onnx2tf 1.27.6__tar.gz → 1.27.8__tar.gz

{onnx2tf-1.27.6/onnx2tf.egg-info → onnx2tf-1.27.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnx2tf
-Version: 1.27.6
+Version: 1.27.8
 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
@@ -334,7 +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:1.27.6
+  ghcr.io/pinto0309/onnx2tf:1.27.8
 
   or
 
@@ -342,7 +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:1.27.6
+  docker.io/pinto0309/onnx2tf:1.27.8
 
   or
 

{onnx2tf-1.27.6 → onnx2tf-1.27.8}/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:1.27.6
+  ghcr.io/pinto0309/onnx2tf:1.27.8
 
   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:1.27.6
+  docker.io/pinto0309/onnx2tf:1.27.8
 
   or
 

{onnx2tf-1.27.6 → onnx2tf-1.27.8}/onnx2tf/__init__.py

@@ -1,3 +1,3 @@
 from onnx2tf.onnx2tf import convert, main
 
-__version__ = '1.27.6'
+__version__ = '1.27.8'

{onnx2tf-1.27.6 → onnx2tf-1.27.8}/onnx2tf/onnx2tf.py

@@ -1947,6 +1947,8 @@ def convert(
                     validated_onnx_tensor: np.ndarray = checked_value[0]
                     matched_flg: int = checked_value[1]
                     max_abs_err: Any = checked_value[2]
+                    onnx_shape_tf_shape: str = checked_value[3]
+
                     message = ''
                     if matched_flg == 0:
                         message = \
@@ -1960,11 +1962,11 @@ def convert(
                     elif matched_flg == 2:
                         message = \
                             Color.GREEN(f'validate_result') + ': ' +\
-                            Color.REVERSE(f'{Color.BLUE} Skipped (Deleted or Shape Unmatched) ')
+                            Color.REVERSE(f'{Color.BLUE} Skipped (Deleted or Shape Unmatched) {onnx_shape_tf_shape}')
                     print(
                         Color.GREEN(f'INFO:') + ' '+
-                        Color.GREEN(f'onnx_output_name') + f': {onnx_output_name} '+
-                        Color.GREEN(f'tf_output_name') + f': {tf_output_name} '+
+                        Color.GREEN(f'onnx_output_name') + f': {re.sub("^wa/", "/", onnx_output_name)} '+
+                        # Color.GREEN(f'tf_output_name') + f': {tf_output_name} '+
                         Color.GREEN(f'shape') + f': {validated_onnx_tensor.shape} '+
                         Color.GREEN(f'dtype') + f': {validated_onnx_tensor.dtype} '+
                         f'{message}'

{onnx2tf-1.27.6 → onnx2tf-1.27.8}/onnx2tf/ops/AveragePool.py

@@ -167,11 +167,46 @@ def make_node(
         input_tensor=input_tensor
     )
 
-    func = math.ceil if ceil_mode else math.floor
-    output_spatial_shape = [
-        func((i + pb + pe - d * (k - 1) - 1) / s + 1)
-        for i, pb, pe, k, d, s in zip(input_tensor_shape[1:-1], pads[:len(pads) // 2], pads[len(pads) // 2:], kernel_shape, dilations, strides)
-    ]
+    if not is_known_shape:
+        def compute_output_spatial_shape_from_tensor(input_tensor, pads, kernel_shape, dilations, strides, ceil_mode=False):
+            input_shape = tf.shape(input_tensor)  # Get dynamic shape
+            input_spatial = input_shape[1:-1]  # Extract spatial dimensions only (NHWC format)
+
+            pad_begin = pads[:len(pads) // 2]
+            pad_end = pads[len(pads) // 2:]
+
+            round_func = tf.math.ceil if ceil_mode else tf.math.floor
+
+            output_spatial = []
+            for i, pb, pe, k, d, s in zip(tf.unstack(input_spatial), pad_begin, pad_end, kernel_shape, dilations, strides):
+                i = tf.cast(i, tf.float32)
+                pb = tf.constant(pb, dtype=tf.float32)
+                pe = tf.constant(pe, dtype=tf.float32)
+                k = tf.constant(k, dtype=tf.float32)
+                d = tf.constant(d, dtype=tf.float32)
+                s = tf.constant(s, dtype=tf.float32)
+
+                numerator = i + pb + pe - d * (k - 1) - 1
+                raw_output = numerator / s + 1
+                output_dim = tf.cast(round_func(raw_output), tf.int32)
+                output_spatial.append(output_dim)
+
+            return output_spatial
+
+        output_spatial_shape = compute_output_spatial_shape_from_tensor(
+            input_tensor=input_tensor,
+            pads=pads,
+            kernel_shape=kernel_shape,
+            dilations=dilations,
+            strides=strides,
+            ceil_mode=ceil_mode
+        )
+    else:
+        func = math.ceil if ceil_mode else math.floor
+        output_spatial_shape = [
+            func((i + pb + pe - d * (k - 1) - 1) / s + 1)
+            for i, pb, pe, k, d, s in zip(input_tensor_shape[1:-1], pads[:len(pads) // 2], pads[len(pads) // 2:], kernel_shape, dilations, strides)
+        ]
 
