mct-nightly 1.7.1.31122022.post351__py3-none-any.whl → 1.8.0.1042023.post423__py3-none-any.whl

model_compression_toolkit/exporter/model_wrapper/keras/quantizers/fq_quantizer.py

@@ -1,99 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================\
-
-from typing import Dict, Any
-import numpy as np
-import tensorflow as tf
-from keras.engine.base_layer import Layer
-from tensorflow import TensorShape
-from tensorflow_model_optimization.python.core.quantization.keras.quantizers import Quantizer
-
-from model_compression_toolkit.core.common.target_platform import QuantizationMethod
-
-
-class FakeQuantQuantizer(Quantizer):
-    """
-    Quantizer using TensorFlow fake quant layer to quantize activations.
-    """
-
-    def __init__(self,
-                 nbits: int,
-                 min_range: np.ndarray,
-                 max_range: np.ndarray,
-                 quantization_method: QuantizationMethod):
-        """
-
-        Args:
-            nbits: Number of bits to quantize.
-            min_range: Min quantization range.
-            max_range: Max quantization range.
-            quantization_method: Quantization method that is used (POT, Uniform, etc.)
-
-        """
-        self.nbits = nbits
-        self.min_range = tf.Variable(min_range,
-                                     trainable=False,
-                                     dtype=tf.float32)
-        self.max_range = tf.Variable(max_range,
-                                     trainable=False,
-                                     dtype=tf.float32)
-        self.quantization_method = quantization_method
-
-
-    def get_config(self) -> Dict[str, Any]:
-        """
-
-        Returns: Configuration of this FakeQuantQuantizer
-
-        """
-        return {"nbits": self.nbits,
-                "min_range": self.min_range.numpy(),
-                "max_range": self.max_range.numpy(),
-                "quantization_method": self.quantization_method,
-                }
-
-    def build(self, tensor_shape: TensorShape, name: str, layer: Layer) -> dict:
-        """
-        Add variables under layer's scope.
-
-        Args:
-            tensor_shape: Shape of tensor which needs to be quantized.
-            name: Name of tensor.
-            layer: Layer to add variables to.
-
-        Returns:
-            Dictionary with new layer's variables.
-        """
-        return {}
-
-    def __call__(self, inputs, training, weights, **kwargs):
-        """
-        Apply quantization to the input tensor.
-
-        Args:
-            inputs: Input tensor to be quantized.
-            training: Whether the graph is currently training.
-            weights: Dictionary of weights the quantizer can use to quantize the tensor. This contains the weights created in the `build` function.
-            **kwargs: Additional variables which may be passed to the quantizer.
-
-        Returns:
-            Quantized tensor.
-        """
-        with tf.name_scope('FakeQuant'):
-            return tf.quantization.fake_quant_with_min_max_vars(inputs,
-                                                                min=self.min_range,
-                                                                max=self.max_range,
-                                                                num_bits=self.nbits)
-

model_compression_toolkit/exporter/model_wrapper/keras/quantizers/weights_uniform_quantizer.py

