mct-nightly 2.2.0.20241106.458__py3-none-any.whl → 2.2.0.20241108.459__py3-none-any.whl

{mct_nightly-2.2.0.20241106.458.dist-info → mct_nightly-2.2.0.20241108.459.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.2.0.20241106.458
+Version: 2.2.0.20241108.459
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct_nightly-2.2.0.20241106.458.dist-info → mct_nightly-2.2.0.20241108.459.dist-info}/RECORD

@@ -1,4 +1,4 @@
-model_compression_toolkit/__init__.py,sha256=GTkn7rteG7LZ-9q9GZqbdyesxj5ZTiAgyOI4bdnUc6A,1573
+model_compression_toolkit/__init__.py,sha256=sV3EGoGSi45kJn43REtpqcqQutybRbJUFwnOwC2OByU,1573
 model_compression_toolkit/constants.py,sha256=i4wYheBkIdQmsQA-axIpcT3YiSO1USNc-jaNiNE8w6E,3920
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -8,10 +8,10 @@ model_compression_toolkit/core/__init__.py,sha256=tnDtL9KmT0vsOU27SsJ19TKDEbIH-t
 model_compression_toolkit/core/analyzer.py,sha256=X-2ZpkH1xdXnISnw1yJvXnvV-ssoUh-9LkLISSWNqiY,3691
 model_compression_toolkit/core/graph_prep_runner.py,sha256=7-b7Jd5jBVaXOWg5nSqbEyzBtdaGDbCxs8aqMV6GZ6I,11287
 model_compression_toolkit/core/quantization_prep_runner.py,sha256=OtL6g2rTC5mfdKrkzm47EPPW-voGGVYMYxpy2_sfu1U,6547
-model_compression_toolkit/core/runner.py,sha256=lahkYyfdsb3HJPJ5Lui7hp4vVWyIOJLXJQ5ATxiIyos,14264
+model_compression_toolkit/core/runner.py,sha256=IavCZRVG9RisEKvFDxz27WDRKrfIG03YKXKv3tcagPo,14700
 model_compression_toolkit/core/common/__init__.py,sha256=Wh127PbXcETZX_d1PQqZ71ETK3J9XO5A-HpadGUbj6o,1447
 model_compression_toolkit/core/common/base_substitutions.py,sha256=xDFSmVVs_iFSZfajytI0cuQaNRNcwHX3uqOoHgVUvxQ,1666
-model_compression_toolkit/core/common/framework_implementation.py,sha256=7k66e0b06eLnFLmu67onWPiM2lJfhWiuyQZPsRJm3lk,21294
+model_compression_toolkit/core/common/framework_implementation.py,sha256=IkMydCj6voau7dwkYLYA_Ka_EFUKP3GKQdpYN6b1fgc,22163
 model_compression_toolkit/core/common/framework_info.py,sha256=1ZMMGS9ip-kSflqkartyNRt9aQ5ub1WepuTRcTy-YSQ,6337
 model_compression_toolkit/core/common/memory_computation.py,sha256=ixoSpV5ZYZGyzhre3kQcvR2sNA8KBsPZ3lgbkDnw9Cs,1205
 model_compression_toolkit/core/common/model_builder_mode.py,sha256=jll9-59OPaE3ug7Y9-lLyV99_FoNHxkGZMgcm0Vkpss,1324
@@ -105,8 +105,8 @@ model_compression_toolkit/core/common/quantization/candidate_node_quantization_c
 model_compression_toolkit/core/common/quantization/core_config.py,sha256=yxCzWqldcHoe8GGxrH0tp99bhrc5jDT7SgZftnMUUBE,2374
 model_compression_toolkit/core/common/quantization/debug_config.py,sha256=zJP2W9apUPX9RstpPWWK71wr9xJsg7j-s7lGV4_bQdc,1510
 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=YycYN8_JMzvSR3pTVm5dT5x4zP3yBHn0Z9agnwrvOKI,26395
-model_compression_toolkit/core/common/quantization/quantization_config.py,sha256=BTDa1Izpdd4Z4essxTWP42V87f8mdq9vdKdVhE8vibo,3818
+model_compression_toolkit/core/common/quantization/node_quantization_config.py,sha256=HmtyIQCQqhay-8oqU3rUHOeK6VhTtH9nuW24HigCUo0,26517
+model_compression_toolkit/core/common/quantization/quantization_config.py,sha256=nBqwNhbDbWQGYbfazLPHrP_ZCCnjbL-k5q58T8yIAcc,3917
