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

model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/keras/quantizers/activation_inferable_quantizers/activation_lut_pot_inferable_quantizer.py

@@ -0,0 +1,148 @@
+# Copyright 2023 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 warnings
+from typing import List
+
+import numpy as np
+
+from model_compression_toolkit.core.common.logger import Logger
+from model_compression_toolkit.core.common.constants import FOUND_TF
+
+from model_compression_toolkit.core.common.target_platform import QuantizationMethod
+from model_compression_toolkit.quantizers_infrastructure import QuantizationTarget
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.base_inferable_quantizer import mark_quantizer
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.constants import MULTIPLIER_N_BITS, EPS
+
+if FOUND_TF:
+    import tensorflow as tf
+    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.quantizer_utils import lut_quantizer
+
+
+    @mark_quantizer(quantization_target=QuantizationTarget.Activation,
+                    quantization_method=[QuantizationMethod.LUT_POT_QUANTIZER],
+                    quantizer_type=None)
+    class ActivationLutPOTInferableQuantizer(BaseKerasInferableQuantizer):
+        """
+        Class for quantizing activations using lut power-of-two quantizer
+        """
+
+        def __init__(self,
+                     num_bits: int,
+                     cluster_centers: np.ndarray,
+                     threshold: List[float],
+                     signed: bool,
+                     multiplier_n_bits: int = MULTIPLIER_N_BITS,
+                     eps: float = EPS):
+            """
+            Initialize the quantizer with the specified parameters.
+
+            Args:
+                num_bits: number of bits to use for quantization
+                cluster_centers: the cluster centers to assign the activations
+                threshold: threshold for quantizing activations
+                signed: whether or not to use signed quantization
+                multiplier_n_bits: Number of bits that determines the quantization range
+                eps: Small value for numerical stability in division
+            """
+            # Call the superclass constructor with the given parameters, along with the target of Activation
+            # quantization
+            super(ActivationLutPOTInferableQuantizer, self).__init__()
+
+            assert isinstance(threshold, list), f'Expected threshold to be of type list but is {type(threshold)}'
+            assert all([isinstance(x, (float, np.float32, tf.float32)) for x in
+                        threshold]), f'Expected threshold list to contain float or np.float values but found ' \
+                                     f'{[type(x) for x in threshold]}'
+
+            # In activation per-channel quantization is not supported thus we expect a single threshold value.
+            assert len(threshold) == 1, f'In per-tensor quantization threshold should be of ' \
+                                        f'length 1 but is {len(threshold)}'
+
+            is_threshold_pot = np.all([int(np.log2(x)) == np.log2(x) for x in threshold])
+            assert is_threshold_pot, f'Expected threshold to be power of 2 but is {threshold}'
+
+            self.threshold = threshold[0]
+
+            assert len(np.unique(cluster_centers)) <= 2 ** num_bits, \
+                f'Expected num of cluster centers to be less or equal than {2 ** num_bits} ' \
+                f'but got {len(cluster_centers)}'
+
+            assert not np.any(cluster_centers - cluster_centers.astype(int)), f'Expected cluster centers to be integers'
+
+            if signed:
+                assert np.all((-1 * (2 ** (multiplier_n_bits - int(signed))) <= cluster_centers) &
+                              (cluster_centers <= (2 ** (multiplier_n_bits - int(signed)) - 1))), \
+                    f'Expected cluster centers in the quantization range'
+            else:
+                assert np.all(cluster_centers <= (2 ** multiplier_n_bits)), \
+                    f'Expected cluster centers in the quantization range'
+
+            # num_bits must be less than multiplier_n_bits
+            assert num_bits <= multiplier_n_bits, f'Look-Up-Table bit configuration has {num_bits} bits. It must be ' \
+                                                  f'less then {multiplier_n_bits}'
+            if num_bits == multiplier_n_bits:
+                warnings.warn("Num of bits equal to multiplier n bits, Please be aware LUT quantizier may be "
+                              "inefficient in that case, consider using SymmetricInferableQuantizer instead")
+
+            # If unsigned activation quantization, all cluster_centers must have the same sign
+            if not signed:
+                assert np.all(cluster_centers >= 0), f'Expected unsigned cluster centers in unsigned activation ' \
+                                                     f'quantization '
+
+            self.num_bits = num_bits
+            self.cluster_centers = cluster_centers
+            self.signed = signed
+            self.multiplier_n_bits = multiplier_n_bits
+            self.eps = eps
+
+        def __call__(self, inputs: tf.Tensor) -> tf.Tensor:
+            """
+            Quantize the given inputs using the quantizer parameters.
+
+            Args:
+                inputs: input tensor to quantize
+
+            Returns:
+                quantized tensor.
+            """
+            assert inputs.dtype == tf.float32, f'Input tensor was expected to be a float tensor but is of type ' \
+                                               f'{inputs.dtype}'
+
+            return lut_quantizer(inputs, cluster_centers=self.cluster_centers, signed=self.signed,
+                                 threshold=self.threshold, multiplier_n_bits=self.multiplier_n_bits, eps=self.eps)
+
+        def get_config(self):
+            """
+            Return a dictionary with the configuration of the quantizer.
+
+            Returns:
+                Dictionary with the following keys: 'num_bits', 'cluster_centers', 'threshold', 'signed',
+                'multiplier_n_bits', 'eps'
+            """
+            return {'num_bits': self.num_bits,
+                    'cluster_centers': self.cluster_centers,
+                    'threshold': self.threshold,
+                    'signed': self.signed,
+                    'multiplier_n_bits': self.multiplier_n_bits,
+                    'eps': self.eps}
+
+
+else:
+    class ActivationLutPOTInferableQuantizer:  # pragma: no cover
+        def __init__(self, *args, **kwargs):
+            Logger.error('Installing tensorflow and tensorflow_model_optimization is mandatory '
+                         'when using ActivationLutPOTInferableQuantizer. '
+                         'Could not find Tensorflow package.')