@@ -1,105 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-
-import numpy as np
-import tensorflow as tf
-from keras.engine.base_layer import Layer
-from tensorflow import TensorShape
-from tensorflow_model_optimization.python.core.quantization.keras.quantizers import Quantizer
-
-from model_compression_toolkit.core.common.quantization.quantizers.quantizers_helpers import uniform_quantize_tensor, \
-    fix_range_to_include_zero
-from model_compression_toolkit.core.common.target_platform import QuantizationMethod
-
-
-class WeightsUniformQuantizer(Quantizer):
-    """
-    Quantizer for weights quantization.
-    """
-
-    def __init__(self,
-                 nbits: int,
-                 min_range: np.ndarray,
-                 max_range: np.ndarray,
-                 weight: np.ndarray,
-                 quantization_method: QuantizationMethod):
-        """
-
-        Args:
-            nbits: Number of bits to quantize.
-            min_range: Min quantization range.
-            max_range: Max quantization range.
-            weight: Tensor of weights to quantize.
-            quantization_method: Quantization method that is used (POT, Uniform, etc.)
-        """
-
-        super().__init__()
-
-        min_range, max_range = fix_range_to_include_zero(np.array(min_range),
-                                                         np.array(max_range),
-                                                         nbits)
-        self.weight = uniform_quantize_tensor(weight,
-                                              range_min=min_range,
-                                              range_max=max_range,
-                                              n_bits=nbits).astype("float32")
-        self.nbits = nbits
-        self.min_range = min_range
-        self.max_range = max_range
-        self.delta = (self.max_range - self.min_range) / (2 ** self.nbits - 1)
-        self.quantization_method = quantization_method
-
-    def __call__(self, inputs, training, weights, **kwargs):
-        """
-        Apply quantization to the input tensor.
-
-        Args:
-            inputs: Input tensor to be quantized.
-            training: Whether the graph is currently training.
-            weights: Dictionary of weights the quantizer can use to quantize the tensor. This contains the weights created in the `build` function.
-            **kwargs: Additional variables which may be passed to the quantizer.
-
-        Returns:
-            Quantized tensor.
-        """
-        with tf.name_scope('WeightsUniformQuant'):
-            return self.weight
-
-    def get_config(self):
-        """
-
-        Returns: Configuration of this WeightsUniformQuantizer
-
-        """
-        cfg = {"nbits": self.nbits,
-               "min_range": self.min_range,
-               "max_range": self.max_range,
-               "weight": self.weight,
-               "quantization_method": self.quantization_method
-               }
-        return cfg
-
-    def build(self, tensor_shape: TensorShape, name: str, layer: Layer) -> dict:
-        """
-        Add variables under layer's scope.
-
-        Args:
-            tensor_shape: Shape of tensor which needs to be quantized.
-            name: Name of tensor.
-            layer: Layer to add variables to.
-
-        Returns:
-            Dictionary with new layer's variables.
-        """
-        return {}

model_compression_toolkit/exporter/model_wrapper/pytorch/builder/node_to_quantize_config.py

@@ -1,61 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-from model_compression_toolkit.core.common import BaseNode
-from model_compression_toolkit.core.pytorch.back2framework.quantization_wrapper.wrapper_quantize_config import \
-    WrapperQuantizeConfig
-from model_compression_toolkit.exporter.model_wrapper.pytorch.builder.node_to_quantizer import \
-    get_weights_quantizer_for_node, \
-    get_activations_quantizer_for_node
-from model_compression_toolkit.exporter.model_wrapper.pytorch.wrappers_quantize_configs.activation_quantize_config \
-    import \
-    ActivationQuantizeConfig
-from model_compression_toolkit.exporter.model_wrapper.pytorch.wrappers_quantize_configs \
-    .no_quantization_quantize_config import \
-    NoQuantizationQuantizeConfig
-from model_compression_toolkit.exporter.model_wrapper.pytorch.wrappers_quantize_configs \
-    .weights_activation_quantize_config import \
-    WeightsActivationQuantizeConfig
-from model_compression_toolkit.exporter.model_wrapper.pytorch.wrappers_quantize_configs.weights_quantize_config \
-    import \
-    WeightsQuantizeConfig
-
-
-def get_quantization_config(node: BaseNode) -> WrapperQuantizeConfig:
-    """
-    Create a WrapperQuantizeConfig to wrap a layer for its corresponding node.
-
-    Args:
-        node: Node to create a WrapperQuantizeConfig for.
-
-    Returns:
-        WrapperQuantizeConfig to use for wrapping the layer from the passed node.
-
-    """
-
-    if node.is_activation_quantization_enabled() and node.is_weights_quantization_enabled():
-        weight_quantizers = get_weights_quantizer_for_node(node)
-        activation_quantizers = get_activations_quantizer_for_node(node)
-        return WeightsActivationQuantizeConfig(weight_quantizers=weight_quantizers,
-                                               activation_quantizers=activation_quantizers)
-
-    elif not node.is_weights_quantization_enabled() and node.is_activation_quantization_enabled():
-        activation_quantizers = get_activations_quantizer_for_node(node)
-        return ActivationQuantizeConfig(activation_quantizers=activation_quantizers)
-
-    elif not node.is_weights_quantization_enabled() and not node.is_activation_quantization_enabled():
-        return NoQuantizationQuantizeConfig()
-
-    weight_quantizers = get_weights_quantizer_for_node(node)
-    return WeightsQuantizeConfig(weight_quantizers=weight_quantizers)

model_compression_toolkit/exporter/model_wrapper/pytorch/quantizers/fq_quantizer.py