 model_compression_toolkit/core/common/quantization/quantization_fn_selection.py,sha256=eyosbVdnCwed7oMQ19tqnh0VoyGZ_UAuD_UnNoXyBpo,2210
 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
@@ -128,10 +128,12 @@ model_compression_toolkit/core/common/quantization/quantizers/lut_kmeans_quantiz
 model_compression_toolkit/core/common/quantization/quantizers/quantizers_helpers.py,sha256=iEoWUPFQMcvZXHtLMe2_7L7IK25XcKiY6-d1_gArZs0,11880
 model_compression_toolkit/core/common/quantization/quantizers/uniform_quantizers.py,sha256=wXExWHf5-0He7L4bpvFpKlx7FG4u3DAfNZiXPpOs_SQ,5521
 model_compression_toolkit/core/common/statistics_correction/__init__.py,sha256=sw7LOPN1bM82o3SkMaklyH0jw-TLGK0-fl2Wq73rffI,697
+model_compression_toolkit/core/common/statistics_correction/apply_activation_bias_correction_to_graph.py,sha256=Aw2N7FSO7p1Kmh-tUjajV9pqrjMJQtgF5etG0WV9Le8,4440
 model_compression_toolkit/core/common/statistics_correction/apply_bias_correction_to_graph.py,sha256=xSWVDOODgbN0k4mjJWWtpawilOsqdm4O7Uw2hbA75EA,4669
 model_compression_toolkit/core/common/statistics_correction/apply_second_moment_correction_to_graph.py,sha256=C_nwhhitTd1pCto0nHZPn3fjIMOeDD7VIciumTR3s6k,5641
+model_compression_toolkit/core/common/statistics_correction/compute_activation_bias_correction_of_graph.py,sha256=ov9-WYktWKqRquibwyARR81QVT9TfPWAoTTfnKOQSd0,9273
 model_compression_toolkit/core/common/statistics_correction/compute_bias_correction_of_graph.py,sha256=LaGhYES7HgIDf9Bi2KAG_mBzAWuum0J6AGmAFPC8wwo,10478
-model_compression_toolkit/core/common/statistics_correction/statistics_correction.py,sha256=5dzNtzDMmmLETgAU23k8Cu7q0q3z1EyS-46_Yx-aS7s,5519
+model_compression_toolkit/core/common/statistics_correction/statistics_correction.py,sha256=E0ZA4edimJwpHh9twI5gafcoJ9fX5F1JX2QUOkUOKEw,6250
 model_compression_toolkit/core/common/substitutions/__init__.py,sha256=sw7LOPN1bM82o3SkMaklyH0jw-TLGK0-fl2Wq73rffI,697
 model_compression_toolkit/core/common/substitutions/apply_substitutions.py,sha256=k-bifmakHIYZeZS-4T1QpZ1Et6AwAijMRgAKs7hmMKc,1390
 model_compression_toolkit/core/common/substitutions/batchnorm_folding.py,sha256=wLlTT7sqUffKHwOrMG2VV5SktQkkP54l8taW1Fq0mh0,13392
@@ -155,7 +157,7 @@ model_compression_toolkit/core/keras/constants.py,sha256=dh4elQWt6Q6NYRht5k5RiiO
 model_compression_toolkit/core/keras/custom_layer_validation.py,sha256=f-b14wuiIgitBe7d0MmofYhDCTO3IhwJgwrh-Hq_t_U,1192
 model_compression_toolkit/core/keras/data_util.py,sha256=JdomIJZfep0QYPtx2jlg0xJ40cd9S_I7BakaWQi0wKw,2681
 model_compression_toolkit/core/keras/default_framework_info.py,sha256=PYcER89eEXjKtR0T7-2Y4f7cckqoD5OQbpHePoRkMec,5030
-model_compression_toolkit/core/keras/keras_implementation.py,sha256=Tn4_rkcx9bH3x-pEoUbGu94S7_nj3Hl3BfvL8SPIL3g,30957
+model_compression_toolkit/core/keras/keras_implementation.py,sha256=Hi8seiFJdFqgYGGC003Y4879JQ7rmVZe8YiJ76T7FDE,32133
 model_compression_toolkit/core/keras/keras_model_validation.py,sha256=1wNV2clFdC9BzIELRLSO2uKf0xqjLqlkTJudwtCeaJk,1722
