onnx2tf 1.27.0__tar.gz → 1.27.2__tar.gz

{onnx2tf-1.27.0/onnx2tf.egg-info → onnx2tf-1.27.2}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: onnx2tf
-Version: 1.27.0
+Version: 1.27.2
 Summary: Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC). The purpose of this tool is to solve the massive Transpose extrapolation problem in onnx-tensorflow (onnx-tf).
 Home-page: https://github.com/PINTO0309/onnx2tf
 Author: Katsuya Hyodo
@@ -18,6 +18,7 @@ Dynamic: description
 Dynamic: description-content-type
 Dynamic: home-page
 Dynamic: license
+Dynamic: license-file
 Dynamic: platform
 Dynamic: requires-python
 Dynamic: summary
@@ -280,7 +281,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
 
 ## Environment
 - Linux / Windows
-- onnx==1.16.1
+- onnx==1.17.0
 - onnxruntime==1.18.1
 - onnx-simplifier==0.4.33 or 0.4.30 `(onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:Slice, node name: /xxxx/Slice): [ShapeInferenceError] Inferred shape and existing shape differ in rank: (x) vs (y))`
 - onnx_graphsurgeon
@@ -333,7 +334,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  ghcr.io/pinto0309/onnx2tf:1.27.0
+  ghcr.io/pinto0309/onnx2tf:1.27.2
 
   or
 
@@ -341,7 +342,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  docker.io/pinto0309/onnx2tf:1.27.0
+  docker.io/pinto0309/onnx2tf:1.27.2
 
   or
 
@@ -1512,7 +1513,7 @@ See: https://github.com/tensorflow/tfjs/tree/master/tfjs-converter
 When converting to CoreML, process as follows. The `-k` option is for conversion while maintaining the input channel order in ONNX's NCHW format.
 
 ```bash
-pip install coremltools
+pip install coremltools==8.2
 
 onnx2tf -i mobilenetv2-12.onnx -k input -ois input:1,3,224,224 -osd
 ```
@@ -1525,7 +1526,7 @@ model = ct.convert(
     model=FOLDER_PATH,
     source='tensorflow',
 )
-model.save(f'{FOLDER_PATH}/model.mlmodel')
+model.save(f'{FOLDER_PATH}/model.mlpackage')
 ```
 
 See: https://github.com/apple/coremltools

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/README.md

@@ -256,7 +256,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
 
 ## Environment
 - Linux / Windows
-- onnx==1.16.1
+- onnx==1.17.0
 - onnxruntime==1.18.1
 - onnx-simplifier==0.4.33 or 0.4.30 `(onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:Slice, node name: /xxxx/Slice): [ShapeInferenceError] Inferred shape and existing shape differ in rank: (x) vs (y))`
 - onnx_graphsurgeon
@@ -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:1.27.0
+  ghcr.io/pinto0309/onnx2tf:1.27.2
 
   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:1.27.0
+  docker.io/pinto0309/onnx2tf:1.27.2
 
   or
 
@@ -1488,7 +1488,7 @@ See: https://github.com/tensorflow/tfjs/tree/master/tfjs-converter
 When converting to CoreML, process as follows. The `-k` option is for conversion while maintaining the input channel order in ONNX's NCHW format.
 
 ```bash
-pip install coremltools
+pip install coremltools==8.2
 
 onnx2tf -i mobilenetv2-12.onnx -k input -ois input:1,3,224,224 -osd
 ```
@@ -1501,7 +1501,7 @@ model = ct.convert(
     model=FOLDER_PATH,
     source='tensorflow',
 )
-model.save(f'{FOLDER_PATH}/model.mlmodel')
+model.save(f'{FOLDER_PATH}/model.mlpackage')
 ```
 
 See: https://github.com/apple/coremltools

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/__init__.py

@@ -1,3 +1,3 @@
 from onnx2tf.onnx2tf import convert, main
 
-__version__ = '1.27.0'
+__version__ = '1.27.2'

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/ops/Add.py

@@ -130,9 +130,9 @@ def make_node(
     )
 
     # Workaround for ConvInteger
-    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64]:
+    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64, tf.float16]:
         input_tensor_2 = tf.cast(input_tensor_2, dtype=tf.float32)
-    elif input_tensor_1.dtype in [tf.int32, tf.int64] and input_tensor_2.dtype == tf.float32:
+    elif input_tensor_1.dtype in [tf.int32, tf.int64, tf.float16] and input_tensor_2.dtype == tf.float32:
         input_tensor_1 = tf.cast(input_tensor_1, dtype=tf.float32)
 
