mct-nightly 2.3.0.20250427.609__py3-none-any.whl → 2.3.0.20250429.622__py3-none-any.whl

{mct_nightly-2.3.0.20250427.609.dist-info → mct_nightly-2.3.0.20250429.622.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250427.609
+Version: 2.3.0.20250429.622
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License
@@ -51,7 +51,7 @@ ______________________________________________________________________
 </p>
 <p align="center">
   <a href="https://sony.github.io/model_optimization#prerequisites"><img src="https://img.shields.io/badge/pytorch-2.2%20%7C%202.3%20%7C%202.4%20%7C%202.5-blue" /></a>
-  <a href="https://sony.github.io/model_optimization#prerequisites"><img src="https://img.shields.io/badge/tensorflow-2.12%20%7C%202.13%20%7C%202.14%20%7C%202.15-blue" /></a>
+  <a href="https://sony.github.io/model_optimization#prerequisites"><img src="https://img.shields.io/badge/tensorflow-02.14%20%7C%202.15-blue" /></a>
   <a href="https://sony.github.io/model_optimization#prerequisites"><img src="https://img.shields.io/badge/python-3.9%20%7C%203.10%20%7C%203.11%20%7C%203.12-blue" /></a>
   <a href="https://github.com/sony/model_optimization/releases"><img src="https://img.shields.io/github/v/release/sony/model_optimization" /></a>
   <a href="https://github.com/sony/model_optimization/blob/main/LICENSE.md"><img src="https://img.shields.io/badge/license-Apache%202.0-blue" /></a>
@@ -171,7 +171,7 @@ Currently, MCT is being tested on various Python, Pytorch and TensorFlow version
 | Python 3.12 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python312_pytorch22.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python312_pytorch22.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python312_pytorch23.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python312_pytorch23.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python312_pytorch24.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python312_pytorch24.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python312_pytorch25.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python312_pytorch25.yml) |
 
 |             | TensorFlow 2.14                                                                                                                                                                                                        | TensorFlow 2.15                                                                                                                                                                                                        |
-|-------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+|-------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
 | Python 3.9  | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_keras214.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_keras214.yml)   | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_keras215.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python39_keras215.yml)   |
 | Python 3.10 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_keras214.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_keras214.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_keras215.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python310_keras215.yml) |
 | Python 3.11 | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_keras214.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_keras214.yml) | [![Run Tests](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_keras215.yml/badge.svg)](https://github.com/sony/model_optimization/actions/workflows/run_tests_python311_keras215.yml) |