 model_compression_toolkit/core/keras/keras_node_prior_info.py,sha256=HUmzEXDQ8LGX7uOYSRiLZ2TNbYxLX9J9IeAa6QYlifg,3927
 model_compression_toolkit/core/keras/resource_utilization_data_facade.py,sha256=s56UIgiPipUQRNd2sd1xW6GFfYNMBmrocRCNtvpYLbY,4977
@@ -214,13 +216,14 @@ model_compression_toolkit/core/keras/reader/nested_model/nodes_merger.py,sha256=
 model_compression_toolkit/core/keras/reader/nested_model/outputs_merger.py,sha256=dUzvNVzamauDLjgyjHweWux6T2vRko3anAuPxnaGpX8,2408
 model_compression_toolkit/core/keras/statistics_correction/__init__.py,sha256=9HIBmj8ROdCA-yvkpA8EcN6RHJe_2vEpLLW_gxOJtak,698
 model_compression_toolkit/core/keras/statistics_correction/apply_second_moment_correction.py,sha256=XNCtT9klMcsO1v5KA3MmCq_WgXOIT5QSzbfTOa9T-04,3060
+model_compression_toolkit/core/keras/statistics_correction/keras_compute_activation_bias_correction_of_graph.py,sha256=lq6yw9r1u0ZGA95JFvzsV-HQax66qAkJBmGeKnG9OrM,3409
 model_compression_toolkit/core/keras/visualization/__init__.py,sha256=mjbqLD-KcG3eNeCYpu1GBS7VclGVOQ63x2p6mAAuba4,698
 model_compression_toolkit/core/pytorch/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
 model_compression_toolkit/core/pytorch/constants.py,sha256=YwD_joIF0vK8UG2vW1NVvg36pCNWA0vHOXjAgy_XWn0,2794
 model_compression_toolkit/core/pytorch/data_util.py,sha256=YYbT135HhlTt0q6XdD2JX7AS_L92f_uV2rWq2hsJOCA,6325
 model_compression_toolkit/core/pytorch/default_framework_info.py,sha256=-Vls1P_8Ckm_18nnOsmQkZ71SmzHwtQLbQ383Z4Rb-U,4365
 model_compression_toolkit/core/pytorch/pytorch_device_config.py,sha256=S25cuw10AW3SEN_fRAGRcG_I3wdvvQx1ehSJzPnn-UI,4404
-model_compression_toolkit/core/pytorch/pytorch_implementation.py,sha256=TWA5Eu_85TIoCii1Owx2yx_ECckOnGg7xgQkiueuZPE,28245
+model_compression_toolkit/core/pytorch/pytorch_implementation.py,sha256=4uzO-lXfuitlC3NHx5-k2Fjm8VHa1T7ox9c8DSxYs9M,29437
 model_compression_toolkit/core/pytorch/pytorch_node_prior_info.py,sha256=2LDQ7qupglHQ7o1Am7LWdfYVacfQnl-aW2N6l9det1w,3264
 model_compression_toolkit/core/pytorch/resource_utilization_data_facade.py,sha256=xpKj99OZKT9NT0vKIl_cOe8d89d2gef1gKoNT6PFElE,4989
 model_compression_toolkit/core/pytorch/utils.py,sha256=7VbgcLwtQvdEEc_AJgSOQ3U3KRKCICFPaBirN1fIQxg,3940
@@ -274,6 +277,7 @@ model_compression_toolkit/core/pytorch/reader/node_holders.py,sha256=7XNc7-l1MZP
 model_compression_toolkit/core/pytorch/reader/reader.py,sha256=GEJE0QX8XJFWbYCkbRBtzttZtmmuoACLx8gw9KyAQCE,6015
 model_compression_toolkit/core/pytorch/statistics_correction/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
 model_compression_toolkit/core/pytorch/statistics_correction/apply_second_moment_correction.py,sha256=VgU24J3jf7QComHH7jonOXSkg6mO4TOch3uFkOthZvM,3261
+model_compression_toolkit/core/pytorch/statistics_correction/pytorch_compute_activation_bias_correction_of_graph.py,sha256=N-9QaEaQYUsIoya9Lc0ZDoMZ0fkiT2gFpOd4zXHKP34,3096
 model_compression_toolkit/data_generation/__init__.py,sha256=9xLN7VE3lnYVjoroYfJ24dxK_-kGEbMmMVeS1PPkPEY,1513
