mct-nightly 2.0.0.20240407.442__py3-none-any.whl → 2.0.0.20240409.404__py3-none-any.whl

{mct_nightly-2.0.0.20240407.442.dist-info → mct_nightly-2.0.0.20240409.404.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240407.442
+Version: 2.0.0.20240409.404
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct_nightly-2.0.0.20240407.442.dist-info → mct_nightly-2.0.0.20240409.404.dist-info}/RECORD

@@ -1,9 +1,9 @@
-model_compression_toolkit/__init__.py,sha256=D8v4eaxAbnPL5dY8Em333z-gdWTtL6LPQkUT3wBaQdQ,1573
+model_compression_toolkit/__init__.py,sha256=ALvOQYWLrTHNtxDnpNxy7lyftsvgDpzcoW-wTFtMedY,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
 model_compression_toolkit/core/__init__.py,sha256=TrRgkWpT1AN2Faw1M_1HXyJkJnbxfn9p-RigDZl7pg0,1982
-model_compression_toolkit/core/analyzer.py,sha256=dbsD61pakp_9JXNyAScLdtJvcXny9jr_cMbET0Bd3Sg,2975
+model_compression_toolkit/core/analyzer.py,sha256=X-2ZpkH1xdXnISnw1yJvXnvV-ssoUh-9LkLISSWNqiY,3691
 model_compression_toolkit/core/graph_prep_runner.py,sha256=Ftqm59hT5TGWmSNkY9bFZkVfCacpGyZfCe-6yZR5WY0,10100
 model_compression_toolkit/core/quantization_prep_runner.py,sha256=hFhDkS8GwzXZ7Ho_9qbbb8DAAWs3OONOfMSD5OU_b0o,6153
 model_compression_toolkit/core/runner.py,sha256=NKSC6ujfQPy6dKtJVwxyK2zNDd64eyR5csYy9lBrCPA,11836
@@ -16,7 +16,7 @@ model_compression_toolkit/core/common/model_builder_mode.py,sha256=jll9-59OPaE3u
 model_compression_toolkit/core/common/model_collector.py,sha256=ofcepKtxc3j2Ouz6BpAKXTzPgjABnpRP47ndmJCXAkk,8352
 model_compression_toolkit/core/common/model_validation.py,sha256=LaG8wd6aZl0OJgieE3SeiVDEPxtk8IHq9-3wSnmWhY4,1214
 model_compression_toolkit/core/common/node_prior_info.py,sha256=WXX_PrGVG9M9I_REG5ZzFBohwmV4yf356sZnrja_FLo,2832
-model_compression_toolkit/core/common/similarity_analyzer.py,sha256=2w-q7guEb5bpLY4Vk_TMjR8TzLYEymR3tPFlrVq7K68,8515
+model_compression_toolkit/core/common/similarity_analyzer.py,sha256=98l9ttnXHf6VYxBW4852h2CPJKg3A6nLOovpHn-tnKs,8560
 model_compression_toolkit/core/common/user_info.py,sha256=dSRMnT-oewmdOziIpEuW-s9K7vTSeyUBxT4z9neXurI,1648
 model_compression_toolkit/core/common/back2framework/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
 model_compression_toolkit/core/common/back2framework/base_model_builder.py,sha256=V1oShKzbSkdcTvREn8VnQQBzvm-tTHkWMXqMkYozF2s,2023
@@ -100,11 +100,11 @@ 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=lXONxIOSvYMgkN9M1st4tV1V5JSpijUGxF0hZWRvtUI,26737
+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/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=xnM9O9LshYw3dprqfsnK9mw7ipOEAkI85o20auyfswg,2626
+model_compression_toolkit/core/common/quantization/quantize_graph_weights.py,sha256=N005MSvx8UypVpa7XrxNrB2G732n2wHj3RmLyjTgd3I,2728
 model_compression_toolkit/core/common/quantization/quantize_node.py,sha256=cdzGNWfT4MRogIU8ehs0tr3lVjnzAI-jeoS9b4TwVBo,2854