model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/keras/quantizers/activation_inferable_quantizers/activation_pot_inferable_quantizer.py

@@ -0,0 +1,65 @@
+# Copyright 2023 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
+
+import numpy as np
+
+from model_compression_toolkit.core.common.logger import Logger
+from model_compression_toolkit.core.common.constants import FOUND_TF
+
+from model_compression_toolkit.core.common.target_platform import QuantizationMethod
+from model_compression_toolkit.quantizers_infrastructure import QuantizationTarget
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.base_inferable_quantizer import mark_quantizer
+
+
+if FOUND_TF:
+    from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.quantizers.activation_inferable_quantizers.activation_symmetric_inferable_quantizer \
+        import ActivationSymmetricInferableQuantizer
+
+    @mark_quantizer(quantization_target=QuantizationTarget.Activation,
+                    quantization_method=[QuantizationMethod.POWER_OF_TWO],
+                    quantizer_type=None)
+    class ActivationPOTInferableQuantizer(ActivationSymmetricInferableQuantizer):
+        """
+        Class for quantizing activations using power-of-two quantizer
+        """
+
+        def __init__(self,
+                     num_bits: int,
+                     threshold: List[float],
+                     signed: bool):
+            """
+            Initialize the quantizer with the specified parameters.
+
+            Args:
+                num_bits: number of bits to use for quantization
+                threshold: threshold for quantizing activations
+                signed: whether or not to use signed quantization
+            """
+            # Call the superclass constructor with the given parameters, along with the target of Activation
+            # quantization
+            super(ActivationPOTInferableQuantizer, self).__init__(num_bits=num_bits,
+                                                                  threshold=threshold,
+                                                                  signed=signed)
+
+            is_threshold_pot = np.all([int(np.log2(x)) == np.log2(x) for x in self.threshold.flatten()])
+            assert is_threshold_pot, f'Expected threshold to be power of 2 but is {self.threshold}'
+
+else:
+    class ActivationPOTInferableQuantizer:  # pragma: no cover
+        def __init__(self, *args, **kwargs):
+            Logger.error('Installing tensorflow and tensorflow_model_optimization is mandatory '
+                         'when using ActivationPOTInferableQuantizer. '
+                         'Could not find Tensorflow package.')

