mct-nightly 1.8.0.8032023.post421__py3-none-any.whl → 1.8.0.8052023.post414__py3-none-any.whl

model_compression_toolkit/{core → target_platform_capabilities}/tpc_models/qnnpack_tpc/target_platform_capabilities.py

@@ -13,7 +13,8 @@
 # limitations under the License.
 # ==============================================================================
 
-from model_compression_toolkit.core.common.constants import FOUND_TF, FOUND_TORCH, TENSORFLOW, PYTORCH, LATEST
+from model_compression_toolkit.constants import FOUND_TF, FOUND_TORCH, TENSORFLOW, PYTORCH
+from model_compression_toolkit.target_platform_capabilities.constants import LATEST
 
 
 ###############################
@@ -21,8 +22,8 @@ from model_compression_toolkit.core.common.constants import FOUND_TF, FOUND_TORC
 ###############################
 keras_tpc_models_dict = None
 if FOUND_TF:
-    from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tpc_keras import get_keras_tpc as get_keras_tpc_v1
-    from model_compression_toolkit.core.tpc_models.qnnpack_tpc.latest import get_keras_tpc_latest
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.qnnpack_tpc.v1.tpc_keras import get_keras_tpc as get_keras_tpc_v1
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.qnnpack_tpc.latest import get_keras_tpc_latest
 
     # Keras: TPC versioning
     keras_tpc_models_dict = {'v1': get_keras_tpc_v1(),
@@ -33,9 +34,9 @@ if FOUND_TF:
 ###############################
 pytorch_tpc_models_dict = None
 if FOUND_TORCH:
-    from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tpc_pytorch import \
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.qnnpack_tpc.v1.tpc_pytorch import \
         get_pytorch_tpc as get_pytorch_tpc_v1
-    from model_compression_toolkit.core.tpc_models.qnnpack_tpc.latest import get_pytorch_tpc_latest
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.qnnpack_tpc.latest import get_pytorch_tpc_latest
 
     # Pytorch: TPC versioning
     pytorch_tpc_models_dict = {'v1': get_pytorch_tpc_v1(),

model_compression_toolkit/{core → target_platform_capabilities}/tpc_models/qnnpack_tpc/v1/tp_model.py

@@ -15,7 +15,10 @@
 from typing import List, Tuple
 
 import model_compression_toolkit as mct
-from model_compression_toolkit.core.common.target_platform import OpQuantizationConfig, TargetPlatformModel
+from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
+    TargetPlatformModel
+from model_compression_toolkit.target_platform_capabilities.target_platform.quantization_format import \
+    QuantizationFormat
 
 tp = mct.target_platform
 
@@ -120,4 +123,7 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         tp.Fusing([conv, relu])
         tp.Fusing([linear, relu])
 
+        # Set quantization format to fakely quant
+        generated_tpc.set_quantization_format(QuantizationFormat.FAKELY_QUANT)
+
     return generated_tpc

model_compression_toolkit/{core → target_platform_capabilities}/tpc_models/qnnpack_tpc/v1/tpc_keras.py

@@ -15,7 +15,7 @@
 import tensorflow as tf
 
 from packaging import version
-from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1 import __version__ as TPC_VERSION
+from model_compression_toolkit.target_platform_capabilities.tpc_models.qnnpack_tpc.v1 import __version__ as TPC_VERSION
 
 if version.parse(tf.__version__) < version.parse("2.6"):
     from tensorflow.keras.layers import Conv2D, DepthwiseConv2D, Conv2DTranspose, Dense, BatchNormalization, ReLU, \
@@ -23,7 +23,7 @@ if version.parse(tf.__version__) < version.parse("2.6"):
 else:
     from keras.layers import Conv2D, DepthwiseConv2D, Conv2DTranspose, Dense, BatchNormalization, ReLU, Activation
 
-from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tp_model import get_tp_model
+from model_compression_toolkit.target_platform_capabilities.tpc_models.qnnpack_tpc.v1.tp_model import get_tp_model
 import model_compression_toolkit as mct
 
 tp = mct.target_platform

model_compression_toolkit/{core → target_platform_capabilities}/tpc_models/qnnpack_tpc/v1/tpc_pytorch.py

@@ -16,9 +16,9 @@ import torch
 from torch.nn import Conv2d, Linear, BatchNorm2d, ConvTranspose2d, Hardtanh, ReLU, ReLU6
 from torch.nn.functional import relu, relu6, hardtanh
 