 model_compression_toolkit/core/common/quantization/set_node_quantization_config.py,sha256=9BEv2l0z2trDEsr40VB8tO3ToBA_b2sd_jH9uqZ5Wo8,11503
 model_compression_toolkit/core/common/quantization/quantization_params_generation/__init__.py,sha256=eCDGwsWYLU6z7qbEVb4TozMW_nd5VEP_iCJ6PcvyEPw,1486
@@ -142,7 +142,7 @@ model_compression_toolkit/core/common/substitutions/virtual_activation_weights_c
 model_compression_toolkit/core/common/substitutions/weights_activation_split.py,sha256=h85L2VlDOqbLd-N98wA3SdYWiblBgSsPceNuLanJd70,4737
 model_compression_toolkit/core/common/visualization/__init__.py,sha256=mjbqLD-KcG3eNeCYpu1GBS7VclGVOQ63x2p6mAAuba4,698
 model_compression_toolkit/core/common/visualization/final_config_visualizer.py,sha256=6I10jKLesB-RQKaXA75Xgz2wPvylQUrnPtCcQZIynGo,6371
-model_compression_toolkit/core/common/visualization/nn_visualizer.py,sha256=jjsOng-fLxeQFQNshIsOu_w1d5a3fJ359Hcnt85Te-o,5921
+model_compression_toolkit/core/common/visualization/nn_visualizer.py,sha256=HOq7AObkmEZiDSZXUMJDAEJzUY-fSXUT0AMgwiyH7dg,7388
 model_compression_toolkit/core/common/visualization/tensorboard_writer.py,sha256=4E4ZXZmqusGIJ4XQNH8FFt07htAHgT3gy5E7wPIaVBI,21951
 model_compression_toolkit/core/keras/__init__.py,sha256=mjbqLD-KcG3eNeCYpu1GBS7VclGVOQ63x2p6mAAuba4,698
 model_compression_toolkit/core/keras/constants.py,sha256=Uv3c0UdW55pIVQNW_1HQlgl-dHXREkltOLyzp8G1mTQ,3163
@@ -336,9 +336,9 @@ model_compression_toolkit/gptq/common/gptq_training.py,sha256=rLA1xlOO-6gWfmc2dL
 model_compression_toolkit/gptq/keras/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
 model_compression_toolkit/gptq/keras/gptq_keras_implementation.py,sha256=axBwnCSjq5xk-xGymOwSOqjp39It-CVtGcCTRTf0E_4,1248
 model_compression_toolkit/gptq/keras/gptq_loss.py,sha256=rbRkF15MYd6nq4G49kcjb_dPTa-XNq9cTkrb93mXawo,6241
-model_compression_toolkit/gptq/keras/gptq_training.py,sha256=6TseqtBzZkLqyc3hiRVdA1dv01us6Y_Su05CBboGjjc,18438
+model_compression_toolkit/gptq/keras/gptq_training.py,sha256=OhYfH6zxRHrRhCde0lbcV9Hu2oeDD9RXh-O8vOPgLbs,18875
 model_compression_toolkit/gptq/keras/graph_info.py,sha256=5IvgGlJlgOmQYmldjdCBv7tuzAoY0HazatG5Pedrg0Q,4639
-model_compression_toolkit/gptq/keras/quantization_facade.py,sha256=yRgiikyeELar4jlsdcf5pO9HQcxiyhKiAXY3lsMixew,13913
+model_compression_toolkit/gptq/keras/quantization_facade.py,sha256=zAkzWpWP9_aobWgMo_BlUm7-4fR5dHvoGx0sDqs2rZg,14299
 model_compression_toolkit/gptq/keras/quantizer/__init__.py,sha256=-DK1CDXvlsnEbki4lukZLpl6Xrbo91_jcqxXlG5Eg6Q,963
 model_compression_toolkit/gptq/keras/quantizer/base_keras_gptq_quantizer.py,sha256=2YU-x4-Q5f6hkUJf0tw6vcwdNwRMHdefrFjhhyHYsvA,4782
 model_compression_toolkit/gptq/keras/quantizer/quant_utils.py,sha256=Vt7Qb8i4JsE4sFtcjpfM4FTXTtfV1t6SwfoNH8a_Iaw,5055
@@ -353,9 +353,9 @@ model_compression_toolkit/gptq/keras/quantizer/ste_rounding/symmetric_ste.py,sha
 model_compression_toolkit/gptq/pytorch/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
