onnx2tf 1.24.0__py3-none-any.whl → 1.25.9__py3-none-any.whl

onnx2tf/ops/Gather.py

@@ -64,13 +64,79 @@ def make_node(
     indices = tf_layers_dict[graph_node_input_2.name]['tf_node'] \
         if isinstance(graph_node_input_2, gs.Variable) else graph_node_input_2
 
+    axis = graph_node.attrs.get("axis", 0)
+    axis = convert_axis(
+        axis=axis,
+        tensor_rank=len(input_tensor.shape),
+        before_op_output_shape_trans=before_op_output_shape_trans,
+    )
+
+    # Param replacement - axis
+    axis = replace_parameter(
+        value_before_replacement=axis,
+        param_target='attributes',
+        param_name='axis',
+        **kwargs,
+    )
+
+    nhwc = tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+        if isinstance(graph_node_input_1, gs.Variable) \
+            and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
+
     before_cast_indices = None
     if isinstance(indices, np.ndarray) and indices.ndim == 1 and len(indices) == 1 and indices[0] is not None:
         if indices[0] >= 0:
             before_cast_indices = indices[0]
+            # 直前が Shape だった場合のみの特別なワークアラウンドで、入力がNHWCで確定しているときはindicesを変換する
+            # 1. ind=0 のときはそのまま
+            # 2. ind=1 のときは末尾
+            # 3. ind=2 のときは1
+            # 4. ind=3 のときは2
+            # 5. ind=4 のときは3
+            #
+            # ONNX: ind=2
+            #   0,1,2 -> 0,2,1
+            #   0,1,2,3 -> 0,2,3,1
+            #   0,1,2,3,4 -> 0,2,3,4,1
+            #   0,1,2,3,4,5 -> 0,2,3,4,5,1
+            if nhwc and graph_node.i().op == 'Shape':
+                input_tensor_rank = input_tensor.shape[0]
+                if before_cast_indices == 0:
+                    # batch
+                    pass
+                elif before_cast_indices == 1:
+                    # channel
+                    before_cast_indices = input_tensor_rank - 1
+                else:
+                    # spartial dim
+                    before_cast_indices = before_cast_indices - 1
+
     elif isinstance(indices, np.ndarray) and indices.ndim == 0 and indices is not None:
         if indices >= 0:
             before_cast_indices = int(indices)
+            # 直前が Shape だった場合のみの特別なワークアラウンドで、入力がNHWCで確定しているときはindicesを変換する
+            # 1. ind=0 のときはそのまま
+            # 2. ind=1 のときは末尾
+            # 3. ind=2 のときは1
+            # 4. ind=3 のときは2
+            # 5. ind=4 のときは3
+            #
+            # ONNX: ind=2
+            #   0,1,2 -> 0,2,1
+            #   0,1,2,3 -> 0,2,3,1
+            #   0,1,2,3,4 -> 0,2,3,4,1
+            #   0,1,2,3,4,5 -> 0,2,3,4,5,1
+            if nhwc and graph_node.i().op == 'Shape':
+                input_tensor_rank = input_tensor.shape[0]
+                if before_cast_indices == 0:
+                    # batch
+                    pass
+                elif before_cast_indices == 1:
+                    # channel
+                    before_cast_indices = input_tensor_rank - 1
+                else:
+                    # spartial dim
+                    before_cast_indices = before_cast_indices - 1
 
     simple_indices = None
     if isinstance(indices, np.ndarray) and indices.ndim == 1 and None not in indices:
@@ -81,21 +147,6 @@ def make_node(
     shape = graph_node_output.shape
     dtype = graph_node_output.dtype
 
-    axis = graph_node.attrs.get("axis", 0)
-    axis = convert_axis(
-        axis=axis,
-        tensor_rank=len(input_tensor.shape),
-        before_op_output_shape_trans=before_op_output_shape_trans,
-    )
-
-    # Param replacement - axis
-    axis = replace_parameter(
-        value_before_replacement=axis,
-        param_target='attributes',
-        param_name='axis',
-        **kwargs,
-    )
-
     optimization_for_gpu_delegate: bool = \
         kwargs['optimization_for_gpu_delegate']
 
@@ -141,9 +192,7 @@ def make_node(
         'optype': graph_node.op,
         'shape': shape,
         'dtype': dtype,
-        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
-            if isinstance(graph_node_input_1, gs.Variable) \
-                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
+        'nhwc': nhwc,
     }
 
