intel
diff --git a/‎neural_compressor/tensorflow/algorithms/static_quant/keras.py
Copy file name to clipboard
+407-328Lines changed: 407 additions & 328 deletions b/‎neural_compressor/tensorflow/algorithms/static_quant/keras.py
Copy file name to clipboard
+407-328Lines changed: 407 additions & 328 deletions
diff --git a/‎neural_compressor/tensorflow/keras/layers/__init__.py
Copy file name to clipboard
+2-1Lines changed: 2 additions & 1 deletion b/‎neural_compressor/tensorflow/keras/layers/__init__.py
Copy file name to clipboard
+2-1Lines changed: 2 additions & 1 deletion
diff --git a/‎neural_compressor/tensorflow/keras/layers/conv2d.py
Copy file name to clipboard
+81-10Lines changed: 81 additions & 10 deletions b/‎neural_compressor/tensorflow/keras/layers/conv2d.py
Copy file name to clipboard
+81-10Lines changed: 81 additions & 10 deletions
diff --git a/‎neural_compressor/tensorflow/keras/layers/dense.py
Copy file name to clipboard
+65-5Lines changed: 65 additions & 5 deletions b/‎neural_compressor/tensorflow/keras/layers/dense.py
Copy file name to clipboard
+65-5Lines changed: 65 additions & 5 deletions
@@ -1,7 +1,7 @@
 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 #
-# Copyright (c) 2022 Intel Corporation
+# Copyright (c) 2024 Intel Corporation
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -21,3 +21,4 @@
 from neural_compressor.tensorflow.keras.layers.pool2d import QAvgPool2D, QMaxPool2D
 from neural_compressor.tensorflow.keras.layers.quantizer import DeQuantize, FakeQuant, Quantize
 from neural_compressor.tensorflow.keras.layers.separable_conv2d import QSeparableConv2D
+from neural_compressor.tensorflow.keras.layers.layer_initializer import layer_initializer_dict
@@ -1,7 +1,7 @@
 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 #
-# Copyright (c) 2022 Intel Corporation
+# Copyright (c) 2024 Intel Corporation
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -23,11 +23,7 @@
 
 from neural_compressor.tensorflow.utils import version1_gte_version2
 
-if version1_gte_version2(tf.__version__, "2.16.1"):
-    from keras.src.layers.convolutional.base_conv import BaseConv  # pylint: disable=E0401
-
-    Conv = BaseConv
-elif version1_gte_version2(tf.__version__, "2.13.0"):
+if version1_gte_version2(tf.__version__, "2.13.0"):
     from keras.src.layers.convolutional.base_conv import Conv  # pylint: disable=E0401
 else:
     from keras.layers.convolutional.base_conv import Conv  # pylint: disable=E0401
@@ -36,6 +32,7 @@
 class QConv2D(Conv):
     def __init__(
         self,
+        name,
         filters,
         kernel_size,
         strides=(1, 1),
@@ -52,11 +49,12 @@ def __init__(
         activity_regularizer=None,
         kernel_constraint=None,
         bias_constraint=None,
-        min_value=-10000,
-        max_value=10000,
+        min_value=None,
+        max_value=None,
         **kwargs
     ):
         super(QConv2D, self).__init__(
+            name=name,
             rank=2,
             filters=filters,
             kernel_size=kernel_size,
@@ -76,10 +74,17 @@ def __init__(
             bias_constraint=constraints.get(bias_constraint),
             **kwargs
         )
-        self.min_value = json.loads(min_value)
-        self.max_value = json.loads(max_value)
+        self.min_value = min_value
+        self.max_value = max_value
 
     def call(self, inputs):
+        kernel_size = self.kernel.shape[-1]
+
+        if not self.min_value:
+            self.min_value = [-10000] * kernel_size
+        if not self.max_value:
+            self.max_value = [10000] * kernel_size
+
         # add the Q/DQ here
         kernel, _, _ = quantization.quantize(
             self.kernel, self.min_value, self.max_value, tf.qint8, axis=3, mode="SCALED"
@@ -111,3 +116,69 @@ def call(self, inputs):
     @classmethod
     def from_config(cls, config):
         return cls(**config)
+
+
+def initialize_int8_conv2d(fp32_layer):
+    kwargs = fp32_layer.get_config()
+
+    if "name" in kwargs:
+        del kwargs["name"]
+    if "filters" in kwargs:
+        del kwargs["filters"]
+    if "kernel_size" in kwargs:
+        del kwargs["kernel_size"]
+    if "strides" in kwargs:
+        del kwargs["strides"]
+    if "padding" in kwargs:
+        del kwargs["padding"]
+    if "data_format" in kwargs:
+        del kwargs["data_format"]
+    if "dilation_rate" in kwargs:
+        del kwargs["dilation_rate"]
+    if "groups" in kwargs:
+        del kwargs["groups"]
+    if "activation" in kwargs:
+        del kwargs["activation"]
+    if "use_bias" in kwargs:
+        del kwargs["use_bias"]
+    if "kernel_initializer" in kwargs:
+        del kwargs["kernel_initializer"]
+    if "bias_initializer" in kwargs:
+        del kwargs["bias_initializer"]
+    if "kernel_regularizer" in kwargs:
+        del kwargs["kernel_regularizer"]
+    if "activity_regularizer" in kwargs:
+        del kwargs["activity_regularizer"]
+    if "bias_regularizer" in kwargs:
+        del kwargs["bias_regularizer"]
+    if "kernel_constraint" in kwargs:
+        del kwargs["kernel_constraint"]
+    if "bias_constraint" in kwargs:
+        del kwargs["bias_constraint"]
+    if "min_value" in kwargs:
+        del kwargs["min_value"]
+    if "max_value" in kwargs:
+        del kwargs["max_value"]
+
+    return QConv2D(
+        name=fp32_layer.name,
+        filters=fp32_layer.filters,
+        kernel_size=fp32_layer.kernel_size,
+        strides=fp32_layer.strides,
+        padding=fp32_layer.padding,
+        data_format=fp32_layer.data_format,
+        dilation_rate=fp32_layer.dilation_rate,
+        groups=fp32_layer.groups,
+        activation=fp32_layer.activation,
+        use_bias=fp32_layer.use_bias,
+        kernel_initializer=fp32_layer.kernel_initializer,
+        bias_initializer=fp32_layer.bias_initializer,
+        kernel_regularizer=fp32_layer.kernel_regularizer,
+        bias_regularizer=fp32_layer.bias_regularizer,
+        activity_regularizer=fp32_layer.activity_regularizer,
+        kernel_constraint=fp32_layer.kernel_constraint,
+        bias_constraint=fp32_layer.bias_constraint,
+        min_value=fp32_layer.min_value,
+        max_value=fp32_layer.max_value,
+        **kwargs
+    )
@@ -1,7 +1,7 @@
 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 #
-# Copyright (c) 2022 Intel Corporation
+# Copyright (c) 2024 Intel Corporation
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -26,6 +26,7 @@
 class QDense(Dense):
     def __init__(
         self,
+        name,
         units,
         activation=None,
         use_bias=True,
@@ -36,11 +37,12 @@ def __init__(
         activity_regularizer=None,
         kernel_constraint=None,
         bias_constraint=None,
-        min_value=-10000,
-        max_value=10000,
+        min_value=None,
+        max_value=None,
         **kwargs
     ):
         super(QDense, self).__init__(
+            name=name,
             units=units,
             activation=activation,
             use_bias=use_bias,
@@ -53,10 +55,17 @@ def __init__(
             bias_constraint=bias_constraint,
             **kwargs
         )
-        self.min_value = json.loads(min_value)
-        self.max_value = json.loads(max_value)
+        self.min_value = min_value
+        self.max_value = max_value
 