 model_compression_toolkit/gptq/pytorch/gptq_loss.py,sha256=kDuWw-6zh17wZpYWh4Xa94rpoodf82DksgjQCnL7nBc,2719
 model_compression_toolkit/gptq/pytorch/gptq_pytorch_implementation.py,sha256=tECPTavxn8EEwgLaP2zvxdJH6Vg9jC0YOIMJ7857Sdc,1268
-model_compression_toolkit/gptq/pytorch/gptq_training.py,sha256=ksFXtepAsk-66xk7OPciG05kU9sgAUrWqjOgplsGSnw,15808
+model_compression_toolkit/gptq/pytorch/gptq_training.py,sha256=LN4vOwcMuSSFTSnHDACV9hX_Yd2YIXJRl7WkdODuA0k,16245
 model_compression_toolkit/gptq/pytorch/graph_info.py,sha256=yXJzDd24zfGs2_vfMovxD1WSh1RxXoPxN4GztOf3P5c,3967
-model_compression_toolkit/gptq/pytorch/quantization_facade.py,sha256=bImrTw9rrAVc3VD4nmrXmBo_K4fuf5m5XPPf8ybOThs,12430
+model_compression_toolkit/gptq/pytorch/quantization_facade.py,sha256=-4USg-tep6EQSArcTxBowhMeAuExrBTNLOWgHFpsIy4,12699
 model_compression_toolkit/gptq/pytorch/quantizer/__init__.py,sha256=ZHNHo1yzye44m9_ht4UUZfTpK01RiVR3Tr74-vtnOGI,968
 model_compression_toolkit/gptq/pytorch/quantizer/base_pytorch_gptq_quantizer.py,sha256=TCA1hAc7raPnrjl06sjFtVM4XUtLtuwAhCGX4U3KGZo,4137
 model_compression_toolkit/gptq/pytorch/quantizer/quant_utils.py,sha256=OocYYRqvl7rZ37QT0hTzfJnWGiNCPskg7cziTlR7TRk,3893
@@ -375,9 +375,9 @@ model_compression_toolkit/pruning/pytorch/pruning_facade.py,sha256=cSuvHHCqgr7k9
 model_compression_toolkit/ptq/__init__.py,sha256=Z_hkmTh7aLFei1DJKV0oNVUbrv_Q_0CTw-qD85Xf8UM,904
 model_compression_toolkit/ptq/runner.py,sha256=_c1dSjlPPpsx59Vbg1buhG9bZq__OORz1VlPkwjJzoc,2552
 model_compression_toolkit/ptq/keras/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
-model_compression_toolkit/ptq/keras/quantization_facade.py,sha256=5Doa5Rwer84DRgJxLa2e6aX9B4yGYdmFGgiv71_wD9o,8992
+model_compression_toolkit/ptq/keras/quantization_facade.py,sha256=T1_UqXmOc4I2a6IHkQAlFhGtcAYjsXSApMIdRlvgDvg,10154
 model_compression_toolkit/ptq/pytorch/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
-model_compression_toolkit/ptq/pytorch/quantization_facade.py,sha256=QW7lMRuuonupLPZ2w2PDIQd7qpDZ_euLInhskTc1Yes,7518
+model_compression_toolkit/ptq/pytorch/quantization_facade.py,sha256=eof9bo-Mv_lLY7fFpiVeT5pIde-MuTWkIAqRKH4j9MI,8646
 model_compression_toolkit/qat/__init__.py,sha256=kj2qsZh_Ca7PncsHKcaL5EVT2H8g4hYtvaQ3KFxOkwE,1143
 model_compression_toolkit/qat/common/__init__.py,sha256=6tLZ4R4pYP6QVztLVQC_jik2nES3l4uhML0qUxZrezk,829
 model_compression_toolkit/qat/common/qat_config.py,sha256=zoq0Vb74vCY7WlWD8JH_KPrHDoUHSvMc3gcO53u7L2U,3394
@@ -469,8 +469,8 @@ 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-2.0.0.20240407.442.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.0.0.20240407.442.dist-info/METADATA,sha256=VYYUCAhW6RmEexLHQghz4uGkGFB0JlYg9R4CkDCiEs0,18795
-mct_nightly-2.0.0.20240407.442.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-mct_nightly-2.0.0.20240407.442.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.0.0.20240407.442.dist-info/RECORD,,
+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,,