     # Param replacement

onnx2tf/ops/Mod.py

@@ -157,8 +157,37 @@ def make_node(
         is_scalar_2_rank = tf.rank(input_tensor_2) == 0
         if hasattr(is_scalar_2_rank, 'numpy'):
             is_scalar_2 = is_scalar_2_rank.numpy()
+
         if (is_scalar_1 or is_scalar_2) and graph_node.i().op == 'Gemm':
             pass
+        elif (is_scalar_1 or is_scalar_2) and graph_node.i().op != 'Gemm':
+            first_tensor = None
+            second_tensor = None
+            if is_scalar_1:
+                first_tensor = input_tensor_2
+                second_tensor = input_tensor_1
+            elif is_scalar_2:
+                first_tensor = input_tensor_1
+                second_tensor = input_tensor_2
+            tmp_result = tf.math.mod(first_tensor, second_tensor)
+            tmp_result_shape = tmp_result.shape
+            if first_tensor.shape == tmp_result_shape:
+                pass
+            else:
+                input_tensor_1, input_tensor_2 = \
+                    pre_explicit_broadcast(
+                        input_tensor_1=input_tensor_1,
+                        input_tensor_2=input_tensor_2,
+                    )
+
+                input_tensor_1, input_tensor_2 = \
+                    explicit_broadcast(
+                        const_or_var_1=input_tensor_1,
+                        const_or_var_2=input_tensor_2,
+                        graph_node=graph_node,
+                        tf_layers_dict= tf_layers_dict,
+                    )
+
         else:
             input_tensor_1, input_tensor_2 = \
                 pre_explicit_broadcast(

onnx2tf/ops/Mul.py

@@ -152,14 +152,44 @@ def make_node(
     try:
         is_scalar_1 = False
         is_scalar_2 = False
+        is_partial_scalar = False
         is_scalar_1_rank = tf.rank(input_tensor_1) == 0
         if hasattr(is_scalar_1_rank, 'numpy'):
             is_scalar_1 = is_scalar_1_rank.numpy()
         is_scalar_2_rank = tf.rank(input_tensor_2) == 0
         if hasattr(is_scalar_2_rank, 'numpy'):
             is_scalar_2 = is_scalar_2_rank.numpy()
+
         if (is_scalar_1 or is_scalar_2) and graph_node.i().op == 'Gemm':
             pass
+        elif (is_scalar_1 or is_scalar_2) and graph_node.i().op != 'Gemm':
+            first_tensor = None
+            second_tensor = None
+            if is_scalar_1:
+                first_tensor = input_tensor_2
+                second_tensor = input_tensor_1
+            elif is_scalar_2:
+                first_tensor = input_tensor_1
+                second_tensor = input_tensor_2
+            tmp_result = tf.math.multiply(first_tensor, second_tensor)
+            tmp_result_shape = tmp_result.shape
+            if first_tensor.shape == tmp_result_shape:
+                pass
+            else:
+                input_tensor_1, input_tensor_2 = \
+                    pre_explicit_broadcast(
+                        input_tensor_1=input_tensor_1,
+                        input_tensor_2=input_tensor_2,
+                    )
+
+                input_tensor_1, input_tensor_2 = \
+                    explicit_broadcast(
+                        const_or_var_1=input_tensor_1,
+                        const_or_var_2=input_tensor_2,
+                        graph_node=graph_node,
+                        tf_layers_dict= tf_layers_dict,
+                    )
+
         else:
             input_tensor_1, input_tensor_2 = \
                 pre_explicit_broadcast(

onnx2tf/ops/ReduceL1.py

@@ -106,6 +106,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceL2.py

@@ -106,6 +106,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceLogSum.py

@@ -106,6 +106,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceLogSumExp.py

@@ -106,6 +106,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceMax.py

@@ -113,6 +113,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceMean.py

@@ -107,6 +107,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceMin.py

@@ -107,6 +107,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceProd.py

@@ -107,6 +107,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceSum.py

@@ -106,6 +106,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/ReduceSumSquare.py

@@ -107,6 +107,9 @@ def make_node(
         'optype': graph_node.op,
         'shape': onnx_output_shape,
         'dtype': dtype,
+        'nhwc': tf_layers_dict[graph_node_input_1.name]['nhwc'] \
+            if isinstance(graph_node_input_1, gs.Variable) \
+                and 'nhwc' in tf_layers_dict[graph_node_input_1.name].keys() else False
     }
 
     onnx_tensor_infos_for_validation: Dict[str:np.ndarray] = kwargs['onnx_tensor_infos_for_validation']

onnx2tf/ops/Shape.py

@@ -259,6 +259,8 @@ def make_node(
                                 if consumer_node.op == 'Resize':
                                     consumer_nodes.append(consumer_node)
                                     consumer_count += 1
+                                else:
+                                    break
                             except:
                                 break
                         if consumer_count == 1:

