mct-nightly 2.3.0.20250423.537__py3-none-any.whl → 2.3.0.20250425.557__py3-none-any.whl

{mct_nightly-2.3.0.20250423.537.dist-info → mct_nightly-2.3.0.20250425.557.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mct-nightly
-Version: 2.3.0.20250423.537
+Version: 2.3.0.20250425.557
 Summary: A Model Compression Toolkit for neural networks
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: Apache Software License
@@ -34,7 +34,7 @@ Dynamic: summary
 <div align="center" markdown="1">
 <p>
       <a href="https://sony.github.io/model_optimization/" target="_blank">
-        <img src="https://github.com/sony/model_optimization/blob/main/docsrc/images/mctHeader1-cropped.svg" width="1000"></a>
+        <img src="https://raw.githubusercontent.com/sony/model_optimization/refs/heads/main/docsrc/images/mctHeader1-cropped.svg" width="1000"></a>
 </p>
   
 ______________________________________________________________________
@@ -100,7 +100,7 @@ For further details, please see [Supported features and algorithms](#high-level-
 <div align="center">
 <p align="center">
 
-<img src="https://github.com/sony/model_optimization/blob/main/docsrc/images/mctDiagram_clean.svg" width="800">
+<img src="https://raw.githubusercontent.com/sony/model_optimization/refs/heads/main/docsrc/images/mctDiagram_clean.svg" width="800">
 </p>    
 </div>
 
@@ -181,16 +181,16 @@ Currently, MCT is being tested on various Python, Pytorch and TensorFlow version
 ## <div align="center">Results</div>
 
 <p align="center">
-<img src="https://github.com/sony/model_optimization/blob/main/docsrc/images/Classification.png" width="200">
-<img src="https://github.com/sony/model_optimization/blob/main/docsrc/images/SemSeg.png" width="200">
-<img src="https://github.com/sony/model_optimization/blob/main/docsrc/images/PoseEst.png" width="200">
-<img src="https://github.com/sony/model_optimization/blob/main/docsrc/images/ObjDet.png" width="200">
+<img src="https://raw.githubusercontent.com/sony/model_optimization/refs/heads/main/docsrc/images/Classification.png" width="200">
+<img src="https://raw.githubusercontent.com/sony/model_optimization/refs/heads/main/docsrc/images/SemSeg.png" width="200">
+<img src="https://raw.githubusercontent.com/sony/model_optimization/refs/heads/main/docsrc/images/PoseEst.png" width="200">
+<img src="https://raw.githubusercontent.com/sony/model_optimization/refs/heads/main/docsrc/images/ObjDet.png" width="200">
 
 MCT can quantize an existing 32-bit floating-point model to an 8-bit fixed-point (or less) model without compromising accuracy. 
 Below is a graph of [MobileNetV2](https://pytorch.org/vision/main/models/generated/torchvision.models.mobilenet_v2.html) accuracy on ImageNet vs average bit-width of weights (X-axis), using **single-precision** quantization, **mixed-precision** quantization, and mixed-precision quantization with GPTQ. 
 
 <p align="center">
-<img src="https://github.com/sony/model_optimization/blob/main/docsrc/images/torch_mobilenetv2.png" width="800">
+<img src="https://raw.githubusercontent.com/sony/model_optimization/refs/heads/main/docsrc/images/torch_mobilenetv2.png" width="800">
 
 For more results, please see [1]
 

{mct_nightly-2.3.0.20250423.537.dist-info → mct_nightly-2.3.0.20250425.557.dist-info}/RECORD