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__ = "2.0.0.20240407.000442"
+__version__ = "2.0.0.20240409.000404"

model_compression_toolkit/core/analyzer.py

@@ -30,7 +30,8 @@ from model_compression_toolkit.logger import Logger
 
 def analyzer_model_quantization(representative_data_gen: Callable,
                                 tb_w: TensorboardWriter,
-                                tg: Graph,
+                                float_graph: Graph,
+                                quantized_graph: Graph,
                                 fw_impl: FrameworkImplementation,
                                 fw_info: FrameworkInfo):
     """
@@ -41,23 +42,32 @@ def analyzer_model_quantization(representative_data_gen: Callable,
     Args:
         representative_data_gen: Dataset used for calibration.
         tb_w: TensorBoardWriter object to log events.
-        tg: Graph of quantized model.
+        float_graph: Graph of float model.
+        quantized_graph: Graph of quantized model.
         fw_impl: FrameworkImplementation object with a specific framework methods implementation.
         fw_info: Information needed for quantization about the specific framework.
 
     """
     if tb_w is not None:
-        visual = NNVisualizer(tg,
+        visual = NNVisualizer(float_graph,
+                              quantized_graph,
                               fw_impl=fw_impl,
                               fw_info=fw_info)
-
-        for i, _data in enumerate(representative_data_gen()):
-            if i >= NUM_SAMPLES_DISTANCE_TENSORBOARD:
-                break
-            figure = visual.plot_distance_graph(_data,
-                                                distance_fn=compute_cs,
-                                                convert_to_range=lambda a: 1 - 2 * a)
-            tb_w.add_figure(figure, f'similarity_distance_sample_{i}')
+        if not visual.has_compare_points():
+            Logger.error(f'No comparing points were found to plot analyze similarity.')
         else:
-            Logger.warning(f'Not enough batches in representative dataset to generate {NUM_SAMPLES_DISTANCE_TENSORBOARD} figures')
+            visualized_samples = 0
+            for _data in representative_data_gen():
+                batch_size = _data[0].shape[0]
+                for sample_index in range(batch_size):
+                    if visualized_samples >= NUM_SAMPLES_DISTANCE_TENSORBOARD:
+                        break
+                    figure = visual.plot_distance_graph(_data,
+                                                        sample_index=sample_index,
+                                                        distance_fn=compute_cs,
+                                                        convert_to_range=lambda a: 1 - 2 * a)
+                    tb_w.add_figure(figure, f'similarity_distance_sample_{visualized_samples}')
+                    visualized_samples += 1
+            if visualized_samples < NUM_SAMPLES_DISTANCE_TENSORBOARD:
+                Logger.error(f'Not enough batches in representative dataset to generate {NUM_SAMPLES_DISTANCE_TENSORBOARD} figures')
         tb_w.close()

model_compression_toolkit/core/common/quantization/node_quantization_config.py

@@ -41,24 +41,24 @@ class BaseNodeQuantizationConfig(object):
     Base class for node quantization configuration
     """
 
-    def set_quant_config_attr(self, parameter_name: str, parameter_value: Any,
+    def set_quant_config_attr(self, config_parameter_name: str, config_parameter_value: Any,
                               *args: List[Any], **kwargs: Dict[str, Any]):
         """
         Changes a BaseNodeQuantizationConfig's parameter.
         Note that arg and kwargs are only to allow clean override in the child classes.
 
         Args:
-            parameter_name: parameter name to change.
-            parameter_value: parameter value to change.
+            config_parameter_name: parameter name to change.
+            config_parameter_value: parameter value to change.
             args: A list of additional arguments.
             kwargs: A dictionary with additional key arguments.
 
         """
 
-        if hasattr(self, parameter_name):
-            setattr(self, parameter_name, parameter_value)
+        if hasattr(self, config_parameter_name):
+            setattr(self, config_parameter_name, config_parameter_value)
         else:
-            Logger.warning(f"Parameter {parameter_name} could not be found in the node quantization config and "
+            Logger.warning(f"Parameter {config_parameter_name} could not be found in the node quantization config and "
                            f"was not updated!")
 
     def __repr__(self) -> str:
@@ -521,7 +521,7 @@ class NodeWeightsQuantizationConfig(BaseNodeQuantizationConfig):
                            f"{list(attrs_with_name.keys())}.")
         return attrs_with_name
 
-    def set_quant_config_attr(self, parameter_name: str, parameter_value: Any, attr_name: str = None,
+    def set_quant_config_attr(self, config_parameter_name: str, config_parameter_value: Any, attr_name: str = None,
                               *args: List[Any], **kwargs: Dict[str, Any]):
         """
         This method overrides the parent class set_quant_config_attr to enable setting a specific weights
@@ -529,26 +529,27 @@ class NodeWeightsQuantizationConfig(BaseNodeQuantizationConfig):
 
         Args:
             attr_name: attribute name to change.
-            parameter_name: parameter name to change.
-            parameter_value: parameter value to change.
+            config_parameter_name: parameter name to change.
+            config_parameter_value: parameter value to change.
             args: A list of additional arguments.
             kwargs: A dictionary with additional key arguments.
 
         """
 
         if attr_name is None:
-            super(NodeWeightsQuantizationConfig, self).set_quant_config_attr(parameter_name, parameter_value,
+            super(NodeWeightsQuantizationConfig, self).set_quant_config_attr(config_parameter_name,
+                                                                             config_parameter_value,
                                                                              *args, **kwargs)
         else:
             if self.has_attribute_config(attr_name):
                 attr_cfg = self.get_attr_config(attr_name)
-                if hasattr(attr_cfg, parameter_name):
-                    setattr(attr_cfg, parameter_name, parameter_value)
+                if hasattr(attr_cfg, config_parameter_name):
+                    setattr(attr_cfg, config_parameter_name, config_parameter_value)
                 else:
-                    Logger.warning(f"Parameter {parameter_name} could not be found in the node quantization config of "
+                    Logger.warning(f"Parameter {config_parameter_name} could not be found in the node quantization config of "
                                    f"weights attribute {attr_name} and was not updated!")
             else:
-                Logger.error(f"Weights attribute {attr_name} could not be found to set parameter {parameter_name}.")
+                Logger.error(f"Weights attribute {attr_name} could not be found to set parameter {config_parameter_name}.")
 