     def call(self, inputs):
+        kernel_size = self.kernel.shape[-1]
+
+        if not self.min_value:
+            self.min_value = [-10000] * kernel_size
+        if not self.max_value:
+            self.max_value = [10000] * kernel_size
+
         # add the Q/DQ here
         kernel, _, _ = quantization.quantize(
             self.kernel,
@@ -66,6 +75,7 @@ def call(self, inputs):
             axis=1,
             mode="SCALED",
         )
+
         kernel = quantization.dequantize(
             kernel,
             self.min_value,
@@ -80,3 +90,53 @@ def call(self, inputs):
         if self.activation is not None:
             outputs = self.activation(outputs)
         return outputs
+
+
+def initialize_int8_dense(fp32_layer):
+    kwargs = fp32_layer.get_config()
+
+    if "name" in kwargs:
+        del kwargs["name"]
+    if "units" in kwargs:
+        del kwargs["units"]
+    if "activation" in kwargs:
+        del kwargs["activation"]
+    if "use_bias" in kwargs:
+        del kwargs["use_bias"]
+    if "kernel_initializer" in kwargs:
+        del kwargs["kernel_initializer"]
+    if "bias_initializer" in kwargs:
+        del kwargs["bias_initializer"]
+    if "kernel_regularizer" in kwargs:
+        del kwargs["kernel_regularizer"]
+    if "activity_regularizer" in kwargs:
+        del kwargs["activity_regularizer"]
+    if "bias_regularizer" in kwargs:
+        del kwargs["bias_regularizer"]
+    if "kernel_constraint" in kwargs:
+        del kwargs["kernel_constraint"]
+    if "bias_constraint" in kwargs:
+        del kwargs["bias_constraint"]
+    if "min_value" in kwargs:
+        del kwargs["min_value"]
+    if "max_value" in kwargs:
+        del kwargs["max_value"]
+
+    q_layer = QDense(
+        name=fp32_layer.name,
+        units=fp32_layer.units,
+        activation=fp32_layer.activation,
+        use_bias=fp32_layer.use_bias,
+        kernel_initializer=fp32_layer.kernel_initializer,
+        bias_initializer=fp32_layer.bias_initializer,
+        kernel_regularizer=fp32_layer.kernel_regularizer,
+        bias_regularizer=fp32_layer.bias_regularizer,
+        activity_regularizer=fp32_layer.activity_regularizer,
+        kernel_constraint=fp32_layer.kernel_constraint,
+        bias_constraint=fp32_layer.bias_constraint,
+        min_value=fp32_layer.min_value,
+        max_value=fp32_layer.max_value,
+        **kwargs
+    )
+
+    return q_layer
Original file line number	Diff line number	Diff line change
`@@ -1,7 +1,7 @@`
`1`	`1`	`#!/usr/bin/env python`
`2`	`2`	`# -- coding: utf-8 --`
`3`	`3`	`#`
`4`		`-# Copyright (c) 2022 Intel Corporation`
	`4`	`+# Copyright (c) 2024 Intel Corporation`
`5`	`5`	`#`
`6`	`6`	`# Licensed under the Apache License, Version 2.0 (the "License");`
`7`	`7`	`# you may not use this file except in compliance with the License.`
`@@ -21,3 +21,4 @@`
`21`	`21`	`from neural_compressor.tensorflow.keras.layers.pool2d import QAvgPool2D, QMaxPool2D`
`22`	`22`	`from neural_compressor.tensorflow.keras.layers.quantizer import DeQuantize, FakeQuant, Quantize`
`23`	`23`	`from neural_compressor.tensorflow.keras.layers.separable_conv2d import QSeparableConv2D`
	`24`	`+from neural_compressor.tensorflow.keras.layers.layer_initializer import layer_initializer_dict`