     # onnx padding value is ignored if auto_pad is not 'NOTSET'
     if auto_pad == 'NOTSET':
@@ -218,28 +253,52 @@ def make_node(
     # count nonzero elements in kernel each strides for the case count_include_pad is False
     non_zero_counts = []
 
-    for input_spatial_shape, output_size, kernel, dilation, stride, pads_begin, pads_end \
-            in zip(input_tensor_shape[1:-1], output_spatial_shape, kernel_shape,
-                   dilations, strides, pads[:len(pads) // 2], pads[len(pads) // 2:]):
-        sample_target = np.concatenate([
-            np.zeros(pads_begin),
-            np.ones(input_spatial_shape),
-            np.zeros(pads_end)]
+    if not is_known_shape:
+        def compute_non_zero_counts_loop(input_tensor, output_spatial_shape, kernel_shape, dilations, strides, pads):
+            
+            counts_list = []
+            
+            for dim in range(len(kernel_shape)):
+                k = kernel_shape[dim]
+                counts = [k] * (output_spatial_shape[dim].numpy() if tf.is_tensor(output_spatial_shape[dim]) and hasattr(output_spatial_shape[dim], 'numpy') else output_spatial_shape[dim])
+                counts_list.append(counts)
+                
+            return counts_list
+
+        non_zero_counts = compute_non_zero_counts_loop(
+            input_tensor=input_tensor,
+            output_spatial_shape=output_spatial_shape,
+            kernel_shape=kernel_shape,
+            dilations=dilations,
+            strides=strides,
+            pads=pads
         )
-        sample_kernel = np.zeros((kernel - 1) * dilation + 1)
-        sample_kernel[::dilation] = 1
+    else:
+        for input_spatial_shape, output_size, kernel, dilation, stride, pads_begin, pads_end \
+                in zip(input_tensor_shape[1:-1], output_spatial_shape, kernel_shape,
+                       dilations, strides, pads[:len(pads) // 2], pads[len(pads) // 2:]):
+            sample_target = np.concatenate([
+                np.zeros(pads_begin),
+                np.ones(input_spatial_shape),
+                np.zeros(pads_end)]
+            )
+            sample_kernel = np.zeros((kernel - 1) * dilation + 1)
+            sample_kernel[::dilation] = 1
 
-        counts = []
-        for i in range(output_size):
-            start = i * stride
-            stride_target = sample_target[start:start+len(sample_kernel)]
-            # pad target to match size
-            stride_target = np.concatenate([stride_target, np.zeros(len(sample_kernel) - len(stride_target))])
-            counts.extend(np.convolve(stride_target, sample_kernel, mode='valid'))
+            counts = []
+            for i in range(output_size):
+                start = i * stride
+                stride_target = sample_target[start:start+len(sample_kernel)]
+                # pad target to match size
+                stride_target = np.concatenate([stride_target, np.zeros(len(sample_kernel) - len(stride_target))])
+                counts.extend(np.convolve(stride_target, sample_kernel, mode='valid'))
 
-        non_zero_counts.append(counts)
+            non_zero_counts.append(counts)
 
-    need_multiplier = len(set([i for sublist in non_zero_counts for i in sublist])) != 1
+    if not is_known_shape:
+        need_multiplier = False  # Default to False for dynamic tensors to avoid errors
+    else:
+        need_multiplier = len(set([i for sublist in non_zero_counts for i in sublist])) != 1
 
     # default tensorflow option for count_include_pad is True and cannot control
     # average value should be compensated in cases below
@@ -295,7 +354,9 @@ def make_node(
                 multiplier[-1] = k / (k - extra_pad)
                 average_multiplier.append(multiplier)
         else:
-            for i, k, non_zero_count, extra_pad in enumerate(zip(kernel_shape, non_zero_counts, extra_pads)):
+            for i in range(len(kernel_shape)):
+                k = kernel_shape[i]
+                extra_pad = extra_pads[i]
                 average_multiplier[i][-1] = k / (k - extra_pad)
 