@@ -1,59 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================\
-import numpy as np
-from typing import Dict, Any
-
-from model_compression_toolkit.core.common.constants import RANGE_MIN, RANGE_MAX
-from model_compression_toolkit.core.common.target_platform import QuantizationMethod
-from model_compression_toolkit.core.pytorch.quantizer.fake_quant_builder import uniform_quantization
-
-
-class FakeQuantQuantizer:
-    """
-    Quantizer using TensorFlow fake quant layer to quantize activations.
-    """
-
-    def __init__(self,
-                 nbits: int,
-                 min_range: np.ndarray,
-                 max_range: np.ndarray,
-                 quantization_method: QuantizationMethod):
-        """
-
-        Args:
-            nbits: Number of bits to quantize.
-            min_range: Min quantization range.
-            max_range: Max quantization range.
-            quantization_method: Quantization method that is used (POT, Uniform, etc.)
-
-        """
-        self.nbits = nbits
-        self.min_range = min_range
-        self.max_range = max_range
-        self.quantization_method = quantization_method
-
-    def __call__(self, inputs):
-        """
-        Apply quantization to the input tensor.
-
-        Args:
-            inputs: Input tensor to be quantized.
-
-        Returns:
-            Quantized tensor.
-        """
-        return uniform_quantization(self.nbits,
-                                    {RANGE_MIN: self.min_range,
-                                     RANGE_MAX: self.max_range})(inputs)

model_compression_toolkit/exporter/model_wrapper/pytorch/quantizers/uniform_weights_quantizer.py

@@ -1,67 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-import numpy as np
-import torch
-
-from model_compression_toolkit.core.common.quantization.quantizers.quantizers_helpers import uniform_quantize_tensor
-from model_compression_toolkit.core.common.target_platform import QuantizationMethod
-
-
-class UniformWeightsQuantizer:
-    """
-    Uniform quantizer for a PyTorch weights module.
-    """
-    def __init__(self,
-                 num_bits: int,
-                 min_range: np.ndarray,
-                 max_range: np.ndarray,
-                 quantization_method: QuantizationMethod,
-                 per_channel: bool,
-                 output_channels_axis: int
-                 ):
-        """
-
-        Args:
-            num_bits: Number of bits for quantization.
-            min_range: Min range for quantization.
-            max_range: Max range for quantization.
-            quantization_method: Method the quantized weight was quantized according to.
-            per_channel: Whether the weight was quantized per-channel or not.
-            output_channels_axis: Dimension of channel axis (needed if it was quantized per-channel).
-        """
-
-        super(UniformWeightsQuantizer, self).__init__()
-        self.num_bits = num_bits
-        self.min_range = min_range
-        self.max_range = max_range
-        self.quantization_method = quantization_method
-        self.per_channel = per_channel
-        self.output_channels_axis = output_channels_axis
-
-    def __call__(self, float_weight: np.ndarray, *args, **kwargs) -> np.ndarray:
-        """
-        Quantize tensor according to quantization params in the quantizer.
-
-        Args:
-            float_weight: Weights to quantize.
-
-        Returns:
-            Quantized tensor.
-
-        """
-        return uniform_quantize_tensor(tensor_data=float_weight,
-                                       range_min=self.min_range,
-                                       range_max=self.max_range,
-                                       n_bits=self.num_bits)

model_compression_toolkit/exporter/model_wrapper/pytorch/wrappers_quantize_configs/activation_quantize_config.py