     def __eq__(self, other: Any) -> bool:
         """

model_compression_toolkit/core/common/quantization/quantize_graph_weights.py

@@ -23,7 +23,7 @@ from model_compression_toolkit.core.common.quantization.quantize_node import get
 from model_compression_toolkit.logger import Logger
 
 
-def quantize_graph_weights(graph: Graph) -> Graph:
+def quantize_graph_weights(graph_to_quantize: Graph) -> Graph:
     """
     Get a graph representing a model, and quantize its nodes' weights.
     Each node is quantized according to the passed framework info and quantization configuration.
@@ -31,12 +31,13 @@ def quantize_graph_weights(graph: Graph) -> Graph:
     is calculated and subtracted from the original node's bias. The graph is quantized in-place.
 
     Args:
-        graph: Graph to quantize its nodes.
+        graph_to_quantize: Graph to quantize its nodes.
 
     """
+    _quantized_graph = copy.deepcopy(graph_to_quantize)
     # Iterate over nodes in the graph and quantize each node's weights and activations
     # (according to operators groups in framework info).
-    for n in graph.nodes():
+    for n in _quantized_graph.nodes():
         for attr in n.get_node_weights_attributes():
             if n.is_weights_quantization_enabled(attr):
                 quantized_attr, io_channels_axes = \
@@ -51,4 +52,4 @@ def quantize_graph_weights(graph: Graph) -> Graph:
                 # Set the attribute to be the quantized attribute.
                 n.set_weights_by_keys(attr, quantized_attr)
 
-    return graph
+    return _quantized_graph

model_compression_toolkit/core/common/similarity_analyzer.py

@@ -146,7 +146,10 @@ def compute_mae(float_tensor: np.ndarray,
     return error
 
 
-def compute_cs(float_tensor: np.ndarray, fxp_tensor: np.ndarray, eps: float = 1e-8, batch: bool = False,
+def compute_cs(float_tensor: np.ndarray,
+               fxp_tensor: np.ndarray,
+               eps: float = 1e-8,
+               batch: bool = False,
                axis: int = None) -> float:
     """
     Compute the similarity between two tensor using cosine similarity.

model_compression_toolkit/core/common/visualization/nn_visualizer.py

@@ -12,19 +12,20 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+
 from typing import Tuple, List, Callable
 
 import numpy as np
 from matplotlib import pyplot as plt
 from matplotlib.figure import Figure
 
-from model_compression_toolkit.core.common.quantization.quantize_graph_weights import quantize_graph_weights
 from model_compression_toolkit.core.common import Graph
 from model_compression_toolkit.core.common.framework_implementation import FrameworkImplementation
 from model_compression_toolkit.core.common.framework_info import FrameworkInfo
 from model_compression_toolkit.core.common.graph.base_node import BaseNode
 from model_compression_toolkit.core.common.model_builder_mode import ModelBuilderMode
 from model_compression_toolkit.core.common.similarity_analyzer import compute_cs
+from model_compression_toolkit.logger import Logger
 
 
 def _get_compare_points(input_graph: Graph) -> Tuple[List[BaseNode], List[str]]:
@@ -57,17 +58,21 @@ class NNVisualizer:
 
     def __init__(self,
                  graph_float: Graph,
+                 graph_quantized: Graph,
                  fw_impl: FrameworkImplementation,
                  fw_info: FrameworkInfo):
         """
         Initialize a NNVisualizer object.
         Args:
             graph_float: Float version of the graph.
+            graph_quantized: Quantized version of the graph.
+            fw_impl: Framework implementation with framework-specific methods implementation.
+            fw_info: Framework info with framework-specific information.
 