-from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tp_model import get_tp_model
+from model_compression_toolkit.target_platform_capabilities.tpc_models.qnnpack_tpc.v1.tp_model import get_tp_model
 import model_compression_toolkit as mct
-from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1 import __version__ as TPC_VERSION
+from model_compression_toolkit.target_platform_capabilities.tpc_models.qnnpack_tpc.v1 import __version__ as TPC_VERSION
 
 tp = mct.target_platform
 

model_compression_toolkit/target_platform_capabilities/tpc_models/tflite_tpc/latest/__init__.py

@@ -0,0 +1,22 @@
+# 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.constants import FOUND_TF, FOUND_TORCH
+from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tp_model import get_tp_model, generate_tp_model, get_op_quantization_configs
+if FOUND_TF:
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tpc_keras import get_keras_tpc as get_keras_tpc_latest
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tpc_keras import generate_keras_tpc
+if FOUND_TORCH:
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tpc_pytorch import get_pytorch_tpc as get_pytorch_tpc_latest
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tpc_pytorch import generate_pytorch_tpc

model_compression_toolkit/{core → target_platform_capabilities}/tpc_models/tflite_tpc/target_platform_capabilities.py

@@ -13,7 +13,8 @@
 # limitations under the License.
 # ==============================================================================
 
-from model_compression_toolkit.core.common.constants import FOUND_TF, FOUND_TORCH, TENSORFLOW, PYTORCH, LATEST
+from model_compression_toolkit.constants import FOUND_TF, FOUND_TORCH, TENSORFLOW, PYTORCH
+from model_compression_toolkit.target_platform_capabilities.constants import LATEST
 
 
 ###############################
@@ -21,8 +22,8 @@ from model_compression_toolkit.core.common.constants import FOUND_TF, FOUND_TORC
 ###############################
 keras_tpc_models_dict = None
 if FOUND_TF:
-    from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tpc_keras import get_keras_tpc as get_keras_tpc_v1
-    from model_compression_toolkit.core.tpc_models.tflite_tpc.latest import get_keras_tpc_latest
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tpc_keras import get_keras_tpc as get_keras_tpc_v1
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.latest import get_keras_tpc_latest
 
     # Keras: TPC versioning
     keras_tpc_models_dict = {'v1': get_keras_tpc_v1(),
@@ -33,9 +34,9 @@ if FOUND_TF:
 ###############################
 pytorch_tpc_models_dict = None
 if FOUND_TORCH:
-    from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tpc_pytorch import \
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tpc_pytorch import \
         get_pytorch_tpc as get_pytorch_tpc_v1
-    from model_compression_toolkit.core.tpc_models.tflite_tpc.latest import get_pytorch_tpc_latest
+    from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.latest import get_pytorch_tpc_latest
 
     # Pytorch: TPC versioning
     pytorch_tpc_models_dict = {'v1': get_pytorch_tpc_v1(),

model_compression_toolkit/{core → target_platform_capabilities}/tpc_models/tflite_tpc/v1/tp_model.py

@@ -15,8 +15,12 @@
 from typing import List, Tuple
 
 import model_compression_toolkit as mct
-from model_compression_toolkit.core.common.target_platform import OpQuantizationConfig, TargetPlatformModel
-from model_compression_toolkit.core.common.target_platform.op_quantization_config import QuantizationMethod
+from model_compression_toolkit.target_platform_capabilities.target_platform import OpQuantizationConfig, \
+    TargetPlatformModel
+from model_compression_toolkit.target_platform_capabilities.target_platform.op_quantization_config import \
+    QuantizationMethod
+from model_compression_toolkit.target_platform_capabilities.target_platform.quantization_format import \
+    QuantizationFormat
 
 tp = mct.target_platform
 
@@ -65,7 +69,7 @@ def get_op_quantization_configs() -> Tuple[OpQuantizationConfig, List[OpQuantiza
         weights_multiplier_nbits=None
     )
 
-    mixed_precision_cfg_list = [] # No mixed precision
+    mixed_precision_cfg_list = []  # No mixed precision
 
     return eight_bits, mixed_precision_cfg_list
 
@@ -106,28 +110,28 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         # differently. For more details:
         # https://www.tensorflow.org/lite/performance/quantization_spec#int8_quantized_operator_specifications
         tp.OperatorsSet("NoQuantization",
-                         tp.get_default_quantization_config_options().clone_and_edit(
-                                  quantization_preserving=True))
+                        tp.get_default_quantization_config_options().clone_and_edit(
+                            quantization_preserving=True))
 
         fc = tp.OperatorsSet("FullyConnected",
-                              tp.get_default_quantization_config_options().clone_and_edit(
-                                       weights_per_channel_threshold=False))
+                             tp.get_default_quantization_config_options().clone_and_edit(
+                                 weights_per_channel_threshold=False))
 