@@ -1,52 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-from typing import List, Callable
-
-from model_compression_toolkit.core.pytorch.back2framework.quantization_wrapper.wrapper_quantize_config import \
-    WrapperQuantizeConfig
-
-
-class ActivationQuantizeConfig(WrapperQuantizeConfig):
-    """
-    QuantizationConfig for Fully Quantized model to quantize layer's activations only.
-
-    """
-
-    def __init__(self, activation_quantizers: List[Callable]):
-        """
-
-        Args:
-            activation_quantizers: List of quantizers for the layer's activation quantization.
-        """
-        super().__init__(is_weight_quantized=False,
-                         is_activation_quantized=True)
-
-        self._activation_quantizers = activation_quantizers
-
-    def get_weight_quantizers(self) -> list:
-        """
-
-        Returns: An empty list as this QC does not quantize the layer's weights.
-
-        """
-        return []
-
-    def get_activation_quantizers(self):
-        """
-
-        Returns: List of quantizers to quantize the layer's activations.
-
-        """
-        return self._activation_quantizers

model_compression_toolkit/exporter/model_wrapper/pytorch/wrappers_quantize_configs/no_quantization_quantize_config.py

@@ -1,46 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-
-from model_compression_toolkit.core.pytorch.back2framework.quantization_wrapper.wrapper_quantize_config import \
-    WrapperQuantizeConfig
-
-
-class NoQuantizationQuantizeConfig(WrapperQuantizeConfig):
-    """
-    QuantizationConfig for Fully Quantized model to keep layer's activations and weights un-quantized.
-    """
-
-    def __init__(self):
-        super().__init__(is_weight_quantized=False,
-                         is_activation_quantized=False)
-
-    def get_weight_quantizers(self):
-        """
-
-        Returns: An empty list as this QC does not quantize the layer's weights.
-
-        """
-        return []
-
-    def get_activation_quantizers(self):
-        """
-
-        Returns: An empty list as this QC does not quantize the layer's activations.
-
-        """
-        return []
-
-
-

model_compression_toolkit/exporter/model_wrapper/pytorch/wrappers_quantize_configs/weights_activation_quantize_config.py

@@ -1,54 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-from typing import Callable, List
-
-from model_compression_toolkit.core.pytorch.back2framework.quantization_wrapper.wrapper_quantize_config import \
-    WrapperQuantizeConfig
-
-
-class WeightsActivationQuantizeConfig(WrapperQuantizeConfig):
-    """
-    QuantizationConfig for Fully Quantized model to quantize layer's activations and weights.
-    """
-
-    def __init__(self,
-                 weight_quantizers: List[Callable],
-                 activation_quantizers: List[Callable]):
-        """
-        Args:
-            weight_quantizers: List of quantizers to quantize the layer's weights.
-            activation_quantizers: List of quantizers to quantize the layer's activations.
-        """
-        super().__init__(is_weight_quantized=True,
-                         is_activation_quantized=True)
-
-        self._weight_quantizers = weight_quantizers
-        self._activation_quantizers = activation_quantizers
-
-    def get_weight_quantizers(self) -> List[Callable]:
-        """
-
-        Returns: List of quantizers to quantize the layer's weights.
-
-        """
-        return self._weight_quantizers
-
-    def get_activation_quantizers(self) -> List[Callable]:
-        """
-
-        Returns: List of quantizers to quantize the layer's activations.
-
-        """
-        return self._activation_quantizers

model_compression_toolkit/exporter/model_wrapper/pytorch/wrappers_quantize_configs/weights_quantize_config.py

@@ -1,52 +0,0 @@
-# Copyright 2022 Sony Semiconductor Israel, Inc. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-# ==============================================================================
-from typing import List, Callable
-
-from model_compression_toolkit.core.pytorch.back2framework.quantization_wrapper.wrapper_quantize_config import \
-    WrapperQuantizeConfig
-
-
-class WeightsQuantizeConfig(WrapperQuantizeConfig):
-    """
-    QuantizationConfig for Fully Quantized model to quantize layer's weights only.
-    """
-
-    def __init__(self,
-                 weight_quantizers: List[Callable]):
-        """
-
-        Args:
-            weight_quantizers: List of quantizers to quantize the layer's weights.
-        """
-        super().__init__(is_weight_quantized=True,
-                         is_activation_quantized=False)
-
-        self._weight_quantizers = weight_quantizers
-
-    def get_weight_quantizers(self) -> List[Callable]:
-        """
-
-        Returns: List of quantizers to quantize the layer's weights.
-
-        """
-        return self._weight_quantizers
-
-    def get_activation_quantizer(self):
-        """
-
-        Returns: An empty list as this QC does not quantize the layer's activations.
-
-        """
-        return []