onnx2tf/ops/Sub.py

@@ -156,8 +156,37 @@ def make_node(
         is_scalar_2_rank = tf.rank(input_tensor_2) == 0
         if hasattr(is_scalar_2_rank, 'numpy'):
             is_scalar_2 = is_scalar_2_rank.numpy()
+
         if (is_scalar_1 or is_scalar_2) and graph_node.i().op == 'Gemm':
             pass
+        elif (is_scalar_1 or is_scalar_2) and graph_node.i().op != 'Gemm':
+            first_tensor = None
+            second_tensor = None
+            if is_scalar_1:
+                first_tensor = input_tensor_2
+                second_tensor = input_tensor_1
+            elif is_scalar_2:
+                first_tensor = input_tensor_1
+                second_tensor = input_tensor_2
+            tmp_result = tf.math.subtract(first_tensor, second_tensor)
+            tmp_result_shape = tmp_result.shape
+            if first_tensor.shape == tmp_result_shape:
+                pass
+            else:
+                input_tensor_1, input_tensor_2 = \
+                    pre_explicit_broadcast(
+                        input_tensor_1=input_tensor_1,
+                        input_tensor_2=input_tensor_2,
+                    )
+
+                input_tensor_1, input_tensor_2 = \
+                    explicit_broadcast(
+                        const_or_var_1=input_tensor_1,
+                        const_or_var_2=input_tensor_2,
+                        graph_node=graph_node,
+                        tf_layers_dict= tf_layers_dict,
+                    )
+
         else:
             input_tensor_1, input_tensor_2 = \
                 pre_explicit_broadcast(

onnx2tf/ops/Transpose.py

@@ -160,10 +160,24 @@ def make_node(
                     perm = new_perm
 
     # Preserving Graph Structure (Dict)
+    nwhc = False
+    if nwc_nhwc_ndhwc_keep:
+        nhwc = True
+    elif isinstance(graph_node_input, gs.Variable) \
+        and 'nhwc' in tf_layers_dict[graph_node_input.name].keys():
+        nhwc = tf_layers_dict[graph_node_input.name]['nhwc']
+        if nhwc and not before_op_output_shape_trans and perm == [i for i in range(len(input_tensor_shape))]:
+            nhwc = True
+        else:
+            nhwc = False
+    else:
+        nhwc = False
+
     tf_layers_dict[graph_node_output.name] = {
         'optype': graph_node.op,
         'shape': output_shape,
         'dtype': dtype,
+        'nhwc': nhwc,
         'nwc_nhwc_ndhwc_keep': nwc_nhwc_ndhwc_keep,
     }
 

onnx2tf/utils/common_functions.py

@@ -2185,7 +2185,7 @@ def process_neg_idx(
             and not isinstance(indices_shape[-1], np.ndarray) \
             and not isinstance(indices_shape[-1], tf.Tensor) \
             and tf_keras.backend.is_keras_tensor(indices_shape[-1]):
-            if data_shape != tf.TensorShape([None]):
+            if tf.TensorShape(None) not in data_shape :
                 max_i = tf.cast(
                     tf.strided_slice(
                         input_=data_shape,
@@ -5202,7 +5202,19 @@ def merge_two_consecutive_identical_ops_into_one(
                 tf_type = tf.math.divide
 
     elif tf_func == 'Sub':
-        if (
+        if isinstance(input_tensor_1, np.ndarray) or hasattr(input_tensor_1, 'numpy'):
+            tf_layers_dict[graph_node_output.name]['tf_node'] = \
+                tf.math.subtract(
+                    x=input_tensor_1 \
+                        if not isinstance(input_tensor_1, np.ndarray) \
+                            else tf.convert_to_tensor(input_tensor_1),
+                    y=input_tensor_2 \
+                        if not isinstance(input_tensor_2, np.ndarray) \
+                            else tf.convert_to_tensor(input_tensor_2),
+                    name=graph_node.name,
+                )
+            tf_type = tf.math.subtract
+        elif (
             not isinstance(graph_node_input_1, np.ndarray) \
                 and 'merge_sub' in tf_layers_dict[graph_node_input_1.name] \
                 and tf_layers_dict[graph_node_input_1.name]['merge_sub']
@@ -5411,12 +5423,7 @@ def merge_two_consecutive_identical_ops_into_one(
                         elif next_graph_node_o_op == 'Sub':
                             # 8. `Add` -> `Sub` to `Single-Add`  : `10 + 5 - 8 -> 10 - 3`
                             if isinstance(next_graph_node_input_1, np.ndarray) or hasattr(next_graph_node_input_1, 'numpy'):
-                                if isinstance(input_tensor_1, np.ndarray) or hasattr(input_tensor_1, 'numpy'):
-                                    input_tensor_1 = (input_tensor_1 - next_graph_node_input_1)
-                                elif isinstance(input_tensor_2, np.ndarray) or hasattr(input_tensor_2, 'numpy'):
-                                    input_tensor_2 = (input_tensor_2 - next_graph_node_input_1)
-                                tf_layers_dict[graph_node_output.name]['merge_add'] = True
-                                tf_type = tf.identity
+                                tf_type = tf.math.add
                             elif isinstance(next_graph_node_input_2, np.ndarray) or hasattr(next_graph_node_input_2, 'numpy'):
                                 if isinstance(input_tensor_1, np.ndarray) or hasattr(input_tensor_1, 'numpy'):
                                     input_tensor_1 = (input_tensor_1 - next_graph_node_input_2)