 model_compression_toolkit/data_generation/common/__init__.py,sha256=huHoBUcKNB6BnY6YaUCcFvdyBtBI172ZoUD8ZYeNc6o,696
 model_compression_toolkit/data_generation/common/constants.py,sha256=21e3ZX9WVYojexG2acTgklrBk8ZO9DjJnKpP4KHZC44,1018
@@ -554,24 +558,8 @@ model_compression_toolkit/xquant/pytorch/model_analyzer.py,sha256=b93o800yVB3Z-i
 model_compression_toolkit/xquant/pytorch/pytorch_report_utils.py,sha256=bOc-hFL3gdoSM1Th_S2N_-9JJSlPGpZCTx_QLJHS6lg,3388
 model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
 model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=mkoEktLFFHtEKzzFRn_jCnxjhJolK12TZ5AQeDHzUO8,9767
-tests_pytest/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
-tests_pytest/keras/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
-tests_pytest/keras/core/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
-tests_pytest/keras/core/test_data_util.py,sha256=XSoPu_ci1xy2EtK-3OWGpESr-Meg1GDaxuSvcj3yt-w,3915
-tests_pytest/keras/gptq/__init__.py,sha256=pKAdbTCFM_2BrZXUtTIw0ouKotrWwUDF_hP3rPwCM2k,696
-tests_pytest/keras/gptq/test_gradual_act_quantization.py,sha256=iwKaLI7QQ8H3qj6zmwwfd2ZOwRcCr8T-v_4llSh_chM,4804
-tests_pytest/keras/trainable_infrastructure/__init__.py,sha256=DvaMXJtJZHAqOm96NdfBiNQsbN2sc9bG2kkyY-mpPh8,710
-tests_pytest/keras/trainable_infrastructure/test_linear_annealing.py,sha256=dZjrMHVIiEVRNDYR3a4lZaXF2ElxFx32KAXXQvDz-v8,1793
-tests_pytest/pytorch/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
-tests_pytest/pytorch/core/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
-tests_pytest/pytorch/core/test_data_util.py,sha256=Bg3c21YVfXE1SAUlTao553gXcITTKF4CPeKtl3peBTE,5604
-tests_pytest/pytorch/gptq/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
-tests_pytest/pytorch/gptq/test_annealing_cfg.py,sha256=hGC7L6mp3N1ygcJ3OctgS_Fz2JY75q5aswolJkbHkZM,2208
-tests_pytest/pytorch/gptq/test_gradual_act_quantization.py,sha256=Dg2cg1X8u9Jxm7Y6tlZIGH81EPoW_vYorcdDExdj02w,4630
-tests_pytest/pytorch/trainable_infrastructure/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
-tests_pytest/pytorch/trainable_infrastructure/test_linear_annealing.py,sha256=zErt9tOu7oupjpv08cvd1Cxvdk9qvP7GMUP6EhefK0c,1814
-mct_nightly-2.2.0.20241106.458.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.2.0.20241106.458.dist-info/METADATA,sha256=OvQyuNyvb2ucuyM03TFlWlAicuXkgODpKoR9u4zQ8NI,20830
-mct_nightly-2.2.0.20241106.458.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-mct_nightly-2.2.0.20241106.458.dist-info/top_level.txt,sha256=csdfSXhtRnpWYRzjZ-dRLIhOmM2TEdVXUxG05A5fgb8,39
-mct_nightly-2.2.0.20241106.458.dist-info/RECORD,,
+mct_nightly-2.2.0.20241108.459.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+mct_nightly-2.2.0.20241108.459.dist-info/METADATA,sha256=BKcV37WvzOfa_tSqxVnw6l2_7oSMLjrOFCVEizArBSE,20830
+mct_nightly-2.2.0.20241108.459.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+mct_nightly-2.2.0.20241108.459.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.2.0.20241108.459.dist-info/RECORD,,