{mct_nightly-2.3.0.20250427.609.dist-info → mct_nightly-2.3.0.20250429.622.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.3.0.20250427.609.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=n5rc5Vyh3AmYQB3AKBl-pAmV5IKGzeqqRm3AOQFlSa4,1557
+mct_nightly-2.3.0.20250429.622.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=PrThtvqTwsbfiKBPdckZFWhbBXaRfAATT21j_WbV8pA,1557
 model_compression_toolkit/constants.py,sha256=iJ6vfTjC2oFIZWt8wvHoxEw5YJi3yl0Hd4q30_8q0Zc,3958
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=L3q7tn3Uht0i_7phnlOWMR2Te2zvzrt2HOz9vYEInts,4529
@@ -47,7 +47,7 @@ model_compression_toolkit/core/common/graph/memory_graph/compute_graph_max_cut.p
 model_compression_toolkit/core/common/graph/memory_graph/cut.py,sha256=ZUGgn-vDA7unzc9UWhK2v_2i5nfdkSG1xOpgpDmziEo,2870
 model_compression_toolkit/core/common/graph/memory_graph/max_cut_astar.py,sha256=1TWLVAOlT8g8q_YyOdjm5cQfiSDZ5EHGQcb509Gnzjg,17895
 model_compression_toolkit/core/common/graph/memory_graph/memory_element.py,sha256=ISD2BvJWj5mB91jrFjG8VQb0oOoLBoita_thCZWzCPI,4238
-model_compression_toolkit/core/common/graph/memory_graph/memory_graph.py,sha256=FCzK4HmX4lWI4qGoGv94wpGv7o6_f5wPBfeBPMerZ18,7752
+model_compression_toolkit/core/common/graph/memory_graph/memory_graph.py,sha256=dME1M0hOjBdW5SqUbl1BPxvFRs-ZtDiF1dDGJuWJbl8,7711
 model_compression_toolkit/core/common/hessian/__init__.py,sha256=E7LK3K_1AwMCQokanNc1JODMwUKNOKmwXQiGQ7GO10I,1033
 model_compression_toolkit/core/common/hessian/hessian_info_service.py,sha256=8NDC_WLe3ZnY_v3e_Vz_lseF22lrbvhFmArihpeWfuI,14291
 model_compression_toolkit/core/common/hessian/hessian_info_utils.py,sha256=1axmN0tjJSo_7hUr2d2KMv4y1pBi19cqWSQpi4BbdsA,1458
@@ -66,16 +66,16 @@ model_compression_toolkit/core/common/mixed_precision/configurable_quant_id.py,s
 model_compression_toolkit/core/common/mixed_precision/configurable_quantizer_utils.py,sha256=7dKMi5S0zQZ16m8NWn1XIuoXsKuZUg64G4-uK8-j1PQ,5177
 model_compression_toolkit/core/common/mixed_precision/distance_weighting.py,sha256=-x8edUyudu1EAEM66AuXPtgayLpzbxoLNubfEbFM5kU,2867
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_candidates_filter.py,sha256=6pLUEEIqRTVIlCYQC4JIvY55KAvuBHEX8uTOQ-1Ac4Q,3859
-model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py,sha256=r1t025_QHshyoop-PZvL7x6UuXaeplCCU3h4VNBhJHo,4309
+model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py,sha256=onHgDwfw8CUbZFNU-RYit9eqA6FrzAtFA3akVZ2d7IM,4533
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_ru_helper.py,sha256=-hOMBucYn12ePyLd0b1KxniPOIRu4b53SwEzv0bWToI,4943
 model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_facade.py,sha256=d5-3j2e_rdcQOT7c4s0p7640i3nSetjJ6MgMhhMM7dc,6152
-model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py,sha256=a0lyySRmQ1vKikx5YvDMA4l1Eha-W5BCPYScvDlL_6c,37300
-model_compression_toolkit/core/common/mixed_precision/sensitivity_evaluation.py,sha256=4bkM8pYKvk18cxHbx973Dz6qWrNT0MRm44cuk__qVaI,27297
+model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py,sha256=J8io_axti6gRoch9QR0FmKOP8JSHGeKqX95rf-nG6fI,37719
+model_compression_toolkit/core/common/mixed_precision/sensitivity_evaluation.py,sha256=R3UIO9lKf-lpEGfJOqgpQAXdP1IWMatWxXKYDkhWj_E,28096
 model_compression_toolkit/core/common/mixed_precision/set_layer_to_bitwidth.py,sha256=P8QtKgFXtt5b2RoubzI5OGlCfbEfZsAirjyrkFzK26A,2846
 model_compression_toolkit/core/common/mixed_precision/solution_refinement_procedure.py,sha256=S1ChgxtUjzXJufNWyRbKoNdyNC6fGUjPeComDMx8ZCo,9479
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization.py,sha256=PKkhc5q8pEPnNLXwo3U56EOCfYnPXIvPs0LlCGZOoKU,4426
-model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py,sha256=D2sNbTPMDsDyUE18NUpVJN27AgdwwhpdOJ8UMLmhdPA,40420
+model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py,sha256=PXBuUUuYDmukjhgyrwEe71egpT_iu-LQt5SqddgkRHo,40793
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py,sha256=QQwtl08DiDxUOQGpYPnek_RlZjWm1Ky7tL2ESHXMK78,4050
 model_compression_toolkit/core/common/mixed_precision/search_methods/__init__.py,sha256=sw7LOPN1bM82o3SkMaklyH0jw-TLGK0-fl2Wq73rffI,697
 model_compression_toolkit/core/common/mixed_precision/search_methods/linear_programming.py,sha256=6Z6nQL9UH7B8dbcUR0cuCTEYFOKZAlvOb-SCk_cAZFA,6670
@@ -528,7 +528,7 @@ model_compression_toolkit/xquant/pytorch/model_analyzer.py,sha256=b93o800yVB3Z-i
 model_compression_toolkit/xquant/pytorch/pytorch_report_utils.py,sha256=UVN_S9ULHBEldBpShCOt8-soT8YTQ5oE362y96qF_FA,3950
 model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
 model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=mkoEktLFFHtEKzzFRn_jCnxjhJolK12TZ5AQeDHzUO8,9767
-mct_nightly-2.3.0.20250427.609.dist-info/METADATA,sha256=uT9gv4F4QKU-7z2fkoBSftfQMo-eUFG_n2ANL1X9DbA,25560
-mct_nightly-2.3.0.20250427.609.dist-info/WHEEL,sha256=SmOxYU7pzNKBqASvQJ7DjX3XGUF92lrGhMb3R6_iiqI,91
-mct_nightly-2.3.0.20250427.609.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250427.609.dist-info/RECORD,,
+mct_nightly-2.3.0.20250429.622.dist-info/METADATA,sha256=R9WTA_IVw4cvLIbQbGpmzTl_ujwHc4RNElUajWHnSNE,25101
+mct_nightly-2.3.0.20250429.622.dist-info/WHEEL,sha256=ck4Vq1_RXyvS4Jt6SI0Vz6fyVs4GWg7AINwpsaGEgPE,91
+mct_nightly-2.3.0.20250429.622.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.3.0.20250429.622.dist-info/RECORD,,

