onnx2tf 1.29.0__tar.gz → 1.29.2__tar.gz

{onnx2tf-1.29.0/onnx2tf.egg-info → onnx2tf-1.29.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnx2tf
-Version: 1.29.0
+Version: 1.29.2
 Summary: Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC).
 Home-page: https://github.com/PINTO0309/onnx2tf
 Author: Katsuya Hyodo
@@ -345,7 +345,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  ghcr.io/pinto0309/onnx2tf:1.29.0
+  ghcr.io/pinto0309/onnx2tf:1.29.2
 
   or
 
@@ -353,7 +353,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  docker.io/pinto0309/onnx2tf:1.29.0
+  docker.io/pinto0309/onnx2tf:1.29.2
 
   or
 

{onnx2tf-1.29.0 → onnx2tf-1.29.2}/README.md

@@ -309,7 +309,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  ghcr.io/pinto0309/onnx2tf:1.29.0
+  ghcr.io/pinto0309/onnx2tf:1.29.2
 
   or
 
@@ -317,7 +317,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  docker.io/pinto0309/onnx2tf:1.29.0
+  docker.io/pinto0309/onnx2tf:1.29.2
 
   or
 

{onnx2tf-1.29.0 → onnx2tf-1.29.2}/onnx2tf/__init__.py

@@ -1,3 +1,3 @@
 from onnx2tf.onnx2tf import convert, main
 
-__version__ = '1.29.0'
+__version__ = '1.29.2'

{onnx2tf-1.29.0 → onnx2tf-1.29.2}/onnx2tf/ops/Conv.py

@@ -370,6 +370,20 @@ def make_node(
             )
 
     def depth_conv_bias(input_tensor, input_weights, pad_mode, strides, dilations, input_bias):
+        # tf.nn.depthwise_conv2d uses a different output shape when dilation>1 and stride>1.
+        # Emulate stride>1 by running stride=1 then subsampling to match ONNX.
+        if pad_mode == 'VALID' \
+            and max(dilations) > 1 \
+            and any(s > 1 for s in strides[1:-1]):
+            conv = tf.nn.depthwise_conv2d(
+                input=input_tensor,
+                filter=input_weights,
+                padding=pad_mode,
+                strides=[1, 1, 1, 1],
+                dilations=dilations,
+            )
+            conv = conv[:, ::strides[1], ::strides[2], :]
+            return tf.add(conv, input_bias)
         return \
             tf.add(
                 tf.nn.depthwise_conv2d(
@@ -438,6 +452,19 @@ def make_node(
             )
 
     def depth_conv_nobias(input_tensor, input_weights, pad_mode, strides, dilations):
+        # tf.nn.depthwise_conv2d uses a different output shape when dilation>1 and stride>1.
+        # Emulate stride>1 by running stride=1 then subsampling to match ONNX.
+        if pad_mode == 'VALID' \
+            and max(dilations) > 1 \
+            and any(s > 1 for s in strides[1:-1]):
+            conv = tf.nn.depthwise_conv2d(
+                input=input_tensor,
+                filter=input_weights,
+                padding=pad_mode,
+                strides=[1, 1, 1, 1],
+                dilations=dilations,
+            )
+            return conv[:, ::strides[1], ::strides[2], :]
         return \
             tf.nn.depthwise_conv2d(
                 input=input_tensor,

{onnx2tf-1.29.0 → onnx2tf-1.29.2}/onnx2tf/ops/SequenceEmpty.py

@@ -30,7 +30,7 @@ def make_node(
     tf_layers_dict: dict
         optype, shape, dtype, tensorflow graph
     """
-    sequence_dtype = ONNX_DTYPES_TO_TF_DTYPES(graph_node.attrs.get('dtype', 1)) # Float32
+    sequence_dtype = ONNX_DTYPES_TO_TF_DTYPES[graph_node.attrs.get('dtype', 1)] # Float32
 
     graph_node_output: gs.Variable = graph_node.outputs[0]
     shape = graph_node_output.shape