@@ -1,5 +1,5 @@
-mct_nightly-2.3.0.20250423.537.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-model_compression_toolkit/__init__.py,sha256=NXqhvuAEHQlzpdJpTtyy1rDJW2gxmMwEGZvHIBQE_f0,1557
+mct_nightly-2.3.0.20250425.557.dist-info/licenses/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+model_compression_toolkit/__init__.py,sha256=kXjmCqsExE8DYRx8JTVs4AhVXF1GF94SPL9EcfyLeks,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
@@ -101,7 +101,7 @@ model_compression_toolkit/core/common/pruning/mask/__init__.py,sha256=huHoBUcKNB
 model_compression_toolkit/core/common/pruning/mask/per_channel_mask.py,sha256=77DB1vqq_gHwbUjeCHRaq1Q-V4wEtdVdwkGezcZgToA,5021
 model_compression_toolkit/core/common/pruning/mask/per_simd_group_mask.py,sha256=_LcDAxLeC5I0KdMHS8jib5XxIKO2ZLavXYuSMIPIQBo,5868
 model_compression_toolkit/core/common/quantization/__init__.py,sha256=sw7LOPN1bM82o3SkMaklyH0jw-TLGK0-fl2Wq73rffI,697
-model_compression_toolkit/core/common/quantization/bit_width_config.py,sha256=0HA3CIZW-ZrA55ra-yJXRvAYnoR8i1SjpbnMDKcWYNQ,12819
+model_compression_toolkit/core/common/quantization/bit_width_config.py,sha256=034kgwe0ydyLXsV83KqxKyyHkoUQH06ai0leLyg0p8I,13019
 model_compression_toolkit/core/common/quantization/candidate_node_quantization_config.py,sha256=lyWPvnoX8BmulhLKR20r5gT2_Yan7P40d8EcgDhErPk,4905
 model_compression_toolkit/core/common/quantization/core_config.py,sha256=yxCzWqldcHoe8GGxrH0tp99bhrc5jDT7SgZftnMUUBE,2374
 model_compression_toolkit/core/common/quantization/debug_config.py,sha256=uH45Uq3Tp9FIyMynex_WY2_y-Kv8LuPw2XXZydnpW5A,1649
@@ -112,7 +112,7 @@ model_compression_toolkit/core/common/quantization/quantization_fn_selection.py,
 model_compression_toolkit/core/common/quantization/quantization_params_fn_selection.py,sha256=7eG7dl1TcbdnHwgmvyjarxLs0o6Lw_9VAjXAm4rsiBk,3791
 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=_hhRb5eeFwbtPddu2xdLi7qK1RsxoR7UHUfjO0ICM3Q,30586
+model_compression_toolkit/core/common/quantization/set_node_quantization_config.py,sha256=76BjqZbeRz7zIblKxGp-AUkHOt-KZMlYjsXP8Ph-1jc,30943
 model_compression_toolkit/core/common/quantization/quantization_params_generation/__init__.py,sha256=eCDGwsWYLU6z7qbEVb4TozMW_nd5VEP_iCJ6PcvyEPw,1486
 model_compression_toolkit/core/common/quantization/quantization_params_generation/error_functions.py,sha256=_m-XkEMJMHf0gYwVIXAoHVjdRa2NXt_gYdwBlw76ZR8,24031
 model_compression_toolkit/core/common/quantization/quantization_params_generation/lut_kmeans_params.py,sha256=RL-PklAjGyC-26anSt8fU07a6pB_LBQFQy9o4e9giN0,8739
@@ -145,7 +145,7 @@ model_compression_toolkit/core/common/substitutions/linear_collapsing_substituti
 model_compression_toolkit/core/common/substitutions/remove_identity.py,sha256=TKU1TIU52UIkVnl0EZvWnDhLV9nIVZ4hqi-w1i4NXMk,2637
 model_compression_toolkit/core/common/substitutions/residual_collapsing.py,sha256=N82mso5j3EJQlKt9EMHjjEJ67FmdGQeCfN8U5grOFXo,4830
 model_compression_toolkit/core/common/substitutions/scale_equalization.py,sha256=p57u25qdW2pimxzGwgMXEBV4S-LzXuTVAlIM7830WfU,10966
-model_compression_toolkit/core/common/substitutions/shift_negative_activation.py,sha256=1utreR5CkJYwaJS1LOCQi_EYkedsnxXzyJCnJ-ZeqQ0,30057
+model_compression_toolkit/core/common/substitutions/shift_negative_activation.py,sha256=zCkdyZHEkbxkORmd071_XWajkpIhnDq9D6FyeE4TQjc,30057
 model_compression_toolkit/core/common/substitutions/softmax_shift.py,sha256=R-0ZqhYAuZLEFWHvB2UTPm52L6gWHGdRdEnwGxKSeGI,2625
 model_compression_toolkit/core/common/substitutions/virtual_activation_weights_composition.py,sha256=w43dRmaG96a8SNECgghxoFCTSoZ-vUb33dXGm2PbomE,4251
 model_compression_toolkit/core/common/substitutions/weights_activation_split.py,sha256=gt07lXRUvYunJKiwv_w20zfXhcplSW4oT2C1dqiNNXc,4719
@@ -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-2.3.0.20250423.537.dist-info/METADATA,sha256=PeCeasbP-z1tGgn1g1tpCnmFWMlSQ7HtIfby4JXgX68,25413
-mct_nightly-2.3.0.20250423.537.dist-info/WHEEL,sha256=pxyMxgL8-pra_rKaQ4drOZAegBVuX-G_4nRHjjgWbmo,91
-mct_nightly-2.3.0.20250423.537.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.3.0.20250423.537.dist-info/RECORD,,
+mct_nightly-2.3.0.20250425.557.dist-info/METADATA,sha256=3W1evd1H2JL0wpqb1frzmmX5nz2fqHcMib1y3remRt8,25560
+mct_nightly-2.3.0.20250425.557.dist-info/WHEEL,sha256=SmOxYU7pzNKBqASvQJ7DjX3XGUF92lrGhMb3R6_iiqI,91
+mct_nightly-2.3.0.20250425.557.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.3.0.20250425.557.dist-info/RECORD,,