{mct_nightly-2.3.0.20250427.609.dist-info → mct_nightly-2.3.0.20250429.622.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (79.0.1)
+Generator: setuptools (80.0.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

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__ = "2.3.0.20250427.000609"
+__version__ = "2.3.0.20250429.000622"

model_compression_toolkit/core/common/graph/memory_graph/memory_graph.py

@@ -38,9 +38,6 @@ class MemoryGraph(DirectedBipartiteGraph):
         Args:
             model_graph: A graph representation of a model.
         """
-
-        self.model_graph = model_graph
-
         nodes = list(model_graph.nodes)
         memory_tensors = []
         node_to_tensor = []

model_compression_toolkit/core/common/mixed_precision/mixed_precision_quantization_config.py

@@ -27,6 +27,7 @@ class MixedPrecisionQuantizationConfig:
     Args:
         compute_distance_fn (Callable): Function to compute a distance between two tensors. If None, using pre-defined distance methods based on the layer type for each layer.
         distance_weighting_method (MpDistanceWeighting): MpDistanceWeighting enum value that provides a function to use when weighting the distances among different layers when computing the sensitivity metric.
+        custom_metric_fn (Callable): Function to compute a custom metric. As input gets the model_mp and returns a float value for metric. If None, uses interest point metric.
         num_of_images (int): Number of images to use to evaluate the sensitivity of a mixed-precision model comparing to the float model.
         configuration_overwrite (List[int]): A list of integers that enables overwrite of mixed precision with a predefined one.
         num_interest_points_factor (float): A multiplication factor between zero and one (represents percentage) to reduce the number of interest points used to calculate the distance metric.
@@ -39,6 +40,7 @@ class MixedPrecisionQuantizationConfig:
 
     compute_distance_fn: Optional[Callable] = None
     distance_weighting_method: MpDistanceWeighting = MpDistanceWeighting.AVG
+    custom_metric_fn: Optional[Callable] = None
     num_of_images: int = MP_DEFAULT_NUM_SAMPLES
     configuration_overwrite: Optional[List[int]] = None
     num_interest_points_factor: float = field(default=1.0, metadata={"description": "Should be between 0.0 and 1.0"})

model_compression_toolkit/core/common/mixed_precision/mixed_precision_search_manager.py

@@ -78,14 +78,18 @@ class MixedPrecisionSearchManager:
         self.mp_topo_configurable_nodes = self.mp_graph.get_configurable_sorted_nodes(fw_info)
 
         self.ru_targets = target_resource_utilization.get_restricted_targets()
-        self.ru_helper = MixedPrecisionRUHelper(self.mp_graph, fw_info, fw_impl)
+        self.ru_helper = MixedPrecisionRUHelper(self.original_graph, fw_info, fw_impl)
 
         self.min_ru_config: Dict[BaseNode, int] = self.mp_graph.get_min_candidates_config(fw_info)
         self.max_ru_config: Dict[BaseNode, int] = self.mp_graph.get_max_candidates_config(fw_info)
-        self.min_ru = self.ru_helper.compute_utilization(self.ru_targets, self.min_ru_config)
 
         self.config_reconstruction_helper = ConfigReconstructionHelper(virtual_graph=self.mp_graph,
                                                                        original_graph=self.original_graph)
+        if self.using_virtual_graph:
+            real_min_ru_config: Dict[BaseNode, int] = self.config_reconstruction_helper.reconstruct_config_from_virtual_graph(self.min_ru_config)
+            self.min_ru = self.ru_helper.compute_utilization(self.ru_targets, real_min_ru_config)
+        else:
+            self.min_ru = self.ru_helper.compute_utilization(self.ru_targets, self.min_ru_config)
 
     def search(self) -> Dict[BaseNode, int]:
         """
@@ -165,6 +169,7 @@ class MixedPrecisionSearchManager:
             return self.sensitivity_evaluator.compute_metric(topo_cfg(cfg),
                                                              node_idx,
                                                              topo_cfg(baseline_cfg) if baseline_cfg else None)
+
         if self.using_virtual_graph:
             origin_max_config = self.config_reconstruction_helper.reconstruct_config_from_virtual_graph(
                 self.max_ru_config)
@@ -251,7 +256,8 @@ class MixedPrecisionSearchManager:
                 else:
                     cfg = self.min_ru_config.copy()
                     cfg[node] = candidate_idx
-                    candidate_rus = self.ru_helper.compute_utilization(self.ru_targets, cfg)
+                    real_cfg = self.config_reconstruction_helper.reconstruct_config_from_virtual_graph(cfg)
+                    candidate_rus = self.ru_helper.compute_utilization(self.ru_targets, real_cfg)
 
                 for target, ru in candidate_rus.items():
                     rus_per_candidate[target].append(ru)

model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_calculator.py

@@ -18,6 +18,8 @@ from copy import deepcopy
 from enum import Enum, auto
 from typing import Dict, NamedTuple, Optional, Tuple, List, Iterable, Union, Literal, Sequence
 