{mct_nightly-2.2.0.20241106.458.dist-info → mct_nightly-2.2.0.20241108.459.dist-info}/top_level.txt

@@ -1,2 +1 @@
 model_compression_toolkit
-tests_pytest

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__ = "2.2.0.20241106.000458"
+__version__ = "2.2.0.20241108.000459"

model_compression_toolkit/core/common/framework_implementation.py

@@ -64,7 +64,7 @@ class FrameworkImplementation(ABC):
 
         Returns: HessianScoresCalculator to use for the hessian approximation scores computation for this request.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_hessian_scores_calculator method.')  # pragma: no cover
 
     @abstractmethod
@@ -77,7 +77,7 @@ class FrameworkImplementation(ABC):
         Returns:
             Numpy array converted from the input tensor.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s to_numpy method.')  # pragma: no cover
 
     @abstractmethod
@@ -90,7 +90,7 @@ class FrameworkImplementation(ABC):
         Returns:
             Framework's tensor converted from the input Numpy array.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s to_tensor method.')  # pragma: no cover
 
     @abstractmethod
@@ -106,7 +106,7 @@ class FrameworkImplementation(ABC):
         Returns:
             Graph representing the input model.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s model_reader method.')  # pragma: no cover
 
     @abstractmethod
@@ -131,7 +131,7 @@ class FrameworkImplementation(ABC):
         Returns:
             A tuple with the model and additional relevant supporting objects.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s model_builder method.')  # pragma: no cover
 
     @abstractmethod
@@ -148,7 +148,7 @@ class FrameworkImplementation(ABC):
         Returns:
             The frameworks model's output.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s run_model_inference method.')  # pragma: no cover
 
     @abstractmethod
@@ -167,9 +167,28 @@ class FrameworkImplementation(ABC):
         Returns:
             Graph after SNC.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s apply_shift_negative_correction method.')  # pragma: no cover
 
+    @abstractmethod
+    def compute_activation_bias_correction(self,
+                                           graph: Graph,
+                                           quant_config: QuantizationConfig,
+                                           fw_info: FrameworkInfo) -> Graph:
+        """
+        Compute activation bias correction on a graph.
+
+        Args:
+            graph: Graph to apply activation bias correction on.
+            quant_config: QuantizationConfig of how the model should be quantized.
+            fw_info: FrameworkInfo object with information about the specific framework's model.
+
+        Returns:
+            Graph after activation bias correction computing.
+        """
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
+                                  f'framework\'s compute_activation_bias_correction method.')  # pragma: no cover
+
     @abstractmethod
     def get_substitutions_channel_equalization(self,
                                                quant_config: QuantizationConfig,
@@ -184,7 +203,7 @@ class FrameworkImplementation(ABC):
         Returns:
             A list of the framework substitutions used after we collect statistics.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_substitutions_channel_equalization method.')  # pragma: no cover
 
     @abstractmethod
@@ -194,7 +213,7 @@ class FrameworkImplementation(ABC):
         Returns: A list of the framework substitutions used to prepare the graph.
 
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_substitutions_prepare_graph method.')  # pragma: no cover
 
     @abstractmethod
@@ -208,7 +227,7 @@ class FrameworkImplementation(ABC):
         Returns: A list of the framework substitutions used before we collect statistics.
 