         tp.OperatorsSet("L2Normalization",
-                         tp.get_default_quantization_config_options().clone_and_edit(
-                                  fixed_zero_point=0, fixed_scale=1 / 128))
+                        tp.get_default_quantization_config_options().clone_and_edit(
+                            fixed_zero_point=0, fixed_scale=1 / 128))
         tp.OperatorsSet("LogSoftmax",
-                         tp.get_default_quantization_config_options().clone_and_edit(
-                                  fixed_zero_point=127, fixed_scale=16 / 256))
+                        tp.get_default_quantization_config_options().clone_and_edit(
+                            fixed_zero_point=127, fixed_scale=16 / 256))
         tp.OperatorsSet("Tanh",
-                         tp.get_default_quantization_config_options().clone_and_edit(
-                                  fixed_zero_point=0, fixed_scale=1 / 128))
+                        tp.get_default_quantization_config_options().clone_and_edit(
+                            fixed_zero_point=0, fixed_scale=1 / 128))
         tp.OperatorsSet("Softmax",
-                         tp.get_default_quantization_config_options().clone_and_edit(
-                                  fixed_zero_point=-128, fixed_scale=1 / 256))
+                        tp.get_default_quantization_config_options().clone_and_edit(
+                            fixed_zero_point=-128, fixed_scale=1 / 256))
         tp.OperatorsSet("Logistic",
-                         tp.get_default_quantization_config_options().clone_and_edit(
-                                  fixed_zero_point=-128, fixed_scale=1 / 256))
+                        tp.get_default_quantization_config_options().clone_and_edit(
+                            fixed_zero_point=-128, fixed_scale=1 / 256))
 
         conv2d = tp.OperatorsSet("Conv2d")
         kernel = tp.OperatorSetConcat(conv2d, fc)
@@ -140,7 +144,8 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         bias_add = tp.OperatorsSet("BiasAdd")
         add = tp.OperatorsSet("Add")
         squeeze = tp.OperatorsSet("Squeeze",
-                                   qc_options=tp.get_default_quantization_config_options().clone_and_edit(quantization_preserving=True))
+                                  qc_options=tp.get_default_quantization_config_options().clone_and_edit(
+                                      quantization_preserving=True))
         # ------------------- #
         # Fusions
         # ------------------- #
@@ -152,4 +157,7 @@ def generate_tp_model(default_config: OpQuantizationConfig,
         tp.Fusing([batch_norm, activations_to_fuse])
         tp.Fusing([batch_norm, add, activations_to_fuse])
 
+        # Set quantization format to int8
+        generated_tpc.set_quantization_format(QuantizationFormat.INT8)
+
     return generated_tpc

model_compression_toolkit/{core → target_platform_capabilities}/tpc_models/tflite_tpc/v1/tpc_keras.py

@@ -24,11 +24,11 @@ else:
 
 from tensorflow.python.keras.layers.core import SlicingOpLambda
 from tensorflow.python.ops.image_ops_impl import ResizeMethod
-from model_compression_toolkit.core.common.target_platform.targetplatform2framework.attribute_filter import Eq
+from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework.attribute_filter import Eq
 
-from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tp_model import get_tp_model
+from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tp_model import get_tp_model
 import model_compression_toolkit as mct
-from model_compression_toolkit.core.tpc_models.tflite_tpc.v1 import __version__ as TPC_VERSION
+from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1 import __version__ as TPC_VERSION
 
 tp = mct.target_platform
 

model_compression_toolkit/{core → target_platform_capabilities}/tpc_models/tflite_tpc/v1/tpc_pytorch.py

@@ -15,11 +15,11 @@
 import torch
 from torch.nn import AvgPool2d, MaxPool2d
 from torch.nn.functional import avg_pool2d, max_pool2d, interpolate
-from model_compression_toolkit.core.common.target_platform.targetplatform2framework.attribute_filter import Eq
+from model_compression_toolkit.target_platform_capabilities.target_platform.targetplatform2framework.attribute_filter import Eq
 
