mct-nightly 2.0.0.20240519.436__py3-none-any.whl → 2.0.0.20240521.418__py3-none-any.whl

{mct_nightly-2.0.0.20240519.436.dist-info → mct_nightly-2.0.0.20240521.418.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.0.0.20240519.436
+Version: 2.0.0.20240521.418
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct_nightly-2.0.0.20240519.436.dist-info → mct_nightly-2.0.0.20240521.418.dist-info}/RECORD

@@ -1,4 +1,4 @@
-model_compression_toolkit/__init__.py,sha256=-xNRcIW41rS3RlCyrt3JZJuXSUMGB1BG9UiBEhAtXcE,1573
+model_compression_toolkit/__init__.py,sha256=hotkHO8sD6YxUKPsjLhBTyAQ5cyxVWS4Br7x_RXz5UM,1573
 model_compression_toolkit/constants.py,sha256=b63Jk_bC7VXEX3Qn9TZ3wUvrNKD8Mkz8zIuayoyF5eU,3828
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -31,7 +31,7 @@ model_compression_toolkit/core/common/fusion/__init__.py,sha256=Rf1RcYmelmdZmBV5
 model_compression_toolkit/core/common/fusion/layer_fusing.py,sha256=lOubqpc18TslhXZijWUJQAa1c3jIB2S-M-5HK78wJPQ,5548
 model_compression_toolkit/core/common/graph/__init__.py,sha256=Xr-Lt_qXMdrCnnOaUS_OJP_3iTTGfPCLf8_vSrQgCs0,773
 model_compression_toolkit/core/common/graph/base_graph.py,sha256=lmIw0srKiwCvz7KWqfwKTxyQHDy3s6rWMIXzFAa1UMo,38326
-model_compression_toolkit/core/common/graph/base_node.py,sha256=IaHrvgtUZvPshhUjS4j2lqIP22Mlny82BQZaL2iVJew,29079
+model_compression_toolkit/core/common/graph/base_node.py,sha256=WGkSxjvbRLQBfFT_yrSQRnlmUpwtqkUfpVwrhLgMw5k,29338
 model_compression_toolkit/core/common/graph/edge.py,sha256=buoSEUZwilWBK3WeBKpJ-GeDaUA1SDdOHxDpxU_bGpk,3784
 model_compression_toolkit/core/common/graph/functional_node.py,sha256=71_4TrCdqR_r0mtgxmAyqI05iP5YoQQGeSmDgynuzTw,3902
 model_compression_toolkit/core/common/graph/graph_matchers.py,sha256=CrDoHYq4iPaflgJWmoJ1K4ziLrRogJvFTVWg8P0UcDU,4744
@@ -72,7 +72,7 @@ model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/resource_utilization_data.py,sha256=MhmYgsybWJ2dlCwbZjDKJB8QlDHHy--zOBs2DwJB2e0,13820
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/ru_aggregation_methods.py,sha256=ttc8wPa_9LZansutQ2f1ss-RTzgTv739wy3qsdLzyyk,4217
 model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/ru_functions_mapping.py,sha256=QhuqaECEGLnYC08iD6-2XXcU7NXbPzYf1sQcjYlGak8,1682
-model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/ru_methods.py,sha256=TUGv_4j9-WUWZceoqLdObH_61tAQnHIRyquxm09x9wI,21473
+model_compression_toolkit/core/common/mixed_precision/resource_utilization_tools/ru_methods.py,sha256=WC1EHoNuo_lrzy4NRhGJ1cgmJ2IsFsbmP86mrVO3AVA,21506
 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=7ceagUJVk3GgLrD8HnmXMgZTRLjEvZW_RnzDR3ahNDM,16592
 model_compression_toolkit/core/common/network_editors/__init__.py,sha256=vZmu55bYqiaOQs3AjfwWDXHmuKZcLHt-wm7uR5fPEqg,1307
@@ -126,7 +126,7 @@ model_compression_toolkit/core/common/quantization/quantizers/uniform_quantizers
 model_compression_toolkit/core/common/statistics_correction/__init__.py,sha256=sw7LOPN1bM82o3SkMaklyH0jw-TLGK0-fl2Wq73rffI,697
 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_bias_correction_of_graph.py,sha256=0vGu7FAbuBpusaWxNbBEJHHTCAH_nigV9YeAFGAhghU,10145
+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/substitutions/__init__.py,sha256=sw7LOPN1bM82o3SkMaklyH0jw-TLGK0-fl2Wq73rffI,697
 model_compression_toolkit/core/common/substitutions/apply_substitutions.py,sha256=k-bifmakHIYZeZS-4T1QpZ1Et6AwAijMRgAKs7hmMKc,1390
@@ -136,7 +136,7 @@ model_compression_toolkit/core/common/substitutions/batchnorm_refusing.py,sha256
 model_compression_toolkit/core/common/substitutions/linear_collapsing.py,sha256=iEtzbWCDXP6EDkTZCtREQ0rpMxhQ2kM9zlcP_0KLq9I,12367
 model_compression_toolkit/core/common/substitutions/linear_collapsing_substitution.py,sha256=uoauhmncQqUBNvD-qCLIXsIbl_IzrbxSKdxiMig-5W4,2406
 model_compression_toolkit/core/common/substitutions/remove_identity.py,sha256=TKU1TIU52UIkVnl0EZvWnDhLV9nIVZ4hqi-w1i4NXMk,2637
-model_compression_toolkit/core/common/substitutions/residual_collapsing.py,sha256=doErjlMq-uSObYMSjA6IywSHb3Hz3QCc0HKU68ccrQ4,4767
+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=cyy4qnlD-v1Gou62oHNDsf1hWLWkYfcjVv1otFrUltY,29865
 model_compression_toolkit/core/common/substitutions/softmax_shift.py,sha256=R-0ZqhYAuZLEFWHvB2UTPm52L6gWHGdRdEnwGxKSeGI,2625
@@ -216,13 +216,13 @@ model_compression_toolkit/core/pytorch/pytorch_device_config.py,sha256=S25cuw10A
 model_compression_toolkit/core/pytorch/pytorch_implementation.py,sha256=sEtlxpWdt0rzuTN3R0bNCC_l75Xy7rIBMUWY7LuhYKI,27351
 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=E6ifk1HdO60k4IRH2EFBzAYWtwUlrGqJoQ66nknpHoQ,4983
-model_compression_toolkit/core/pytorch/utils.py,sha256=dRPiteBg2dBNsHwZyYzXiCIAjnelSoeZZsDXlsTw5JQ,2880
+model_compression_toolkit/core/pytorch/utils.py,sha256=OT_mrNEJqPgWLdtQuivKMQVjtJY49cmoIVvbRhANl1w,3004
 model_compression_toolkit/core/pytorch/back2framework/__init__.py,sha256=H_WixgN0elVWf3exgGYsi58imPoYDj5eYPeh6x4yfug,813
 model_compression_toolkit/core/pytorch/back2framework/factory_model_builder.py,sha256=DwNO8WO3JiMawKGKDhlrwCoCjMSBIw5BMbsFFF7eDS4,2279
 model_compression_toolkit/core/pytorch/back2framework/float_model_builder.py,sha256=tLrlUyYhxVKVjkad1ZAtbRra0HedB3iVfIkZ_dYnQ-4,3419
-model_compression_toolkit/core/pytorch/back2framework/instance_builder.py,sha256=d7U_8VKGjQjbmvcUKRKsD80woWu_0gh6LXg1NhZNTjY,1348
+model_compression_toolkit/core/pytorch/back2framework/instance_builder.py,sha256=BBHBfTqeWm7L3iDyPBpk0jxvj-rBg1QWI23imkjfIl0,1467
 model_compression_toolkit/core/pytorch/back2framework/mixed_precision_model_builder.py,sha256=D7lU1r9Uq_7fdNuKk2BMF8ho5GrsY-8gyGN6yYoHaVg,15060