         """
 
         self.graph_float = graph_float
-        self.graph_quantized = quantize_graph_weights(graph_float)
+        self.graph_quantized = graph_quantized
         self.fw_impl = fw_impl
         self.fw_info = fw_info
 
@@ -75,6 +80,16 @@ class NNVisualizer:
         self.compare_points, self.compare_points_name = _get_compare_points(self.graph_quantized)
         self.compare_points_float, self.compare_points_name_float = _get_compare_points(self.graph_float)
 
+        if len(self.compare_points) != len(self.compare_points_float):
+            Logger.critical(f"Number of compare points in float and quantized models must be equal but "
+                            f"num of quantized compare points: {len(self.compare_points)} and "
+                            f"num of float compare points: {len(self.compare_points_float)}")
+        if len(self.compare_points_name) != len(self.compare_points_name_float):
+            Logger.critical(f"Number of compare points in float and quantized models must be equal "
+                            f"but num of quantized compare points: {len(self.compare_points_name)}"
+                            f" and num of float compare points: "
+                            f"{len(self.compare_points_name_float)}")
+
         self.quantized_model, _ = self.fw_impl.model_builder(self.graph_quantized,
                                                              mode=ModelBuilderMode.QUANTIZED,
                                                              append2output=self.compare_points,
@@ -85,8 +100,19 @@ class NNVisualizer:
                                                          append2output=self.compare_points_float,
                                                          fw_info=self.fw_info)
 
+    def has_compare_points(self) -> bool:
+        """
+
+        Returns: Whether or not compare points were found.
+
+        """
+        return len(self.compare_points_float) > 0 and len(self.compare_points) > 0 and len(
+            self.compare_points_name_float) > 0 and len(self.compare_points_name) > 0
+
+
     def plot_distance_graph(self,
                             input_image: np.ndarray,
+                            sample_index: int,
                             distance_fn: Callable = compute_cs,
                             convert_to_range: Callable = lambda a: a) -> Figure:
         """
@@ -95,6 +121,7 @@ class NNVisualizer:
 
         Args:
             input_image: Image to use as input to the networks.
+            sample_index: The index of the sample from input_image to use for comparison.
             distance_fn: Distance function to calculate the distance between two tensors.
             convert_to_range: Optional function to move the distance values into a specific range, e.g., when using
                 cosine similarity for distance, use 'lambda a: 1 - 2 * a' to convert the distance values to the range
@@ -108,7 +135,7 @@ class NNVisualizer:
         # to make the difference more noticeable when exists
         new_inputs = []
         for single_input in input_image:
-            img = single_input[0]
+            img = single_input[sample_index]
             new_inputs.append(np.expand_dims(img, axis=0))
 
         # Get outputs
@@ -123,7 +150,7 @@ class NNVisualizer:
 
         # Display the result: distance at every layer's output.
         fig = plt.figure()
-        plt.plot(distance_array)
+        plt.plot(list(range(len(distance_array))), distance_array)
         eps = 0.5
         y_limits = (min(distance_array) - eps, max(distance_array) + eps)
         plt.ylim(y_limits)

model_compression_toolkit/gptq/keras/gptq_training.py

@@ -337,12 +337,16 @@ class KerasGPTQTrainer(GPTQTrainer):
                 node = node[0]
                 kernel_attribute = get_kernel_attribute_name_for_gptq(layer_type=node.type,
                                                                       fw_info=self.fw_info)
+                # TODO: only kernel attributes are currently trained in GPTQ, so only the kernel weights need to be updated.
+                #  To enable GPTQ for other attributes, this code needs to be modified.
                 weights, weight_quant_config, activation_quant_config = \
                     layer.weights_quantizers[kernel_attribute].update_layer_quantization_params(layer)
                 for weight_attr, weight in weights.items():
                     node.set_weights_by_keys(weight_attr, weight.numpy())
-                for config_attr, config_value in weight_quant_config.items():
-                    node.final_weights_quantization_cfg.set_quant_config_attr(config_attr, config_value)
+                for config_parameter_name, config_parameter_value in weight_quant_config.items():
+                    node.final_weights_quantization_cfg.set_quant_config_attr(config_parameter_name,
+                                                                              config_parameter_value,
+                                                                              attr_name=kernel_attribute)
                 for config_attr, config_value in activation_quant_config.items():
                     node.final_activation_quantization_cfg.set_quant_config_attr(config_attr, config_value)
                 if self.gptq_config.train_bias:

model_compression_toolkit/gptq/keras/quantization_facade.py

@@ -12,10 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+import copy
 
 from typing import Callable, Tuple
 from packaging import version
 
+from model_compression_toolkit.core.common.quantization.quantize_graph_weights import quantize_graph_weights
 from model_compression_toolkit.core.common.visualization.tensorboard_writer import init_tensorboard_writer
 from model_compression_toolkit.gptq.common.gptq_constants import REG_DEFAULT
 from model_compression_toolkit.logger import Logger
@@ -210,6 +212,8 @@ if FOUND_TF:
                                                                   target_resource_utilization=target_resource_utilization,
                                                                   tb_w=tb_w)
 