+from model_compression_toolkit.core.common.fusion.graph_fuser import GraphFuser
+
 from model_compression_toolkit.constants import FLOAT_BITWIDTH
 from model_compression_toolkit.core import FrameworkInfo
 from model_compression_toolkit.core.common import Graph, BaseNode
@@ -145,8 +147,14 @@ class ResourceUtilizationCalculator:
                 raise RuntimeError("Failed to calculate activation memory cuts for graph.")
             cuts = [cut for cut in cuts if cut.mem_elements.elements]
             # cache cuts nodes for future use, so do not filter by target
-            self._cuts = {cut: [self.graph.find_node_by_name(m.node_name)[0] for m in cut.mem_elements.elements]
-                          for cut in cuts}
+            self._cuts = {
+                cut: [
+                    node
+                    for m in cut.mem_elements.elements
+                    for node in (self.graph.fusing_info.get_fused_nodes(m.node_name) or (self.graph.find_node_by_name(m.node_name)[0],))
+                ]
+                for cut in cuts
+            }
         return self._cuts
 
     def compute_resource_utilization(self,
@@ -580,7 +588,9 @@ class ResourceUtilizationCalculator:
 
     def _compute_cuts(self):
         """ Compute activation cuts of the graph. """
-        memory_graph = MemoryGraph(deepcopy(self.graph))
+        # Compute memory graph on fused graph with fused nodes
+        graph = GraphFuser().apply_node_fusion(self.graph)
+        memory_graph = MemoryGraph(deepcopy(graph))
         _, _, cuts = compute_graph_max_cut(memory_graph)
         return cuts
 

model_compression_toolkit/core/common/mixed_precision/sensitivity_evaluation.py

@@ -89,6 +89,9 @@ class SensitivityEvaluation:
         self.interest_points = get_mp_interest_points(graph,
                                                       fw_impl.count_node_for_mixed_precision_interest_points,
                                                       quant_config.num_interest_points_factor)
+        # If using a custom metric - return only model outputs
+        if self.quant_config.custom_metric_fn is not None:
+            self.interest_points = []
 
         # We use normalized MSE when not running hessian-based. For Hessian-based normalized MSE is not needed
         # because hessian weights already do normalization.
@@ -96,6 +99,9 @@ class SensitivityEvaluation:
         self.ips_distance_fns, self.ips_axis = self._init_metric_points_lists(self.interest_points, use_normalized_mse)
 
         self.output_points = get_output_nodes_for_metric(graph)
+        # If using a custom metric - return all model outputs
+        if self.quant_config.custom_metric_fn is not None:
+            self.output_points = [n.node for n in graph.get_outputs()]
         self.out_ps_distance_fns, self.out_ps_axis = self._init_metric_points_lists(self.output_points,
                                                                                     use_normalized_mse)
 
@@ -160,7 +166,7 @@ class SensitivityEvaluation:
         """
         Compute the sensitivity metric of the MP model for a given configuration (the sensitivity
         is computed based on the similarity of the interest points' outputs between the MP model
-        and the float model).
+        and the float model or a custom metric if given).
 
         Args:
             mp_model_configuration: Bitwidth configuration to use to configure the MP model.
@@ -177,15 +183,21 @@ class SensitivityEvaluation:
                                         node_idx)
 
         # Compute the distance metric
-        ipts_distances, out_pts_distances = self._compute_distance()
+        if self.quant_config.custom_metric_fn is None:
+            ipts_distances, out_pts_distances = self._compute_distance()
+            sensitivity_metric = self._compute_mp_distance_measure(ipts_distances, out_pts_distances,
+                                              self.quant_config.distance_weighting_method)
+        else:
+            sensitivity_metric = self.quant_config.custom_metric_fn(self.model_mp)
+            if not isinstance(sensitivity_metric, (float, np.floating)):
+                raise TypeError(f'The custom_metric_fn is expected to return float or numpy float, got {type(sensitivity_metric).__name__}')
 
         # Configure MP model back to the same configuration as the baseline model if baseline provided
         if baseline_mp_configuration is not None:
             self._configure_bitwidths_model(baseline_mp_configuration,
                                             node_idx)
 
-        return self._compute_mp_distance_measure(ipts_distances, out_pts_distances,
-                                                 self.quant_config.distance_weighting_method)
+        return sensitivity_metric
 
     def _init_baseline_tensors_list(self):
         """