     # Preserving Graph Structure (Dict)

{onnx2tf-1.27.6 → onnx2tf-1.27.8}/onnx2tf/utils/common_functions.py

@@ -1,6 +1,7 @@
 import math
 import os
 import io
+import re
 import sys
 import copy
 import json
@@ -280,7 +281,7 @@ def print_node_info(func):
             if graph_input is not None:
                 debug(
                     Color.GREEN(f'INFO:') + ' '+
-                    Color.GREEN(f'input_op_name') + f': {graph_input.name} '+
+                    Color.GREEN(f'input_op_name') + f': {re.sub("^wa/", "/", graph_input.name)} '+
                     Color.GREEN(f'shape') + f': {graph_input.shape} '+
                     Color.GREEN(f'dtype') + f': {graph_input.dtype}'
                 )
@@ -294,18 +295,18 @@ def print_node_info(func):
                 )
                 debug(
                     Color.GREEN(f'INFO:') + ' ' + Color.MAGENTA(f'onnx_op_type') + ': '+
-                    f'{graph_node.op}' + Color.MAGENTA(' onnx_op_name') + f': {graph_node.name}')
+                    f'{graph_node.op}' + Color.MAGENTA(' onnx_op_name') + f': {re.sub("^wa/", "/", graph_node.name)}')
                 for idx, graph_node_input in enumerate(graph_node.inputs):
                     debug(
                         Color.GREEN(f'INFO:') + ' '+
-                        Color.CYAN(f' input_name.{idx+1}') + f': {graph_node_input.name} '+
+                        Color.CYAN(f' input_name.{idx+1}') + f': {re.sub("^wa/", "/", graph_node_input.name)} '+
                         Color.CYAN(f'shape') + f': {graph_node_input.shape} '+
                         Color.CYAN(f'dtype') + f': {graph_node_input.dtype}'
                     )
                 for idx, graph_node_output in enumerate(graph_node.outputs):
                     debug(
                         Color.GREEN(f'INFO:') + ' '+
-                        Color.CYAN(f' output_name.{idx+1}') + f': {graph_node_output.name} '+
+                        Color.CYAN(f' output_name.{idx+1}') + f': {re.sub("^wa/", "/", graph_node_output.name)} '+
                         Color.CYAN(f'shape') + f': {graph_node_output.shape} '+
                         Color.CYAN(f'dtype') + f': {graph_node_output.dtype}'
                     )
@@ -4033,18 +4034,19 @@ def onnx_tf_tensor_validation(
 
     Returns
     ----------
-    check_results: Dict[str, List[np.ndarray, int, float|int]]
+    check_results: Dict[str, List[np.ndarray, int, float|int], str]
         Tensor Comparison Results
         {
             onnx_output_name: [
                 onnx_tensor,
                 matched_flg, <--- 0: Unmatched, 1: Matched, 2: Skipped (Deleted or Shape Unmatched),
                 max_abs_err,
+                onnx_shape_tf_shape,
             ]
         }
     """
     check_results = {
-        k: [v[0], False, 0.0] \
+        k: [v[0], False, 0.0, ""] \
             for k, v in output_pairs.items()
     }
 
@@ -4120,9 +4122,12 @@ def onnx_tf_tensor_validation(
             # If there was no match between ONNX and TensorFlow output shapes.
             check_results[names_pair][1] = 2
             check_results[names_pair][2] = max_abs_err
+            check_results[names_pair][3] = \
+                f"onnx.shape:{onnx_tensor.shape if hasattr(onnx_tensor, 'shape') else 'None'}/tf.shape:{tf_tensor.shape if hasattr(tf_tensor, 'shape') else 'None'}"
         else:
             check_results[names_pair][1] = validate_result
             check_results[names_pair][2] = max_abs_err
+            check_results[names_pair][3] = ""
 
     return check_results
 

{onnx2tf-1.27.6 → onnx2tf-1.27.8/onnx2tf.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: onnx2tf
-Version: 1.27.6
+Version: 1.27.8
 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
@@ -334,7 +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:1.27.6
+  ghcr.io/pinto0309/onnx2tf:1.27.8
 
   or
 
@@ -342,7 +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:1.27.6
+  docker.io/pinto0309/onnx2tf:1.27.8
 
   or