model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/keras/quantizers/activation_inferable_quantizers/activation_symmetric_inferable_quantizer.py

@@ -0,0 +1,86 @@
+# Copyright 2023 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
+
+import numpy as np
+
+from model_compression_toolkit.core.common.constants import FOUND_TF
+
+from model_compression_toolkit.core.common.target_platform import QuantizationMethod
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.base_inferable_quantizer import mark_quantizer, \
+    QuantizationTarget
+
+
+if FOUND_TF:
+    import tensorflow as tf
+    from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.quantizers.activation_inferable_quantizers.activation_uniform_inferable_quantizer \
+        import ActivationUniformInferableQuantizer
+
+    @mark_quantizer(quantization_target=QuantizationTarget.Activation,
+                    quantization_method=[QuantizationMethod.SYMMETRIC],
+                    quantizer_type=None)
+    class ActivationSymmetricInferableQuantizer(ActivationUniformInferableQuantizer):
+
+        """
+        Class for quantizing activations using a symmetric quantizer
+        """
+
+        def __init__(self,
+                     num_bits: int,
+                     threshold: List[float],
+                     signed: bool):
+            """
+            Initialize the quantizer with the specified parameters.
+
+            Args:
+                num_bits: number of bits to use for quantization
+                threshold: threshold for quantizing activations
+                signed: whether or not to use signed quantization
+            """
+            assert isinstance(threshold, list), f'Expected threshold to be of type list but is {type(threshold)}'
+            assert all([isinstance(x, (float, np.float32, tf.float32)) for x in
+                        threshold]), f'Expected threshold list to contain float or np.float values but found ' \
+                                     f'{[type(x) for x in threshold]}'
+
+            self.threshold = np.asarray(threshold)
+            self.signed = signed
+
+            delta = self.threshold / (2 ** (num_bits - int(self.signed)))
+            # In activation quantization is per-tensor only - thus we pass the threshold as a list with a len of 1
+            min_range = list(-self.threshold) if self.signed else [0.0]
+            max_range = list(self.threshold - delta)
+
+            super(ActivationSymmetricInferableQuantizer, self).__init__(num_bits=num_bits,
+                                                                        min_range=min_range,
+                                                                        max_range=max_range)
+
+        def get_config(self):
+            """
+            Return a dictionary with the configuration of the quantizer.
+
+            Returns:
+                Dictionary with the following keys: 'num_bits', 'signed', 'threshold'
+            """
+            return {'num_bits': self.num_bits,
+                    'signed': self.signed,
+                    'threshold': self.threshold}
+
+
+else:
+    class ActivationSymmetricInferableQuantizer:  # pragma: no cover
+        def __init__(self, *args, **kwargs):
+            raise Exception('Installing tensorflow and tensorflow_model_optimization is mandatory '
+                            'when using ActivationSymmetricInferableQuantizer. '
+                            'Could not find Tensorflow package.')

model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/keras/quantizers/activation_inferable_quantizers/activation_uniform_inferable_quantizer.py