     # Disable unnecessary Transpose

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/ops/Cast.py

@@ -96,10 +96,16 @@ def make_node(
         tf_layers_dict[graph_node_output.name].pop('nhwc')
 
     # Suppression of FlexCast generation
-    # Float32 -> Float64
+    # Float64 -> Float32
+    # Float16 -> Float32
     if input_tensor.dtype == tf.float32 \
         and to == tf.float64:
         to = tf.float32
+    elif isinstance(graph_node.inputs[0], gs.Variable) \
+        and hasattr(graph_node.inputs[0], "dtype") \
+        and graph_node.inputs[0].dtype == np.float32 \
+        and to == tf.float16:
+        to = tf.float32
 
     # Generation of TF OP
     tf_layers_dict[graph_node_output.name]['tf_node'] = \

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/ops/Div.py

@@ -116,9 +116,9 @@ def make_node(
     )
 
     # Workaround for ConvInteger
-    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64]:
+    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64, tf.float16]:
         input_tensor_2 = tf.cast(input_tensor_2, dtype=tf.float32)
-    elif input_tensor_1.dtype in [tf.int32, tf.int64] and input_tensor_2.dtype == tf.float32:
+    elif input_tensor_1.dtype in [tf.int32, tf.int64, tf.float16] and input_tensor_2.dtype == tf.float32:
         input_tensor_1 = tf.cast(input_tensor_1, dtype=tf.float32)
 
     # Disable unnecessary Transpose

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/ops/MatMul.py

@@ -128,6 +128,12 @@ def make_node(
     output_dtype = NUMPY_DTYPES_TO_TF_DTYPES[dtype] \
         if isinstance(dtype, np.dtype) else dtype
 
+    # Workaround for Float16
+    if input_tensor_1.dtype == tf.float32 and output_dtype in [tf.int32, tf.int64, tf.float16]:
+        output_dtype = tf.float32
+    elif output_dtype and input_tensor_2.dtype == tf.float32:
+        output_dtype = tf.float32
+
     # Shape Unmatch Error Mitigation Measures
     # Search for and transpose shapes that do not cause shape unmatch errors
     min_abs_err = sys.maxsize

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/ops/Mod.py

@@ -115,9 +115,9 @@ def make_node(
     )
 
     # Workaround for ConvInteger
-    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64]:
+    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64, tf.float16]:
         input_tensor_2 = tf.cast(input_tensor_2, dtype=tf.float32)
-    elif input_tensor_1.dtype in [tf.int32, tf.int64] and input_tensor_2.dtype == tf.float32:
+    elif input_tensor_1.dtype in [tf.int32, tf.int64, tf.float16] and input_tensor_2.dtype == tf.float32:
         input_tensor_1 = tf.cast(input_tensor_1, dtype=tf.float32)
 
     # Disable unnecessary Transpose

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/ops/Mul.py

@@ -120,9 +120,9 @@ def make_node(
     )
 
     # Workaround for ConvInteger
-    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64]:
+    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64, tf.float16]:
         input_tensor_2 = tf.cast(input_tensor_2, dtype=tf.float32)
-    elif input_tensor_1.dtype in [tf.int32, tf.int64] and input_tensor_2.dtype == tf.float32:
+    elif input_tensor_1.dtype in [tf.int32, tf.int64, tf.float16] and input_tensor_2.dtype == tf.float32:
         input_tensor_1 = tf.cast(input_tensor_1, dtype=tf.float32)
 
     # Disable unnecessary Transpose

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/ops/Split.py

@@ -158,15 +158,56 @@ def make_node(
                 if idx == axis:
                     end_.append(split_idx + 1)
                 elif input_tensor_shape[idx] is None:
-                    end_.append(-1)
+                    end_.append(0)
                 else:
                     end_.append(input_tensor_shape[idx])
 