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_substitutions_pre_statistics_collection method.')  # pragma: no cover
 
     @abstractmethod
@@ -216,7 +235,7 @@ class FrameworkImplementation(ABC):
         """
         Returns: linear collapsing substitution
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_linear_collapsing_substitution method.')  # pragma: no cover
 
     @abstractmethod
@@ -224,7 +243,7 @@ class FrameworkImplementation(ABC):
         """
         Returns: conv2d add const collapsing substitution
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_op2d_add_const_collapsing_substitution method.')  # pragma: no cover
 
     @abstractmethod
@@ -239,7 +258,7 @@ class FrameworkImplementation(ABC):
         Returns:
             A list of the framework substitutions used for statistics correction.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_substitutions_statistics_correction method.')  # pragma: no cover
 
     @abstractmethod
@@ -247,7 +266,7 @@ class FrameworkImplementation(ABC):
         """
         Returns: A list of the framework substitutions used for residual collapsing
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_residual_collapsing_substitution method.')  # pragma: no cover
 
 
@@ -263,7 +282,7 @@ class FrameworkImplementation(ABC):
         Returns:
             A list of the framework substitutions used after we collect statistics.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_substitutions_post_statistics_collection method.')  # pragma: no cover
 
     @abstractmethod
@@ -272,7 +291,7 @@ class FrameworkImplementation(ABC):
         Returns: A list of Keras substitutions used to build a virtual graph with composed activation-weights pairs.
         """
 
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_substitutions_virtual_weights_activation_coupling '
                              f'method.')  # pragma: no cover
 
@@ -288,7 +307,7 @@ class FrameworkImplementation(ABC):
         Returns:
             A list of the framework substitutions used after we apply second moment statistics.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_substitutions_after_second_moment_correction '
                              f'method.')  # pragma: no cover
 
@@ -316,7 +335,7 @@ class FrameworkImplementation(ABC):
             A function that computes the metric.
         """
 
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_sensitivity_evaluator method.')  # pragma: no cover
 
     def get_node_prior_info(self, node: BaseNode,
@@ -334,7 +353,7 @@ class FrameworkImplementation(ABC):
             NodePriorInfo with information about the node.
         """
 
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_node_prior_info method.')  # pragma: no cover
 
     def count_node_for_mixed_precision_interest_points(self, node: BaseNode) -> bool:
@@ -345,7 +364,7 @@ class FrameworkImplementation(ABC):
         Returns: True if the node should be considered an interest point, False otherwise.
         """
 
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s count_node_for_mixed_precision_interest_points method.')  # pragma: no cover
 
     def get_mp_node_distance_fn(self, n: BaseNode,
@@ -364,7 +383,7 @@ class FrameworkImplementation(ABC):
         Returns: A distance function between two tensors and a axis on which the distance is computed (if exists).
         """
 
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_mp_node_distance_fn method.')  # pragma: no cover
 
 
@@ -381,7 +400,7 @@ class FrameworkImplementation(ABC):
 
         """
 
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s is_output_node_compatible_for_hessian_score_computation method.')  # pragma: no cover
 
     @abstractmethod
@@ -398,7 +417,7 @@ class FrameworkImplementation(ABC):
         Returns: The MAC count of the operation
         """
 
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_node_mac_operations method.')  # pragma: no cover
 
     @abstractmethod
@@ -419,7 +438,7 @@ class FrameworkImplementation(ABC):
         Returns:
             A Graph after second moment correction.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s apply_second_moment_correction method.')  # pragma: no cover
 
     @abstractmethod
@@ -436,7 +455,7 @@ class FrameworkImplementation(ABC):
         Returns:
             The output of the model inference on the given input.
         """
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s sensitivity_eval_inference method.')  # pragma: no cover
 
     def get_inferable_quantizers(self, node: BaseNode):
@@ -452,9 +471,9 @@ class FrameworkImplementation(ABC):
 
         """
 
-        raise NotImplementedError(f'{self.__class__.__name__} have to implement the '
+        raise NotImplementedError(f'{self.__class__.__name__} has to implement the '
                              f'framework\'s get_inferable_quantizers method.')  # pragma: no cover
-    
+
     @staticmethod
     def convert_data_gen_to_dataloader(data_gen_fn: Callable[[], Generator], batch_size: int):
         """

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -95,7 +95,9 @@ class NodeActivationQuantizationConfig(BaseNodeQuantizationConfig):
         self.activation_error_method = qc.activation_error_method
         self.activation_n_bits = op_cfg.activation_n_bits
         self.relu_bound_to_power_of_2 = qc.relu_bound_to_power_of_2
+        self.activation_bias_correction_term = None
         self.enable_activation_quantization = op_cfg.enable_activation_quantization
+        self.quantization_preserving = op_cfg.quantization_preserving
         self.signedness = op_cfg.signedness
         self.activation_channel_equalization = qc.activation_channel_equalization
         self.input_scaling = qc.input_scaling

model_compression_toolkit/core/common/quantization/quantization_config.py

@@ -84,6 +84,8 @@ class QuantizationConfig:
     shift_negative_threshold_recalculation: bool = False
     shift_negative_params_search: bool = False
     concat_threshold_update: bool = False
+    activation_bias_correction: bool = False
+    activation_bias_correction_threshold: float = 0.0
 
 
 # Default quantization configuration the library use.

model_compression_toolkit/core/common/statistics_correction/apply_activation_bias_correction_to_graph.py

@@ -0,0 +1,81 @@
+# 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 model_compression_toolkit.core import CoreConfig, QuantizationConfig
+from model_compression_toolkit.core.common import BaseNode, Graph
+from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
+from model_compression_toolkit.core.common.quantization.node_quantization_config import WeightsAttrQuantizationConfig
+from model_compression_toolkit.target_platform_capabilities.target_platform import AttributeQuantizationConfig
+
+
+def apply_activation_bias_correction_to_graph(graph: Graph,
+                                              core_config: CoreConfig,
+                                              fw_impl: FrameworkImplementation) -> Graph:
+    """
+    Get a graph, where each node has a final activation quantization configuration (with an activation bias
+    correction term in it), and apply the activation bias correction for each node in the graph.
+
+    Args:
+        graph: Graph to apply activation bias correction to.
+        core_config: CoreConfig containing parameters of how the model should be quantized.
+        fw_impl: FrameworkImplementation object with a specific framework methods implementation.
+
+    Returns:
+        Graph with activation bias correction apply to it's nodes.
+    """
+
+    for n in graph.nodes:
+        # Activation bias correction is only relevant for nodes with kernel op
+        kernel_attr = graph.fw_info.get_kernel_op_attributes(n.type)[0]
+        if core_config.quantization_config.activation_bias_correction and kernel_attr is not None and \
+                n.final_activation_quantization_cfg.activation_bias_correction_term is not None:
+            # If activation bias correction is enabled in n.quantization_cfg, an activation bias correction term was
+            # calculated during model preparation, and is used now in the node's bias term.
+            _apply_activation_bias_correction_to_node(n, fw_impl, core_config.quantization_config)
+    return graph
+
+
+def _apply_activation_bias_correction_to_node(node: BaseNode,
+                                              fw_impl: FrameworkImplementation,
+                                              qc: QuantizationConfig):
+    """
+    Set new bias to node using the activation bias correction term that is stored in the
+    final activation quantization configuration.
+
+    Args:
+        node: Node to set its corrected bias after activation bias correction.
+        fw_impl: FrameworkImplementation object with a specific framework methods implementation.
+        qc: QuantizationConfig containing parameters of how the model should be quantized.
+
+    """
+    correction = node.final_activation_quantization_cfg.activation_bias_correction_term
+    bias = node.get_weights_by_keys(fw_impl.constants.BIAS)  # get original bias from node's weights
+
+    if bias is None:
+        # If the layer has no bias, we set the bias as -correction.
+        node.set_weights_by_keys(fw_impl.constants.BIAS, - correction)
+
+        # Mark the use_bias attribute of the node.
+        node.framework_attr[fw_impl.constants.USE_BIAS] = True
+
+        # Configure the quantization of the bias as disabled.
+        node.final_weights_quantization_cfg.set_attr_config(fw_impl.constants.BIAS,
+                                                            WeightsAttrQuantizationConfig(
+                                                                qc,
+                                                                AttributeQuantizationConfig(
+                                                                    enable_weights_quantization=False)))
+    else:
+        # If the layer has bias, we subtract the correction from original bias
+        node.set_weights_by_keys(fw_impl.constants.BIAS, bias - correction)