-from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tp_model import get_tp_model
+from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1.tp_model import get_tp_model
 import model_compression_toolkit as mct
-from model_compression_toolkit.core.tpc_models.tflite_tpc.v1 import __version__ as TPC_VERSION
+from model_compression_toolkit.target_platform_capabilities.tpc_models.tflite_tpc.v1 import __version__ as TPC_VERSION
 
 tp = mct.target_platform
 

model_compression_toolkit/core/tpc_models/default_tpc/latest/__init__.py

@@ -1,25 +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.constants import FOUND_TF, FOUND_TORCH
-
-from model_compression_toolkit.core.tpc_models.default_tpc.v4.tp_model import get_tp_model, generate_tp_model, get_op_quantization_configs
-
-if FOUND_TF:
-    from model_compression_toolkit.core.tpc_models.default_tpc.v4.tpc_keras import get_keras_tpc as get_keras_tpc_latest
-    from model_compression_toolkit.core.tpc_models.default_tpc.v4.tpc_keras import generate_keras_tpc
-
-if FOUND_TORCH:
-    from model_compression_toolkit.core.tpc_models.default_tpc.v4.tpc_pytorch import get_pytorch_tpc as get_pytorch_tpc_latest
-    from model_compression_toolkit.core.tpc_models.default_tpc.v4.tpc_pytorch import generate_pytorch_tpc

model_compression_toolkit/core/tpc_models/qnnpack_tpc/latest/__init__.py

@@ -1,22 +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.constants import FOUND_TF, FOUND_TORCH
-from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tp_model import get_tp_model, generate_tp_model, get_op_quantization_configs
-if FOUND_TF:
-    from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tpc_keras import get_keras_tpc as get_keras_tpc_latest
-    from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tpc_keras import generate_keras_tpc
-if FOUND_TORCH:
-    from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tpc_pytorch import get_pytorch_tpc as get_pytorch_tpc_latest
-    from model_compression_toolkit.core.tpc_models.qnnpack_tpc.v1.tpc_pytorch import generate_pytorch_tpc

model_compression_toolkit/core/tpc_models/tflite_tpc/latest/__init__.py

@@ -1,22 +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.constants import FOUND_TF, FOUND_TORCH
-from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tp_model import get_tp_model, generate_tp_model, get_op_quantization_configs
-if FOUND_TF:
-    from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tpc_keras import get_keras_tpc as get_keras_tpc_latest
-    from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tpc_keras import generate_keras_tpc
-if FOUND_TORCH:
-    from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tpc_pytorch import get_pytorch_tpc as get_pytorch_tpc_latest
-    from model_compression_toolkit.core.tpc_models.tflite_tpc.v1.tpc_pytorch import generate_pytorch_tpc

model_compression_toolkit/gptq/common/gptq_quantizer_config.py

@@ -1,93 +0,0 @@
-# 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 Any, List, Callable
-
-from model_compression_toolkit.core.common import Logger
-from model_compression_toolkit.gptq.common.gptq_constants import REG_DEFAULT, REGULARIZATION_VALUES
-
-
-class GPTQQuantizerConfig:
-    """
-    A base class to define specific quantizer configuration for GPTQ quantizer.
-    """
-
-    def __init__(self):
-        self.n_batches = None
-
-    def get_regularization_value(self, fxp_model: Any, **kwargs) -> float:
-        """
-        Computes a regularization value for the quantizer's loss (if needed).
-        In the base class it only returns 0, to be used for GPTQ quantizers that don't require regularization.
-
-        Args:
-            fxp_model: The quantized model that is being trained.
-            **kwargs: Additional arguments for the quantizer regularization computation.
-
-        Returns: The regularization value.
-        """
-
-        return 0
-
-    def set_num_batches(self, num_batches: int):
-        """
-        Allows to set the number of batches that the quantizer uses for training (in each epoch).
-
-        Args:
-            num_batches: number of batches to be set.
-
-        """
-        self.n_batches = num_batches
-
-
-class SoftQuantizerConfig(GPTQQuantizerConfig):
-    def __init__(self, entropy_regularization: float = REG_DEFAULT):
-        """
-        Initializes an object that holds the arguments that are needed for soft rounding quantizer.
-
-        Args:
-            entropy_regularization (float): A floating point number that defines the gumbel entropy regularization factor.
-        """
-
-        super().__init__()
-        self.entropy_regularization = entropy_regularization
-
-
-    def get_regularization_value(self, fxp_model: Any, **kwargs) -> float:
-        """
-        Computes a regularization value for the soft quantizer.
-
-        Args:
-            fxp_model: The quantized model that is being trained.
-            **kwargs: Additional arguments for the quantizer regularization computation.
-
-        Returns: The regularization value.
-        """
-
-        soft_rounding_reg_values = kwargs.get(REGULARIZATION_VALUES)
-
-        if soft_rounding_reg_values is None:
-            Logger.error("No regularization values has been provided for computing the regularization "  # pragma: no cover
-                         "of the soft quantizer.")
-        if not isinstance(soft_rounding_reg_values, List):
-            Logger.error("The provided regularization values parameter of the soft quantizer "  # pragma: no cover
-                         "is not compatible (should be a list).")
-
-        reg = 0
-
-        for sq in soft_rounding_reg_values:
-            reg += sq
-
-        return self.entropy_regularization * reg
-