@@ -0,0 +1,111 @@
+# Copyright 2023 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
+
+import numpy as np
+
+from model_compression_toolkit.core.common.logger import Logger
+from model_compression_toolkit.core.common.constants import FOUND_TF
+from model_compression_toolkit.core.common.target_platform import QuantizationMethod
+from model_compression_toolkit.quantizers_infrastructure import QuantizationTarget
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.base_inferable_quantizer import mark_quantizer
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.quant_utils import \
+    adjust_range_to_include_zero
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.validation_functions import \
+    validate_uniform_min_max_ranges, validate_adjusted_min_max_ranges
+
+if FOUND_TF:
+    import tensorflow as tf
+    from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.quantizers.base_keras_inferable_quantizer import BaseKerasInferableQuantizer
+
+    @mark_quantizer(quantization_target=QuantizationTarget.Activation,
+                    quantization_method=[QuantizationMethod.UNIFORM],
+                    quantizer_type=None)
+    class ActivationUniformInferableQuantizer(BaseKerasInferableQuantizer):
+        """
+        Class for quantizing activations using an uniform quantizer
+        """
+
+        def __init__(self,
+                     num_bits: int,
+                     min_range: List[float],
+                     max_range: List[float],
+                     ):
+            """
+            Initialize the quantizer with the specified parameters.
+
+            Args:
+                num_bits: number of bits to use for quantization
+                min_range: min range for quantizing activations
+                max_range: max range for quantizing activations
+            """
+            super(ActivationUniformInferableQuantizer, self).__init__()
+
+            # Validate ranges properties
+            validate_uniform_min_max_ranges(min_range,
+                                            max_range)
+
+            # In activation per-channel quantization is not supported thus we expect a single min/max value.
+            assert len(min_range) == 1, f'In per-tensor quantization min_range should be of length 1 but is {len(min_range)}'
+            assert len(max_range) == 1, f'In per-tensor quantization max_range should be of length 1 but is {len(max_range)}'
+
+            self.num_bits = num_bits
+
+            # Convert min/max to numpy arrays
+            min_range, max_range = np.asarray(min_range), np.asarray(max_range)
+            _min_range, _max_range = adjust_range_to_include_zero(min_range, max_range, num_bits)
+            validate_adjusted_min_max_ranges(min_range=min_range,
+                                             max_range=max_range,
+                                             adj_min=_min_range,
+                                             adj_max=_max_range)
+
+            self.max_range = _max_range[0]
+            self.min_range = _min_range[0]
+
+
+        def __call__(self, inputs:tf.Tensor) -> tf.Tensor:
+            """
+            Quantize the given inputs using the quantizer parameters.
+
+            Args:
+                inputs: input tensor to quantize
+
+            Returns:
+                quantized tensor.
+            """
+            assert inputs.dtype==tf.float32, f'Input tensor was expected to be a float tensor but is of type {inputs.dtype}'
+
+            return tf.quantization.fake_quant_with_min_max_vars(inputs,
+                                                                min=self.min_range,
+                                                                max=self.max_range,
+                                                                num_bits=self.num_bits)
+
+        def get_config(self):
+            """
+            Return a dictionary with the configuration of the quantizer.
+
+            Returns:
+                Dictionary with the following keys: 'num_bits', 'min_range', 'max_range'
+            """
+            return {'num_bits': self.num_bits,
+                    'min_range': self.min_range,
+                    'max_range': self.max_range}
+
+else:
+    class ActivationUniformInferableQuantizer:  # pragma: no cover
+        def __init__(self, *args, **kwargs):
+            Logger.error('Installing tensorflow and tensorflow_model_optimization is mandatory '
+                         'when using ActivationUniformInferableQuantizer. '
+                         'Could not find Tensorflow package.')

model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/keras/quantizers/base_keras_inferable_quantizer.py

@@ -0,0 +1,56 @@
+# Copyright 2023 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 abc import abstractmethod
+
+from model_compression_toolkit.core.common.constants import FOUND_TF
+from model_compression_toolkit.quantizers_infrastructure import BaseInferableQuantizer
+
+if FOUND_TF:
+    import tensorflow as tf
+
+    class BaseKerasInferableQuantizer(BaseInferableQuantizer):
+        def __init__(self):
+            """
+            This class is a base quantizer for Keras quantizers for inference only.
+            """
+            super(BaseKerasInferableQuantizer, self).__init__()
+
+        @abstractmethod
+        def get_config(self):
+            """
+            Return a dictionary with the configuration of the quantizer.
+            """
+            raise NotImplemented(f'{self.__class__.__name__} did not implement get_config')  # pragma: no cover
+
+        @abstractmethod
+        def __call__(self, inputs: tf.Tensor):
+            """
+            Quantize the given inputs using the quantizer parameters.
+
+            Args:
+                inputs: input tensor to quantize
+
+            Returns:
+                quantized tensor.
+            """
+            raise NotImplemented(f'{self.__class__.__name__} did not implement __call__')  # pragma: no cover
+else:
+    class BaseKerasInferableQuantizer:  # pragma: no cover
+        def __init__(self, *args, **kwargs):
+            raise Exception('Installing tensorflow and tensorflow_model_optimization is mandatory '
+                            'when using BaseKerasInferableQuantizer. '
+                            'Could not find Tensorflow package.')
+
+

model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/keras/quantizers/constants.py