{mct_nightly-2.3.0.20250423.537.dist-info → mct_nightly-2.3.0.20250425.557.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (79.0.0)
+Generator: setuptools (79.0.1)
 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__ = "2.3.0.20250423.000537"
+__version__ = "2.3.0.20250425.000557"

model_compression_toolkit/core/common/quantization/bit_width_config.py

@@ -20,6 +20,8 @@ from model_compression_toolkit.core.common.matchers.node_matcher import BaseNode
 from model_compression_toolkit.logger import Logger
 
 from model_compression_toolkit.core.common.graph.base_node import WeightAttrT
+from model_compression_toolkit.target_platform_capabilities.constants import POS_ATTR
+
 
 @dataclass
 class ManualBitWidthSelection:
@@ -221,9 +223,10 @@ class BitWidthConfig:
                     if isinstance(attr_str, str) and isinstance(manual_bit_width_selection.attr, str):
                         if attr_str.find(manual_bit_width_selection.attr) != -1:
                             attr.append(attr_str)
-                    elif isinstance(attr_str, int) and isinstance(manual_bit_width_selection.attr, int):
-                        if attr_str == manual_bit_width_selection.attr:
-                            attr.append(attr_str)
+                    # this is a positional attribute, so it needs to be handled separately.
+                    # Search manual_bit_width_selection's attribute that contain the POS_ATTR string.
+                    elif isinstance(attr_str, int) and POS_ATTR in manual_bit_width_selection.attr:
+                        attr.append(POS_ATTR)
                 if len(attr) == 0:
                     Logger.critical(f'The requested attribute {manual_bit_width_selection.attr} to change the bit width for {n} does not exist.')
 

model_compression_toolkit/core/common/quantization/set_node_quantization_config.py

@@ -119,11 +119,16 @@ def filter_node_qco_by_graph(node: BaseNode,
             _next_nodes.extend(graph.get_next_nodes(n))
         next_nodes.append(n)
 
-    if len(next_nodes):
-        next_nodes_qc_options = [_node.get_qco(fqc) for _node in next_nodes]
-        next_nodes_supported_input_bitwidth = min([max_input_activation_n_bits(op_cfg)
+    if len(next_nodes) == 0:
+        return _base_config, _node_qc_options
+    next_nodes_qc_options = [_node.get_qco(fqc) for _node in next_nodes]
+    all_next_nodes_supported_input_bitwidth = [max_input_activation_n_bits(op_cfg)
                                                    for qc_opts in next_nodes_qc_options
-                                                   for op_cfg in qc_opts.quantization_configurations])
+                                                   for op_cfg in qc_opts.quantization_configurations
+                                               if op_cfg.enable_activation_quantization or op_cfg.quantization_preserving
+                                               ]
+    if len(all_next_nodes_supported_input_bitwidth):
+        next_nodes_supported_input_bitwidth = min(all_next_nodes_supported_input_bitwidth)
 
         # Filter node's QC options that match next nodes input bit-width.
         _node_qc_options = [_option for _option in _node_qc_options
@@ -205,7 +210,7 @@ def set_quantization_configs_to_node(node: BaseNode,
                 # Preserving the quantization of more than 1 previous node is ambiguous, so disable it.
                 Logger.info(f"Disabling Quantization-Preserving for node {node.name} because it has more than 1 input activations.")
                 candidate_qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT
-            elif not prev_nodes[0].is_quantization_preserving() or not prev_nodes[0].is_activation_quantization_enabled():
+            elif not prev_nodes[0].is_quantization_preserving() and not prev_nodes[0].is_activation_quantization_enabled():
                 # Preserving the quantization of an unquantized node isn't possible, so disable it.
                 Logger.info(f"Disabling Quantization-Preserving for node {node.name} because previous node activation quantization is disabled.")
                 candidate_qc.activation_quantization_cfg.quant_mode = ActivationQuantizationMode.NO_QUANT

model_compression_toolkit/core/common/substitutions/shift_negative_activation.py

@@ -343,6 +343,13 @@ def shift_negative_function(graph: Graph,
     graph.set_out_stats_collector_to_node(add_node, add_node_stats_collector)
     graph.shift_stats_collector(add_node, np.array(shift_value))
 
+    set_quantization_configs_to_node(fw_info=fw_info,
+                                     node=add_node,
+                                     graph=graph,
+                                     quant_config=core_config.quantization_config,
+                                     fqc=graph.fqc,
+                                     mixed_precision_enable=core_config.is_mixed_precision_enabled)
+
     if padding is not None:
         pad_node = create_pad_node(op2d_node.name,
                                    add_node.name,
@@ -373,13 +380,6 @@ def shift_negative_function(graph: Graph,
 
         op2d_node.input_shape = pad_node.output_shape
 
-    set_quantization_configs_to_node(fw_info=fw_info,
-                                     node=add_node,
-                                     graph=graph,
-                                     quant_config=core_config.quantization_config,
-                                     fqc=graph.fqc,
-                                     mixed_precision_enable=core_config.is_mixed_precision_enabled)
-
     original_non_linear_activation_nbits = non_linear_node_cfg_candidate.activation_n_bits
     # The non-linear node's output should be float, so we approximate it by using 16bits quantization.
     for candidate_qc in non_linear_node.candidates_quantization_cfg: