mct-nightly 2.2.0.20241123.518__py3-none-any.whl → 2.2.0.20241125.521__py3-none-any.whl

{mct_nightly-2.2.0.20241123.518.dist-info → mct_nightly-2.2.0.20241125.521.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.2.0.20241123.518
+Version: 2.2.0.20241125.521
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct_nightly-2.2.0.20241123.518.dist-info → mct_nightly-2.2.0.20241125.521.dist-info}/RECORD

@@ -1,4 +1,4 @@
-model_compression_toolkit/__init__.py,sha256=-xH9WmXx3mA1ga3UmoUb7EqBNN0auFU_wQv6dSA5WVw,1573
+model_compression_toolkit/__init__.py,sha256=V7JxE-NIGj2DUXx77l3Ho4NbCRrjcf7p_ma6-PrKRlM,1573
 model_compression_toolkit/constants.py,sha256=i4wYheBkIdQmsQA-axIpcT3YiSO1USNc-jaNiNE8w6E,3920
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -176,7 +176,7 @@ model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm_reconstruction.py,sha256=GR1a3mCZpNUu4WxixJXF_aSm57phAdxaRoHecNx3hxw,3168
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/batchnorm_refusing.py,sha256=5df_xGfXkqNub4xVRnCWQvSohWqdv12axjJ6edVU2H0,2478
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/concat_threshold_update.py,sha256=Hl4LEQ_bw_Vpmf3ZqHujYUqVdvTNsPlEMvr9dZhwg2U,2806
-model_compression_toolkit/core/keras/graph_substitutions/substitutions/conv_funcs_to_layer.py,sha256=sf4nrcQckDlvfA1CNXCXcp35vHO-TVPrc9tXK2vDV6I,11198
+model_compression_toolkit/core/keras/graph_substitutions/substitutions/conv_funcs_to_layer.py,sha256=RwzqSksGNmN1KPH8RTJzpCSjGgxvtT9kqqPqsjbGPqs,11631
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/dwconv_to_conv.py,sha256=R3U7cjc2E0zheMem16GHygp5jZFGSaomkNOTxTjcAgw,5794
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/input_scaling.py,sha256=V6hp67CkS_A3WqdsjLjs0ETtdZAOo4P9mhy4aT7W5FE,5940
 model_compression_toolkit/core/keras/graph_substitutions/substitutions/linear_collapsing.py,sha256=AvquvVVVT8-ioeVn-gjqysK4L41L3I7TlNOEDfWjViY,8185
@@ -559,8 +559,8 @@ model_compression_toolkit/xquant/pytorch/model_analyzer.py,sha256=b93o800yVB3Z-i
 model_compression_toolkit/xquant/pytorch/pytorch_report_utils.py,sha256=bOc-hFL3gdoSM1Th_S2N_-9JJSlPGpZCTx_QLJHS6lg,3388
 model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
 model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=mkoEktLFFHtEKzzFRn_jCnxjhJolK12TZ5AQeDHzUO8,9767
-mct_nightly-2.2.0.20241123.518.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.2.0.20241123.518.dist-info/METADATA,sha256=Cu9zX56iaIeLO2feGNpwHp1m3r6sbwKCmUF-ZY5VPUg,26473
-mct_nightly-2.2.0.20241123.518.dist-info/WHEEL,sha256=bFJAMchF8aTQGUgMZzHJyDDMPTO3ToJ7x23SLJa1SVo,92
-mct_nightly-2.2.0.20241123.518.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.2.0.20241123.518.dist-info/RECORD,,
+mct_nightly-2.2.0.20241125.521.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+mct_nightly-2.2.0.20241125.521.dist-info/METADATA,sha256=P2SQ21pjaU6q--__nRskZCmLp3jjCliPQPLOBCm7VnM,26473
+mct_nightly-2.2.0.20241125.521.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+mct_nightly-2.2.0.20241125.521.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.2.0.20241125.521.dist-info/RECORD,,

{mct_nightly-2.2.0.20241123.518.dist-info → mct_nightly-2.2.0.20241125.521.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: bdist_wheel (0.45.0)
+Generator: bdist_wheel (0.45.1)
 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__ = "2.2.0.20241123.000518"
+__version__ = "2.2.0.20241125.000521"

model_compression_toolkit/core/keras/graph_substitutions/substitutions/conv_funcs_to_layer.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+from typing import Optional, Tuple
 
 import numpy as np
 import tensorflow as tf
@@ -30,7 +31,7 @@ from model_compression_toolkit.core.common.graph.graph_matchers import NodeOpera
 from model_compression_toolkit.constants import REUSE, REUSE_GROUP
 from model_compression_toolkit.core.keras.constants import KERNEL, BIAS, USE_BIAS, FILTERS, PADDING, \
     KERNEL_SIZE, DEPTH_MULTIPLIER, STRIDES, DILATIONS, DILATION_RATE, DEPTHWISE_KERNEL, RATE, \
-    ACTIVATION, LINEAR
+    ACTIVATION, LINEAR, DATA_FORMAT, GROUPS, CHANNELS_FORMAT_FIRST, CHANNELS_FORMAT_LAST
 
 
 def extract_bias_node_data(_node: FunctionalNode, _graph: Graph) -> np.ndarray:
@@ -136,35 +137,35 @@ class Conv2dFuncToConv2dLayer(common.BaseSubstitution):
         conv_fw_attr = {FILTERS: k.shape[3], KERNEL_SIZE: k.shape[:2], ACTIVATION: LINEAR}
         if len(conv_func_node.op_call_args) > 0:
             Logger.critical(f"node {conv_func_node.name} expected to have only kwargs but got args={conv_func_node.op_call_args}.")  # pragma: no cover
-        if STRIDES in conv_func_node.op_call_kwargs:
-            strides = conv_func_node.op_call_kwargs[STRIDES]
-            if len(strides) == 4:
-                if strides[0] > 1 or strides[3] > 1:
-                    # Non-standard strides -> skip substitution.
-                    return graph  # pragma: no cover
-                conv_fw_attr[STRIDES] = strides[1:3]
-            else:
-                conv_fw_attr[STRIDES] = strides
-        if PADDING in conv_func_node.op_call_kwargs:
-            padding = conv_func_node.op_call_kwargs[PADDING]
-            if not isinstance(padding, str):
-                # Non-standard padding, Layer only support either 'valid' or 'same' -> skip substitution.
-                return graph  # pragma: no cover
-            conv_fw_attr[PADDING] = padding
-        if DILATIONS in conv_func_node.op_call_kwargs and conv_func_node.op_call_kwargs[DILATIONS] is not None:
-            dilations = conv_func_node.op_call_kwargs[DILATIONS]
-            if isinstance(dilations, (list, tuple)) and len(dilations) == 4:
-                if dilations[0] > 1 or dilations[3] > 1:
-                    # Non-standard dilations -> skip substitution.
-                    return graph  # pragma: no cover
-                conv_fw_attr[DILATION_RATE] = dilations[1:3]
-            else:
-                conv_fw_attr[DILATION_RATE] = dilations
+
+        strides = self._parse_tf_stride_dilation(conv_func_node, STRIDES)
+        if strides is None:
+            # Non-standard strides -> skip substitution.
+            return graph
+        conv_fw_attr[STRIDES] = strides
+
+        padding = conv_func_node.op_call_kwargs.get(PADDING) or 'VALID'
+        if not isinstance(padding, str):
+            # Non-standard padding, Layer only support either 'valid' or 'same' -> skip substitution.
+            return graph  # pragma: no cover
+        conv_fw_attr[PADDING] = padding
+
+        dilations = self._parse_tf_stride_dilation(conv_func_node, DILATIONS)
+        if dilations is None:
+            # Non-standard dilations -> skip substitution.
+            return graph
+        conv_fw_attr[DILATION_RATE] = dilations
+
         if b is None:
             conv_fw_attr[USE_BIAS] = False
         else:
             weights[BIAS] = b
 
+        data_format = conv_func_node.op_call_kwargs.get(DATA_FORMAT, 'NHWC')
+        conv_fw_attr[DATA_FORMAT] = {'NHWC': CHANNELS_FORMAT_LAST, 'NCHW': CHANNELS_FORMAT_FIRST}[data_format]
+
+        conv_fw_attr[GROUPS] = 1
+
         _reuse_params = {REUSE: conv_func_node.reuse, REUSE_GROUP: conv_func_node.reuse_group}
         conv_node = BaseNode(conv_func_node.name, conv_fw_attr, conv_func_node.input_shape, conv_func_node.output_shape,
                              weights, Conv2D, **_reuse_params)
@@ -172,6 +173,31 @@ class Conv2dFuncToConv2dLayer(common.BaseSubstitution):
         replace_conv_node(graph, conv_node, conv_func_node, remove_add_node=b is not None)
         return graph
 
+    def _parse_tf_stride_dilation(self, node, key) -> Optional[Tuple[int, int]]:
+        """
+        Extract stride/dilation param from tf node and convert it to keras format (suitable for Conv2D).
+
+        Args:
+            node: node
+            key: param key
+
+        Returns:
+            Parsed value or None if non-standard.
+        """
+        v = node.op_call_kwargs.get(key)
+        if v is None:
+            return 1, 1
+        if isinstance(v, int):
+            return v, v
+        if len(v) == 1:
+            return v[0], v[0]
+        if len(v) == 4:
+            if v[0] > 1 and v[-1] > 1:
+                return None
+            else:
+                return v[1:3]
+        return tuple(v)
+
 
 class DwConv2dFuncToDwConv2dLayer(common.BaseSubstitution):
     """