mct-nightly 2.0.0.20240409.404__py3-none-any.whl → 2.0.0.20240410.422__py3-none-any.whl

{mct_nightly-2.0.0.20240409.404.dist-info → mct_nightly-2.0.0.20240410.422.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240409.404
+Version: 2.0.0.20240410.422
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct_nightly-2.0.0.20240409.404.dist-info → mct_nightly-2.0.0.20240410.422.dist-info}/RECORD

@@ -1,4 +1,4 @@
-model_compression_toolkit/__init__.py,sha256=ALvOQYWLrTHNtxDnpNxy7lyftsvgDpzcoW-wTFtMedY,1573
+model_compression_toolkit/__init__.py,sha256=c33LV9Kt6hpVEoLixt_I5rqhtSzRBPSrdmFEifg-VHU,1573
 model_compression_toolkit/constants.py,sha256=KW_HUEPmQEYqCvWGyORqkYxpvO7w5LViB5J5D-pm_6o,3648
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -100,8 +100,8 @@ model_compression_toolkit/core/common/quantization/candidate_node_quantization_c
 model_compression_toolkit/core/common/quantization/core_config.py,sha256=KYdyfSmjSL4ye24nKlC_c4_AxYb14qoqaeMnZj4-8kE,2257
 model_compression_toolkit/core/common/quantization/debug_config.py,sha256=HtkMmneN-EmAzgZK4Vp4M8Sqm5QKdrvNyyZMpaVqYzY,1482
 model_compression_toolkit/core/common/quantization/filter_nodes_candidates.py,sha256=fwF4VILaX-u3ZaFd81xjbJuhg8Ef-JX_KfMXW0TPV-I,7136
-model_compression_toolkit/core/common/quantization/node_quantization_config.py,sha256=h_cgpvT50gYgO8T363-Zw_b2jfqo3uoa7TqnSuig7I4,26947
-model_compression_toolkit/core/common/quantization/quantization_config.py,sha256=hQMKm55EXS1oV-Upt6IQtsYhpuhMvYeWRJhh6lhv_Ko,6699
+model_compression_toolkit/core/common/quantization/node_quantization_config.py,sha256=TCgpvtfyzFUedv4sZ6sKzsTyikaVl2ixLj_aHPSC2r0,27014
+model_compression_toolkit/core/common/quantization/quantization_config.py,sha256=BieZDv9oc-Mc78S_LRMGo-s_2acbqiLE0ewaSE1v2VY,6818
 model_compression_toolkit/core/common/quantization/quantization_fn_selection.py,sha256=T1nVWdRJfBQ_iuMQYQSIkjfkR-2n3lAOKGAz_rUZZN0,2190
 model_compression_toolkit/core/common/quantization/quantization_params_fn_selection.py,sha256=MwIOBZ4BlZSTIOG75PDvlI3JmZ6t8YjPc1VP9Adei60,3847
 model_compression_toolkit/core/common/quantization/quantize_graph_weights.py,sha256=N005MSvx8UypVpa7XrxNrB2G732n2wHj3RmLyjTgd3I,2728
@@ -148,7 +148,7 @@ model_compression_toolkit/core/keras/__init__.py,sha256=mjbqLD-KcG3eNeCYpu1GBS7V
 model_compression_toolkit/core/keras/constants.py,sha256=Uv3c0UdW55pIVQNW_1HQlgl-dHXREkltOLyzp8G1mTQ,3163
 model_compression_toolkit/core/keras/custom_layer_validation.py,sha256=f-b14wuiIgitBe7d0MmofYhDCTO3IhwJgwrh-Hq_t_U,1192
 model_compression_toolkit/core/keras/default_framework_info.py,sha256=Ha4HTHuiw_KTS5Po1Xnv6GyK9eprpDhYWf-eooS62Ys,4961
-model_compression_toolkit/core/keras/keras_implementation.py,sha256=NDHLl19I-xQrQGcsAwTcFjnIjCRn31xaPrqDYm8g_dg,29027
+model_compression_toolkit/core/keras/keras_implementation.py,sha256=RS2UEtZ_anZeDxz7Zv6sNv7v9tFVct6d9KVrUlxTGpo,29309
 model_compression_toolkit/core/keras/keras_model_validation.py,sha256=1wNV2clFdC9BzIELRLSO2uKf0xqjLqlkTJudwtCeaJk,1722
 model_compression_toolkit/core/keras/keras_node_prior_info.py,sha256=Aqh31wOPaiZcJIOm-uJwzev0eTMdJyXaOk97rs4z7BU,3879
 model_compression_toolkit/core/keras/resource_utilization_data_facade.py,sha256=Xmk2ZL5CaYdb7iG62HdtZ1F64vap7ffnrsuR3e3G5hc,4851
@@ -166,6 +166,7 @@ model_compression_toolkit/core/keras/graph_substitutions/substitutions/activatio
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm_folding.py,sha256=9YCNPiK5BD7tLs1meabPhzfb2VsyPxrZM17zMFsW_Fo,8158
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm_reconstruction.py,sha256=GR1a3mCZpNUu4WxixJXF_aSm57phAdxaRoHecNx3hxw,3168
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm_refusing.py,sha256=5df_xGfXkqNub4xVRnCWQvSohWqdv12axjJ6edVU2H0,2478
+model_compression_toolkit/core/keras/graph_substitutions/substitutions/concat_threshold_update.py,sha256=Hl4LEQ_bw_Vpmf3ZqHujYUqVdvTNsPlEMvr9dZhwg2U,2806