+        float_graph = copy.deepcopy(tg)
+
         tg_gptq = gptq_runner(tg,
                               core_config,
                               gptq_config,
@@ -223,7 +227,12 @@ if FOUND_TF:
         del hessian_info_service
 
         if core_config.debug_config.analyze_similarity:
-            analyzer_model_quantization(representative_data_gen, tb_w, tg_gptq, fw_impl, fw_info)
+            analyzer_model_quantization(representative_data_gen,
+                                        tb_w,
+                                        float_graph,
+                                        tg_gptq,
+                                        fw_impl,
+                                        DEFAULT_KERAS_INFO)
 
         return get_exportable_keras_model(tg_gptq)
 

model_compression_toolkit/gptq/pytorch/gptq_training.py

@@ -284,12 +284,16 @@ class PytorchGPTQTrainer(GPTQTrainer):
                 node = node[0]
                 kernel_attribute = get_kernel_attribute_name_for_gptq(layer_type=node.type,
                                                                       fw_info=self.fw_info)
+                # TODO: only kernel attributes are currently trained in GPTQ, so only the kernel weights need to be updated.
+                #  To enable GPTQ for other attributes, this code needs to be modified.
                 weights, weight_quant_config, activation_quant_config = \
                     layer.weights_quantizers[kernel_attribute].update_layer_quantization_params(layer)
                 for weight_attr, weight in weights.items():
                     node.set_weights_by_keys(weight_attr, self.fw_impl.to_numpy(weight))
-                for config_attr, config_value in weight_quant_config.items():
-                    node.final_weights_quantization_cfg.set_quant_config_attr(config_attr, config_value)
+                for config_parameter_name, config_parameter_value in weight_quant_config.items():
+                    node.final_weights_quantization_cfg.set_quant_config_attr(config_parameter_name,
+                                                                              config_parameter_value,
+                                                                              attr_name=kernel_attribute)
                 for config_attr, config_value in activation_quant_config.items():
                     node.final_activation_quantization_cfg.set_quant_config_attr(config_attr, config_value)
                 if self.gptq_config.train_bias and hasattr(layer.layer, BIAS):

model_compression_toolkit/gptq/pytorch/quantization_facade.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+import copy
+
 from typing import Callable
 from model_compression_toolkit.core import common
 from model_compression_toolkit.constants import FOUND_TORCH
@@ -177,6 +179,7 @@ if FOUND_TORCH:
                                                                      tpc=target_platform_capabilities,
                                                                      target_resource_utilization=target_resource_utilization,
                                                                      tb_w=tb_w)
+        float_graph = copy.deepcopy(graph)
 
         # ---------------------- #
         # GPTQ Runner
@@ -192,7 +195,12 @@ if FOUND_TORCH:
                                  hessian_info_service=hessian_info_service)
 
         if core_config.debug_config.analyze_similarity:
-            analyzer_model_quantization(representative_data_gen, tb_w, graph_gptq, fw_impl, DEFAULT_PYTORCH_INFO)
+            analyzer_model_quantization(representative_data_gen,
+                                        tb_w,
+                                        float_graph,
+                                        graph_gptq,
+                                        fw_impl,
+                                        DEFAULT_PYTORCH_INFO)
 
         return get_exportable_pytorch_model(graph_gptq)
 

model_compression_toolkit/ptq/keras/quantization_facade.py

@@ -12,11 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+import copy
 
 from typing import Callable
 
 from model_compression_toolkit.core import CoreConfig
 from model_compression_toolkit.core.analyzer import analyzer_model_quantization
+from model_compression_toolkit.core.common.quantization.quantize_graph_weights import quantize_graph_weights
 from model_compression_toolkit.core.common.visualization.tensorboard_writer import init_tensorboard_writer
 from model_compression_toolkit.logger import Logger
 from model_compression_toolkit.constants import TENSORFLOW, FOUND_TF
@@ -122,8 +124,8 @@ if FOUND_TF:
         if core_config.mixed_precision_enable:
             if not isinstance(core_config.mixed_precision_config, MixedPrecisionQuantizationConfig):
                 Logger.critical("Given quantization config to mixed-precision facade is not of type "
-                                    "MixedPrecisionQuantizationConfig. Please use keras_post_training_quantization "
-                                    "API, or pass a valid mixed precision configuration.")  # pragma: no cover
+                                "MixedPrecisionQuantizationConfig. Please use keras_post_training_quantization "
+                                "API, or pass a valid mixed precision configuration.")  # pragma: no cover
 
         tb_w = init_tensorboard_writer(fw_info)
 
@@ -139,15 +141,30 @@ if FOUND_TF:
                                                target_resource_utilization=target_resource_utilization,
                                                tb_w=tb_w)
 