+            begin_mask_ = np.sum([2**idx if idx != axis else 0 for idx in range(input_tensor_rank)])
+            end_mask_ = np.sum([2**idx if idx != axis else 0 for idx in range(input_tensor_rank)])
+
             splited_tensors.append(
                 tf.strided_slice(
                     input_=input_tensor,
                     begin=begin_,
                     end=end_,
+                    begin_mask=begin_mask_,
+                    end_mask=end_mask_,
+                )
+            )
+    elif isinstance(split, np.ndarray) \
+        and len(list(split)) > 1 \
+        and np.prod(split) != 1 \
+        and np.all(split == split[0]) \
+        and isinstance(input_tensor_shape[axis], int) \
+        and input_tensor_shape[axis] == np.sum(split):
+        # strided_slice - Slice everything in same size
+        # Suppression of FlexSplitV generation
+        # https://github.com/PINTO0309/onnx2tf/issues/751
+        splited_tensors = []
+        split_size = split[0]
+        for split_idx in range(len(list(split))):
+            begin_ = [
+                split_size * split_idx if idx == axis else 0 for idx in range(input_tensor_rank)
+            ]
+            end_ = []
+            for idx in range(input_tensor_rank):
+                if idx == axis:
+                    end_.append(split_size * split_idx + split_size)
+                elif input_tensor_shape[idx] is None:
+                    end_.append(0)
+                else:
+                    end_.append(input_tensor_shape[idx])
+
+            begin_mask_ = np.sum([2**idx if idx != axis else 0 for idx in range(input_tensor_rank)])
+            end_mask_ = np.sum([2**idx if idx != axis else 0 for idx in range(input_tensor_rank)])
+
+            splited_tensors.append(
+                tf.strided_slice(
+                    input_=input_tensor,
+                    begin=begin_,
+                    end=end_,
+                    begin_mask=begin_mask_,
+                    end_mask=end_mask_,
                 )
             )
     elif isinstance(split, np.ndarray) \

{onnx2tf-1.27.0 → onnx2tf-1.27.2}/onnx2tf/ops/Sub.py

@@ -114,9 +114,9 @@ def make_node(
     )
 
     # Workaround for ConvInteger
-    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64]:
+    if input_tensor_1.dtype == tf.float32 and input_tensor_2.dtype in [tf.int32, tf.int64, tf.float16]:
         input_tensor_2 = tf.cast(input_tensor_2, dtype=tf.float32)
-    elif input_tensor_1.dtype in [tf.int32, tf.int64] and input_tensor_2.dtype == tf.float32:
+    elif input_tensor_1.dtype in [tf.int32, tf.int64, tf.float16] and input_tensor_2.dtype == tf.float32:
         input_tensor_1 = tf.cast(input_tensor_1, dtype=tf.float32)
 
     # Disable unnecessary Transpose

{onnx2tf-1.27.0 → onnx2tf-1.27.2/onnx2tf.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: onnx2tf
-Version: 1.27.0
+Version: 1.27.2
 Summary: Self-Created Tools to convert ONNX files (NCHW) to TensorFlow/TFLite/Keras format (NHWC). The purpose of this tool is to solve the massive Transpose extrapolation problem in onnx-tensorflow (onnx-tf).
 Home-page: https://github.com/PINTO0309/onnx2tf
 Author: Katsuya Hyodo
@@ -18,6 +18,7 @@ Dynamic: description
 Dynamic: description-content-type
 Dynamic: home-page
 Dynamic: license
+Dynamic: license-file
 Dynamic: platform
 Dynamic: requires-python
 Dynamic: summary
@@ -280,7 +281,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
 
 ## Environment
 - Linux / Windows
-- onnx==1.16.1
+- onnx==1.17.0
 - onnxruntime==1.18.1
 - onnx-simplifier==0.4.33 or 0.4.30 `(onnx.onnx_cpp2py_export.shape_inference.InferenceError: [ShapeInferenceError] (op_type:Slice, node name: /xxxx/Slice): [ShapeInferenceError] Inferred shape and existing shape differ in rank: (x) vs (y))`
 - onnx_graphsurgeon
@@ -333,7 +334,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  ghcr.io/pinto0309/onnx2tf:1.27.0
+  ghcr.io/pinto0309/onnx2tf:1.27.2
 
   or
 
@@ -341,7 +342,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  docker.io/pinto0309/onnx2tf:1.27.0
+  docker.io/pinto0309/onnx2tf:1.27.2
 
   or
 
@@ -1512,7 +1513,7 @@ See: https://github.com/tensorflow/tfjs/tree/master/tfjs-converter
 When converting to CoreML, process as follows. The `-k` option is for conversion while maintaining the input channel order in ONNX's NCHW format.
 
 ```bash
-pip install coremltools
+pip install coremltools==8.2
 
 onnx2tf -i mobilenetv2-12.onnx -k input -ois input:1,3,224,224 -osd
 ```
@@ -1525,7 +1526,7 @@ model = ct.convert(
     model=FOLDER_PATH,
     source='tensorflow',
 )
-model.save(f'{FOLDER_PATH}/model.mlmodel')
+model.save(f'{FOLDER_PATH}/model.mlpackage')
 ```
 
 See: https://github.com/apple/coremltools