-model_compression_toolkit/core/pytorch/back2framework/pytorch_model_builder.py,sha256=11pUQUxtBM1gCxipEaXn-8A9dVI5rgRiSoFa9i_nw58,18281
+model_compression_toolkit/core/pytorch/back2framework/pytorch_model_builder.py,sha256=5fPlI4BttvQB-gm0iKcWNXZMGjXlcwfXEsxxW0TilTQ,18301
 model_compression_toolkit/core/pytorch/back2framework/quantized_model_builder.py,sha256=qZNNOlNTTV4ZKPG3q5GDXkIVTPUEr8dvxAS_YiMORmg,3456
 model_compression_toolkit/core/pytorch/back2framework/quantization_wrapper/__init__.py,sha256=cco4TmeIDIh32nj9ZZXVkws4dd9F2UDrmjKzTN8G0V0,697
 model_compression_toolkit/core/pytorch/back2framework/quantization_wrapper/quantized_layer_wrapper.py,sha256=q2JDw10NKng50ee2i9faGzWZ-IydnR2aOMGSn9RoZmc,5773
@@ -261,7 +261,7 @@ model_compression_toolkit/core/pytorch/quantizer/__init__.py,sha256=Rf1RcYmelmdZ
 model_compression_toolkit/core/pytorch/quantizer/fake_quant_builder.py,sha256=D8_CEuFqKAhbUgKaRw7Jlxo0zlqgPTMu6CIIIM4LfS0,7045
 model_compression_toolkit/core/pytorch/quantizer/lut_fake_quant.py,sha256=uyeBtNokyDUikk-YkDP_mN_2DX0J5oPm3kSfdSUT2Ck,4420
 model_compression_toolkit/core/pytorch/reader/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
-model_compression_toolkit/core/pytorch/reader/graph_builders.py,sha256=8EieykOysQEMoHZQTLZv9yfkg0dkoFwRtlR3I0rlshg,12626
+model_compression_toolkit/core/pytorch/reader/graph_builders.py,sha256=x5n8KHBqvutqS5l5AillA_FQfhf-2ibP813ixK3Gvy8,12627
 model_compression_toolkit/core/pytorch/reader/node_holders.py,sha256=TaolORuwBZEddWe-q0Mg79Nmswz-Sq3-9-4o8UxFQ50,1028
 model_compression_toolkit/core/pytorch/reader/reader.py,sha256=GEJE0QX8XJFWbYCkbRBtzttZtmmuoACLx8gw9KyAQCE,6015
 model_compression_toolkit/core/pytorch/statistics_correction/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKWF575ofc06JFQSq83Jz99A,696
@@ -483,8 +483,8 @@ 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=MxylaVFPgN7zBiRBy6WV610EA4scLgRJFbMucKvvNDU,2896
-mct_nightly-2.0.0.20240519.436.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.0.0.20240519.436.dist-info/METADATA,sha256=tKiBev23nHrH34I8xfcYLYe523x-prmdwXrf2TQq3J8,18477
-mct_nightly-2.0.0.20240519.436.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-mct_nightly-2.0.0.20240519.436.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.0.0.20240519.436.dist-info/RECORD,,
+mct_nightly-2.0.0.20240521.418.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+mct_nightly-2.0.0.20240521.418.dist-info/METADATA,sha256=J-73dG_pMvZnU4is4fD4RMFUIGnr2pHz2oPayokFBVk,18477
+mct_nightly-2.0.0.20240521.418.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+mct_nightly-2.0.0.20240521.418.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.0.0.20240521.418.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__ = "2.0.0.20240519.000436"
+__version__ = "2.0.0.20240521.000418"