-        tg = ptq_runner(tg, representative_data_gen, core_config, fw_info, fw_impl, tb_w)
+        # At this point, tg is a graph that went through substitutions (such as BN folding) and is
+        # ready for quantization (namely, it holds quantization params, etc.) but the weights are
+        # not quantized yet. For this reason, we use it to create a graph that acts as a "float" graph
+        # for things like similarity analyzer (because the quantized and float graph should have the same
+        # architecture to find the appropriate compare points for similarity computation).
+        similarity_baseline_graph = copy.deepcopy(tg)
+
+        graph_with_stats_correction = ptq_runner(tg,
+                                                 representative_data_gen,
+                                                 core_config,
+                                                 fw_info,
+                                                 fw_impl,
+                                                 tb_w)
 
         if core_config.debug_config.analyze_similarity:
+            quantized_graph = quantize_graph_weights(graph_with_stats_correction)
             analyzer_model_quantization(representative_data_gen,
-                                        tb_w, tg,
+                                        tb_w,
+                                        similarity_baseline_graph,
+                                        quantized_graph,
                                         fw_impl,
                                         fw_info)
 
-        return get_exportable_keras_model(tg)
+        return get_exportable_keras_model(graph_with_stats_correction)
 
 
 

model_compression_toolkit/ptq/pytorch/quantization_facade.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+import copy
+
 from typing import Callable
 
 from model_compression_toolkit.core import common
@@ -26,6 +28,7 @@ from model_compression_toolkit.core.common.mixed_precision.mixed_precision_quant
 from model_compression_toolkit.core.runner import core_runner
 from model_compression_toolkit.ptq.runner import ptq_runner
 from model_compression_toolkit.core.analyzer import analyzer_model_quantization
+from model_compression_toolkit.core.common.quantization.quantize_graph_weights import quantize_graph_weights
 
 
 if FOUND_TORCH:
@@ -90,14 +93,16 @@ if FOUND_TORCH:
 
         """
 
+        fw_info = DEFAULT_PYTORCH_INFO
+
         if core_config.mixed_precision_enable:
             if not isinstance(core_config.mixed_precision_config, MixedPrecisionQuantizationConfig):
                 Logger.critical("Given quantization config to mixed-precision facade is not of type "
-                             "MixedPrecisionQuantizationConfig. Please use "
-                             "pytorch_post_training_quantization API, or pass a valid mixed precision "
-                             "configuration.")  # pragma: no cover
+                                "MixedPrecisionQuantizationConfig. Please use "
+                                "pytorch_post_training_quantization API, or pass a valid mixed precision "
+                                "configuration.")  # pragma: no cover
 
-        tb_w = init_tensorboard_writer(DEFAULT_PYTORCH_INFO)
+        tb_w = init_tensorboard_writer(fw_info)
 
         fw_impl = PytorchImplementation()
 
@@ -105,22 +110,36 @@ if FOUND_TORCH:
         tg, bit_widths_config, _ = core_runner(in_model=in_module,
                                                representative_data_gen=representative_data_gen,
                                                core_config=core_config,
-                                               fw_info=DEFAULT_PYTORCH_INFO,
+                                               fw_info=fw_info,
                                                fw_impl=fw_impl,
                                                tpc=target_platform_capabilities,
                                                target_resource_utilization=target_resource_utilization,
                                                tb_w=tb_w)
 
-        tg = ptq_runner(tg, representative_data_gen, core_config, DEFAULT_PYTORCH_INFO, fw_impl, tb_w)
+        # At this point, tg is a graph that went through substitutions (such as BN folding) and is
+        # ready for quantization (namely, it holds quantization params, etc.) but the weights are
+        # not quantized yet. For this reason, we use it to create a graph that acts as a "float" graph
+        # for things like similarity analyzer (because the quantized and float graph should have the same
+        # architecture to find the appropriate compare points for similarity computation).
+        similarity_baseline_graph = copy.deepcopy(tg)
+
+        graph_with_stats_correction = ptq_runner(tg,
+                                                 representative_data_gen,
+                                                 core_config,
+                                                 fw_info,
+                                                 fw_impl,
+                                                 tb_w)
 
         if core_config.debug_config.analyze_similarity:
+            quantized_graph = quantize_graph_weights(graph_with_stats_correction)
             analyzer_model_quantization(representative_data_gen,
                                         tb_w,
-                                        tg,
+                                        similarity_baseline_graph,
+                                        quantized_graph,
                                         fw_impl,
-                                        DEFAULT_PYTORCH_INFO)
+                                        fw_info)
 
-        return get_exportable_pytorch_model(tg)
+        return get_exportable_pytorch_model(graph_with_stats_correction)
 
 
 else: