uniform, Symmetric, etc.).
  - `identifier`: A unique identifier for the quantizer class. This is a helper property that allows the creation of advanced quantizers for specific tasks.
 
 ## Getting Started
@@ -45,16 +45,11 @@ To install the latest stable release of MCT Quantizer from PyPi, run the followi
 pip install mct-quantizers
 ```
 
-If you prefer to use the nightly package (unstable version), you can install it with the following command:
-```
-pip install mct-quantizers-nightly
-```
-
 #### From Source
 To work with the MCT Quantizers source code, follow these steps:
 ```
-git clone https://github.com/sony/mct_quantizers.git
-cd mct_quantizers
+git clone https://github.com/SonySemiconductorSolutions/mct-quantization-layers.git
+cd mct-quantization-layers
 python setup.py install
 ```
 
@@ -68,7 +63,7 @@ For use with Tensorflow, please install the following package:
 For use with PyTorch, please install the following package:
 [torch](https://pytorch.org/)
 
-You can also use the [requirements](https://github.com/sony/mct_quantizers/blob/main/requirements.txt) file to set up your environment.
+You can also use the [requirements](https://github.com/SonySemiconductorSolutions/mct-quantization-layers/blob/main/requirements.txt) file to set up your environment.
 
 ## License
-[Apache License 2.0](https://github.com/sony/mct_quantizers/blob/main/LICENSE.md).
+[Apache License 2.0](https://github.com/SonySemiconductorSolutions/mct-quantization-layers/blob/main/LICENSE.md).

setup.py

@@ -52,5 +52,5 @@ def get_release_arguments():
           "Topic :: Scientific/Engineering :: Artificial Intelligence"
       ],
       install_requires=read_install_requires(),
-      python_requires='>=3.9'
+      python_requires='>=3.10'
       )

tests/compatibility_tests/keras_comp_tests/base_activation_compatibility_test.py

@@ -57,16 +57,34 @@
 
 def _build_model_with_quantization_holder(act_layer, quant_activation_holder, input_shape, model_name):
     inputs = tf.keras.layers.Input(shape=input_shape)
-    x = tf.keras.layers.Conv2D(filters=3, kernel_size=4)(inputs)
+    # If all conv outputs are negative, the ReLU output will be 0, causing a quantizer validation error.
+    # To include positive and negative values ​​in the conv outputs, we set the conv weights to positive values.
+    # Also set the upper half of the input tensor to positive and the lower half to negative.
+    conv = tf.keras.layers.Conv2D(filters=3, kernel_size=4)
+    conv.build(input_shape=input_shape)
+    kernel, bias = conv.get_weights()
+    conv.set_weights([np.abs(kernel), bias])
+    x = conv(inputs)
     act_output = act_layer(x)
     quant_output = quant_activation_holder(act_output)
     return tf.keras.Model(inputs=inputs, outputs=[quant_output, act_output], name=model_name)
 
 
 def _build_model_with_operator_quantization_holder(act_layer, quant_activation_holder, input_shape, model_name):
     inputs = tf.keras.layers.Input(shape=input_shape)
-    x = tf.keras.layers.Conv2D(filters=3, kernel_size=4)(inputs)
-    y = tf.keras.layers.Conv2D(filters=3, kernel_size=4)(inputs)
+    # If all conv outputs are negative, the ReLU output will be 0, causing a quantizer validation error.
+    # To include positive and negative values ​​in the conv outputs, we set the conv weights to positive values.
+    # Also set the upper half of the input tensor to positive and the lower half to negative.
+    conv1 = tf.keras.layers.Conv2D(filters=3, kernel_size=4)
+    conv1.build(input_shape=input_shape)
+    kernel1, bias1 = conv1.get_weights()
+    conv1.set_weights([np.abs(kernel1), bias1])
+    conv2 = tf.keras.layers.Conv2D(filters=3, kernel_size=4)
+    conv2.build(input_shape=input_shape)
+    kernel2, bias2 = conv2.get_weights()
+    conv2.set_weights([np.abs(kernel2), bias2])
+    x = conv1(inputs)
+    y = conv2(inputs)
     act_output = act_layer([x, y])
     quant_output = quant_activation_holder(act_output)
     return tf.keras.Model(inputs=inputs, outputs=[quant_output, act_output], name=model_name)
@@ -98,7 +116,14 @@ def build_and_save_model(self, quantizer, quantizer_params, layer, model_name, i
         self.assertEqual(len(quant_holder_layer), 1)
 
         # Verifying activation quantization after holder
-        output = model(np.random.randn(1, *input_shape))
+        # If all conv outputs are negative, the ReLU output will be 0, causing a quantizer validation error.
+        # To include positive and negative values ​​in the conv outputs, we set the conv weights to positive values.
+        # Also set the upper half of the input tensor to positive and the lower half to negative.
+        rand_inp = np.random.randn(1, *input_shape)
+        sign = np.ones((1, *input_shape))
+        sign[:, rand_inp.shape[1]//2:, :, :] = -1
+        rand_inp = rand_inp * sign
+        output = model(rand_inp)
         self.assertTrue(np.any(output[0] != output[1]), "Expecting activation layer output to be different "
                                                         "from the activation holder layer output, which should be "
                                                         "quantized.")

tests/compatibility_tests/torch_comp_tests/base_activation_compatibility_test.py

@@ -57,7 +57,13 @@ def _build_model_with_quantization_holder(act_layer, quant_activation_holder, in
     class Model(torch.nn.Module):
         def __init__(self):
             super(Model, self).__init__()
+            # If all conv outputs are negative, the ReLU output will be 0, causing a quantizer validation error.
+            # To include positive and negative values ​​in the conv outputs, we set the conv weights to positive values.
+            # Also set the upper half of the input tensor to positive and the lower half to negative.
             self.conv = torch.nn.Conv2d(in_channels=3, out_channels=3, kernel_size=4)
+            with torch.no_grad():
+                weight = self.conv.weight.abs()
+                self.conv.weight.copy_(weight)
             self.act_layer = act_layer
             self.quant_activation_holder = quant_activation_holder
 
@@ -74,8 +80,16 @@ def _build_model_with_operator_quantization_holder(act_layer, quant_activation_h
     class Model(torch.nn.Module):
         def __init__(self):
             super(Model, self).__init__()
+            # If all conv outputs are negative, the ReLU output will be 0, causing a quantizer validation error.
+            # To include positive and negative values ​​in the conv outputs, we set the conv weights to positive values.
+            # Also set the upper half of the input tensor to positive and the lower half to negative.
             self.conv1 = torch.nn.Conv2d(in_channels=3, out_channels=3, kernel_size=4)
             self.conv2 = torch.nn.Conv2d(in_channels=3, out_channels=3, kernel_size=4)
+            with torch.no_grad():
+                weight1 = self.conv1.weight.abs()
+                self.conv1.weight.copy_(weight1)
+                weight2 = self.conv2.weight.abs()
+                self.conv2.weight.copy_(weight2)
             self.act_layer = act_layer
             self.quant_activation_holder = quant_activation_holder
 
@@ -115,7 +129,13 @@ def build_and_save_model(self, quantizer, quantizer_params, layer, model_name, i
         quant_holder_layer = [_l for _, _l in model.named_modules() if isinstance(_l, PytorchActivationQuantizationHolder)]
         self.assertEqual(len(quant_holder_layer), 1)
 
+        # If all conv outputs are negative, the ReLU output will be 0, causing a quantizer validation error.
+        # To include positive and negative values ​​in the conv outputs, we set the conv weights to positive values.
+        # Also set the upper half of the input tensor to positive and the lower half to negative.
         rand_inp = torch.rand(1, *input_shape).to(BaseActivationQuantizerBuildAndSaveTest.device)
+        sign = torch.ones(1, *input_shape)
+        sign[:, :, rand_inp.shape[2]//2:, :] = -1
+        rand_inp = rand_inp * sign
         model = model.to(BaseActivationQuantizerBuildAndSaveTest.device)
 
         # Verifying activation quantization after holder