@@ -0,0 +1,25 @@
+# Copyright 2023 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.
+# ==============================================================================
+
+# Inferable keras quantizer signature parameters:
+NUM_BITS = 'num_bits'
+SIGNED = 'signed'
+THRESHOLD = 'threshold'
+PER_CHANNEL = 'per_channel'
+MIN_RANGE = 'min_range'
+MAX_RANGE = 'max_range'
+CHANNEL_AXIS = 'channel_axis'
+INPUT_RANK = 'input_rank'
+CLUSTER_CENTERS = 'cluster_centers'

model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/keras/quantizers/weights_inferable_quantizers/__init__.py

@@ -0,0 +1,14 @@
+# Copyright 2023 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.
+# ==============================================================================

model_compression_toolkit/quantizers_infrastructure/inferable_infrastructure/keras/quantizers/weights_inferable_quantizers/weights_lut_pot_inferable_quantizer.py

@@ -0,0 +1,79 @@
+# Copyright 2023 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
+
+import numpy as np
+
+from model_compression_toolkit.core.common.constants import FOUND_TF
+from model_compression_toolkit.core.common.target_platform import QuantizationMethod
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.base_inferable_quantizer import mark_quantizer, \
+    QuantizationTarget
+from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.common.constants import MULTIPLIER_N_BITS, EPS
+
+if FOUND_TF:
+    from model_compression_toolkit.quantizers_infrastructure.inferable_infrastructure.keras.quantizers.weights_inferable_quantizers.\
+        weights_lut_symmetric_inferable_quantizer import WeightsLUTSymmetricInferableQuantizer
+
+
+    @mark_quantizer(quantization_target=QuantizationTarget.Weights,
+                    quantization_method=[QuantizationMethod.LUT_POT_QUANTIZER],
+                    quantizer_type=None)
+    class WeightsLUTPOTInferableQuantizer(WeightsLUTSymmetricInferableQuantizer):
+        """
+        Class for quantizing weights using a lut power-of-two quantizer
+        """
+
+        def __init__(self,
+                     num_bits: int,
+                     cluster_centers: np.ndarray,
+                     threshold: List[float],
+                     per_channel: bool,
+                     channel_axis: int = None,
+                     input_rank: int = None,
+                     multiplier_n_bits: int = MULTIPLIER_N_BITS,
+                     eps: float = EPS):
+            """
+            Initialize the quantizer with the specified parameters.
+
+            Args:
+                num_bits: number of bits to use for quantization
+                cluster_centers: the cluster centers to assign the weights
+                threshold: threshold for quantizing weights
+                per_channel: whether to use per-channel quantization
+                channel_axis: axis along which to apply per-channel quantization
+                input_rank: number of dimensions of input tensor the quantizer quantizes
+                multiplier_n_bits: Number of bits that determines the quantization range
+                eps: Small value for numerical stability in division
+            """
+
+            super(WeightsLUTPOTInferableQuantizer, self).__init__(num_bits=num_bits,
+                                                                  cluster_centers=cluster_centers,
+                                                                  threshold=threshold,
+                                                                  per_channel=per_channel,
+                                                                  channel_axis=channel_axis,
+                                                                  input_rank=input_rank,
+                                                                  multiplier_n_bits=multiplier_n_bits,
+                                                                  eps=eps)
+
+            is_threshold_pot = np.all([int(np.log2(x)) == np.log2(x) for x in self.threshold.flatten()])
+            assert is_threshold_pot, f'Expected threshold to be power of 2 but is {self.threshold}'
+
+
+else:
+    class WeightsLUTPOTInferableQuantizer:  # pragma: no cover
+        def __init__(self, *args, **kwargs):
+            raise Exception('Installing tensorflow and tensorflow_model_optimization is mandatory '
+                            'when using WeightsLUTPOTInferableQuantizer. '
+                            'Could not find Tensorflow package.')