model_compression_toolkit/core/common/graph/base_node.py

@@ -96,6 +96,16 @@ class BaseNode:
         """
         return self.has_activation
 
+    @property
+    def has_positional_weights(self):
+        """
+        Returns has_positional_weights attribute.
+
+        Returns: Whether the node has positional weights.
+
+        """
+        return any(isinstance(key, int) for key in self.weights.keys())
+
     def is_activation_quantization_enabled(self) -> bool:
         """
 

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

@@ -285,7 +285,7 @@ def _bops_utilization(mp_cfg: List[int],
     # Go over all nodes that should be taken into consideration when computing the BOPS utilization.
     bops = []
     for n in graph.get_topo_sorted_nodes():
-        if n.has_kernel_weight_to_quantize(fw_info):
+        if n.has_kernel_weight_to_quantize(fw_info) and not n.has_positional_weights:
             # If node doesn't have weights then its MAC count is 0, and we shouldn't consider it in the BOPS count.
             incoming_edges = graph.incoming_edges(n, sort_by_attr=EDGE_SINK_INDEX)
             if len(incoming_edges) != 1:

model_compression_toolkit/core/common/statistics_correction/compute_bias_correction_of_graph.py

@@ -49,11 +49,16 @@ def compute_bias_correction_of_graph(graph: Graph,
         if fw_info.is_kernel_op(n.type):
             kernel_attr = fw_info.get_kernel_op_attributes(n.type)[0]
             if n.is_weights_quantization_enabled(kernel_attr):
-                _compute_bias_correction_per_candidate_qc(n,
-                                                          kernel_attr,
-                                                          fw_info,
-                                                          graph.get_in_stats_collector(n),
-                                                          fw_impl=fw_impl)
+                # Bias correction is not applied to layers with constant inputs.
+                if n.has_positional_weights:
+                    for candidate_qc in n.candidates_quantization_cfg:
+                        candidate_qc.weights_quantization_cfg.weights_bias_correction = False
+                else:
+                    _compute_bias_correction_per_candidate_qc(n,
+                                                              kernel_attr,
+                                                              fw_info,
+                                                              graph.get_in_stats_collector(n),
+                                                              fw_impl=fw_impl)
     return graph
 
 

model_compression_toolkit/core/common/substitutions/residual_collapsing.py

@@ -67,7 +67,8 @@ class ResidualCollapsing(common.BaseSubstitution):
             return graph
 
         # Check if convolution and residual satisfy the collapsing conditions, otherwise skip substitution
-        if len(graph.get_next_nodes(first_node)) > 1 or len(graph.get_prev_nodes(second_node)) != 2:
+        if (len(graph.get_next_nodes(first_node)) > 1 or len(graph.get_prev_nodes(first_node)) < 1 or
+                len(graph.get_prev_nodes(second_node)) != 2):
             return graph
 
         # Check if Add is residual connection, otherwise skip substitution

model_compression_toolkit/core/pytorch/back2framework/instance_builder.py

@@ -33,7 +33,8 @@ def node_builder(n: BaseNode) -> Module:
 
     framework_attr = copy.copy(n.framework_attr)
     node_instance = n.type(**framework_attr)
-    node_instance.load_state_dict({k: torch.tensor(v) for k, v in n.weights.items()}, strict=False)
+    # Positional weights act as inputs to the PyTorch layer, rather than serving as its weights.
+    node_instance.load_state_dict({k: torch.tensor(v) for k, v in n.weights.items() if isinstance(k, str)}, strict=False)
     set_model(node_instance)
     return node_instance
 

model_compression_toolkit/core/pytorch/back2framework/pytorch_model_builder.py

@@ -71,7 +71,7 @@ def _build_input_tensors_list(node: BaseNode,
             input_tensors = node.insert_positional_weights_to_input_list(input_tensors)
         # convert inputs from positional weights (numpy arrays) to tensors. Must handle each element in the
         # list separately, because in FX the tensors are FX objects and fail to_torch_tensor
-        input_tensors = [to_torch_tensor(t) if isinstance(t, np.ndarray) else t
+        input_tensors = [to_torch_tensor(t, numpy_type=t.dtype) if isinstance(t, np.ndarray) else t
                          for t in input_tensors]
     return input_tensors
 

model_compression_toolkit/core/pytorch/reader/graph_builders.py

@@ -138,6 +138,7 @@ def nodes_builder(model: GraphModule,
                 if input_node in consts_dict:
                     used_consts.add(input_node)
                     weights.update({i: consts_dict[input_node]})
+
                 tensor_meta = input_node.meta
                 if tensor_meta[TYPE] == torch.Tensor:
                     input_shape += [list(tensor_meta[TENSOR_META].shape)]

model_compression_toolkit/core/pytorch/utils.py

@@ -38,11 +38,13 @@ def set_model(model: torch.nn.Module, train_mode: bool = False):
     model.to(device)
 
 
-def to_torch_tensor(tensor):
+def to_torch_tensor(tensor,
+                    numpy_type=np.float32):
     """
     Convert a Numpy array to a Torch tensor.
     Args:
         tensor: Numpy array.
+        numpy_type: The desired data type for the tensor. Default is np.float32.
 
     Returns:
         Torch tensor converted from the input Numpy array.
@@ -55,9 +57,9 @@ def to_torch_tensor(tensor):
     elif isinstance(tensor, tuple):
         return (to_torch_tensor(t) for t in tensor)
     elif isinstance(tensor, np.ndarray):
-        return torch.from_numpy(tensor.astype(np.float32)).to(working_device)
+        return torch.from_numpy(tensor.astype(numpy_type)).to(working_device)
     elif isinstance(tensor, (int, float)):
-        return torch.from_numpy(np.array(tensor).astype(np.float32)).to(working_device)
+        return torch.from_numpy(np.array(tensor).astype(numpy_type)).to(working_device)
     else:
         Logger.critical(f'Unsupported type for conversion to Torch.tensor: {type(tensor)}.')