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/dwconv_to_conv.py,sha256=R3U7cjc2E0zheMem16GHygp5jZFGSaomkNOTxTjcAgw,5794
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/input_scaling.py,sha256=V6hp67CkS_A3WqdsjLjs0ETtdZAOo4P9mhy4aT7W5FE,5940
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/linear_collapsing.py,sha256=dyhqZrxSTclXyarT2JYnI5WPX0OvWR_CQiwddIr632U,8143
@@ -209,7 +210,7 @@ model_compression_toolkit/core/pytorch/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKW
 model_compression_toolkit/core/pytorch/constants.py,sha256=NI-J7REuxn06oEIHsmJ4GqtNC3TbV8xlkJjt5Ar-c4U,2626
 model_compression_toolkit/core/pytorch/default_framework_info.py,sha256=r1XyzUFvrjGcJHQM5ETLsMZIG2yHCr9HMjqf0ti9inw,4175
 model_compression_toolkit/core/pytorch/pytorch_device_config.py,sha256=IoMvTch5awAEPvB6Tg6ANhFGXvfSgv7JLsUBlxpMwk4,4330
-model_compression_toolkit/core/pytorch/pytorch_implementation.py,sha256=dakO4Nj-tFfs53y6dJyXbpoljx2n3ZqmMoB4CFWGNSQ,26868
+model_compression_toolkit/core/pytorch/pytorch_implementation.py,sha256=mT4jd8E1saCpAgrsClufQbnVJ0eYn1xaTQ3teALu4jk,27117
 model_compression_toolkit/core/pytorch/pytorch_node_prior_info.py,sha256=n_B4a6FMwM9D2w8kzy3oenBWZgXNZuIZgTJC6JEuTy0,3250
 model_compression_toolkit/core/pytorch/resource_utilization_data_facade.py,sha256=E6ifk1HdO60k4IRH2EFBzAYWtwUlrGqJoQ66nknpHoQ,4983
 model_compression_toolkit/core/pytorch/utils.py,sha256=dRPiteBg2dBNsHwZyYzXiCIAjnelSoeZZsDXlsTw5JQ,2880
@@ -228,6 +229,7 @@ model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/__init_
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/batchnorm_folding.py,sha256=j3q5DzbH3ys5MPFfSOVnAXdD7-g4XEKj2ADrdihVr30,8292
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/batchnorm_reconstruction.py,sha256=B7aC2TZNrQJ2oQVGBFhKAVqdUU5lYVJSMmwKhjxOHWk,2822
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/batchnorm_refusing.py,sha256=JDWOaNwYrZG0zTwd3HwoZUM3tKu7zPbzLOrqNQsu8xA,2162
+model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/concat_threshold_update.py,sha256=SBrR24ZAnWPftLinv4FuIqdBGjfYtfXbYQJN5mgy5V4,2861
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/const_holder_conv.py,sha256=dYGyb5ebnoeFBF0EaHPQU7CkXvoARdznEEe0laM47LA,3919
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/functional_batch_norm.py,sha256=iX8bLHtw2osP42-peNLTRmbpX3cUxdGsAbEfw7NLpx0,3935
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/functional_layer_norm.py,sha256=zKSgtVw_P9fUvdq4e7P9yaLDPG_vZ0cecM9sVPtm1ns,3799
@@ -469,8 +471,8 @@ model_compression_toolkit/trainable_infrastructure/keras/quantize_wrapper.py,sha
 model_compression_toolkit/trainable_infrastructure/keras/quantizer_utils.py,sha256=MVwXNymmFRB2NXIBx4e2mdJ1RfoHxRPYRgjb1MQP5kY,1797
 model_compression_toolkit/trainable_infrastructure/pytorch/__init__.py,sha256=huHoBUcKNB6BnY6YaUCcFvdyBtBI172ZoUD8ZYeNc6o,696
 model_compression_toolkit/trainable_infrastructure/pytorch/base_pytorch_quantizer.py,sha256=7bbzqJN8ZAycVDvZr_5xC-niTAR5df8f03Kooev_pfg,3047
-mct_nightly-2.0.0.20240409.404.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.0.0.20240409.404.dist-info/METADATA,sha256=uDkh4Eu7g8uMdBVYp8H_rPGwkuhe_aWWK86DgPSBj94,18795
-mct_nightly-2.0.0.20240409.404.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-mct_nightly-2.0.0.20240409.404.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.0.0.20240409.404.dist-info/RECORD,,
+mct_nightly-2.0.0.20240410.422.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+mct_nightly-2.0.0.20240410.422.dist-info/METADATA,sha256=Xx2HTbZkpp4O8bS07IXSnaYSh9ZZTxe61I47ovv9fzE,18795
+mct_nightly-2.0.0.20240410.422.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+mct_nightly-2.0.0.20240410.422.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.0.0.20240410.422.dist-info/RECORD,,

model_compression_toolkit/__init__.py

@@ -27,4 +27,4 @@ from model_compression_toolkit import data_generation
 from model_compression_toolkit import pruning
 from model_compression_toolkit.trainable_infrastructure.keras.load_model import keras_load_quantized_model
 
-__version__ = "2.0.0.20240409.000404"
+__version__ = "2.0.0.20240410.000422"

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -106,6 +106,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
         self.z_threshold = qc.z_threshold
         self.shift_negative_ratio = qc.shift_negative_ratio
         self.shift_negative_threshold_recalculation = qc.shift_negative_threshold_recalculation
+        self.concat_threshold_update = qc.concat_threshold_update
 
     def quantize_node_output(self,
                              tensors: Any) -> Any:
@@ -219,7 +220,7 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
                self.shift_negative_activation_correction == other.shift_negative_activation_correction and \
                self.z_threshold == other.z_threshold and \
                self.shift_negative_ratio == other.shift_negative_ratio and \
-               self.shift_negative_threshold_recalculation == other.shift_negative_threshold_recalculation
+               self.shift_negative_threshold_recalculation == other.shift_negative_threshold_recalculation 
 
     def __hash__(self):
         return hash((self.activation_quantization_fn,

model_compression_toolkit/core/common/quantization/quantization_config.py

@@ -62,7 +62,8 @@ class QuantizationConfig:
                  residual_collapsing: bool = True,
                  shift_negative_ratio: float = 0.05,
                  shift_negative_threshold_recalculation: bool = False,
-                 shift_negative_params_search: bool = False):
+                 shift_negative_params_search: bool = False,
+                 concat_threshold_update: bool = False):
         """
         Class to wrap all different parameters the library quantize the input model according to.
 
@@ -117,6 +118,7 @@ class QuantizationConfig:
         self.shift_negative_ratio = shift_negative_ratio
         self.shift_negative_threshold_recalculation = shift_negative_threshold_recalculation
         self.shift_negative_params_search = shift_negative_params_search
+        self.concat_threshold_update = concat_threshold_update
 
     def __repr__(self):
         return str(self.__dict__)

model_compression_toolkit/core/keras/graph_substitutions/substitutions/concat_threshold_update.py

@@ -0,0 +1,66 @@
+# Copyright 2024 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 tensorflow.keras.layers import Concatenate
+import tensorflow as tf
+
+from model_compression_toolkit.core import common
+from model_compression_toolkit.core.common import Graph, BaseNode
+from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
+from model_compression_toolkit.constants import THRESHOLD
+
+
+
+class ConcatThresholdUpdate(common.BaseSubstitution):
+
+
+    """
+    Find concat layers and match their prior layers thresholds unless prior layer outputs to multiple layers.
+    """
+
+    def __init__(self):
+        """
+        Initialize a threshold_updater object.
+        """
+        concatination_node = NodeOperationMatcher(Concatenate) | \
+            NodeOperationMatcher(tf.concat) 
+        super().__init__(matcher_instance=concatination_node)
+
+    def substitute(self,
+                   graph: Graph,
+                   node: BaseNode) -> Graph:
+        """
+        Update previous layers thresholds to match concatinations quantization thresholds. No change if
+        previous layer outputs to multiple layers. No change in case of uniform quantization. 
+        No change in case of multiple quantization candidates (mixed precision).
+
+
+        Args:
+            graph: Graph we apply the substitution on.
+            node: Node refference to edit previous nodes thresholds.
+
+        Returns:
+            Graph after applying the substitution.
+        """
+
+        if len(node.candidates_quantization_cfg) == 1 and THRESHOLD in node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params:
+            concat_threshold = node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params[THRESHOLD]
+            prev_nodes = graph.get_prev_nodes(node)
+            for prev_node in prev_nodes:
+                if len(graph.get_next_nodes(prev_node))==1 and prev_node.type != Concatenate and prev_node.type != tf.concat:
+                    prev_node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params[THRESHOLD] = concat_threshold
+
+        return graph

model_compression_toolkit/core/keras/keras_implementation.py

@@ -80,7 +80,8 @@ from model_compression_toolkit.core.keras.graph_substitutions.substitutions.line
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.residual_collapsing import \
     keras_residual_collapsing
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.input_scaling import InputScaling, \
-    InputScalingWithPad
+    InputScalingWithPad 
+from model_compression_toolkit.core.keras.graph_substitutions.substitutions.concat_threshold_update import ConcatThresholdUpdate
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.relu_bound_to_power_of_2 import \
     ReLUBoundToPowerOfTwo
 from model_compression_toolkit.core.keras.graph_substitutions.substitutions.multi_head_attention_decomposition import \
@@ -300,8 +301,8 @@ class KerasImplementation(FrameworkImplementation):
         """
         return keras_op2d_add_const_collapsing()
 
-    def get_substitutions_post_statistics_collection(self, quant_config: QuantizationConfig) \
-            -> List[common.BaseSubstitution]:
+    def get_substitutions_post_statistics_collection(self, 
+                                                     quant_config: QuantizationConfig) -> List[common.BaseSubstitution]:
         """
         Return a list of the framework substitutions used after we collect statistics.
 
@@ -317,6 +318,8 @@ class KerasImplementation(FrameworkImplementation):
         if quant_config.input_scaling:
             substitutions_list.append(InputScaling())
             substitutions_list.append(InputScalingWithPad())
+        if quant_config.concat_threshold_update:
+            substitutions_list.append(ConcatThresholdUpdate())
         return substitutions_list
 
 

model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/concat_threshold_update.py

@@ -0,0 +1,69 @@
+# Copyright 2024 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 torch
+
+from model_compression_toolkit.core import common
+from model_compression_toolkit.core.common.graph.base_graph import Graph
+from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
+from model_compression_toolkit.core.common.graph.base_node import BaseNode
+from model_compression_toolkit.constants import THRESHOLD
+
+
+class ConcatThresholdUpdate(common.BaseSubstitution):
+    """
+    Find concat layers and match their prior layers thresholds unless prior layer outputs to multiple layers.
+    """
+
+
+    def __init__(self):
+        """
+        Initialize a threshold_updater object.
+        """
+        concatination_node = NodeOperationMatcher(torch.cat) | \
+            NodeOperationMatcher(torch.concat) 
+        super().__init__(matcher_instance=concatination_node)
+
+    def substitute(self,
+                   graph: Graph,
+                   node: BaseNode) -> Graph:
+        """
+        Update previous layers thresholds to match concatinations quantization thresholds. No change if
+        previous layer outputs to multiple layers. No change in case of uniform quantization. 
+        No change in case of multiple quantization candidates (mixed precision).
+
+
+        Args:
+            graph: Graph we apply the substitution on.
+            node: Node refference to edit previous nodes thresholds.
+
+        Returns:
+            Graph after applying the substitution.
+        """
+    
+        if len(node.candidates_quantization_cfg) == 1 and THRESHOLD in node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params:
+            concat_threshold = node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params[THRESHOLD]
+            prev_nodes = graph.get_prev_nodes(node)
+            for prev_node in prev_nodes:
+                if len(graph.get_next_nodes(prev_node))==1 and prev_node.type != torch.cat and prev_node.type != torch.concat:
+                    prev_node.candidates_quantization_cfg[0].activation_quantization_cfg.activation_quantization_params[THRESHOLD] = concat_threshold
+
+        return graph
+
+
+
+

model_compression_toolkit/core/pytorch/pytorch_implementation.py

@@ -73,6 +73,8 @@ from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.vi
     VirtualActivationWeightsComposition
 from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.weights_activation_split import \
     WeightsActivationSplit
+from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.concat_threshold_update import \
+    ConcatThresholdUpdate
 from model_compression_toolkit.core.pytorch.hessian.activation_trace_hessian_calculator_pytorch import \
     ActivationTraceHessianCalculatorPytorch
 from model_compression_toolkit.core.pytorch.hessian.weights_trace_hessian_calculator_pytorch import \
@@ -302,6 +304,8 @@ class PytorchImplementation(FrameworkImplementation):
             substitutions_list.append(pytorch_softmax_shift())
         if quant_config.input_scaling:
             Logger.critical('Input scaling is currently not supported for Pytorch.')
+        if quant_config.concat_threshold_update:
+            substitutions_list.append(ConcatThresholdUpdate())
         return substitutions_list