[Experimental / WIP] Generate a miniscule model for each OP only, retain the results of test inference, and propagate to all OPs. STEP.1,2

README.md

@@ -34,7 +34,7 @@ Video speed is adjusted approximately 50 times slower than actual speed.
   $ docker run --rm -it \
   -v `pwd`:/workdir \
   -w /workdir \
-  ghcr.io/pinto0309/onnx2tf:1.6.7
+  ghcr.io/pinto0309/onnx2tf:1.7.0
 
   or
 

onnx2tf/__init__.py

@@ -1,3 +1,3 @@
 from onnx2tf.onnx2tf import convert, main
 
-__version__ = '1.6.7'
+__version__ = '1.7.0'

onnx2tf/onnx2tf.py

@@ -662,6 +662,7 @@ def convert(
 
     # Define additional parameters
     additional_parameters = {
+        'input_onnx_file_path': input_onnx_file_path if input_onnx_file_path is not None else None,
         'onnx_graph': onnx_graph,
         'opset': graph.opset,
         'batch_size': batch_size,

onnx2tf/ops/Add.py

@@ -16,7 +16,10 @@
     post_process_transpose,
     disable_unnecessary_transpose,
     shape_unmatched_special_avoidance_workaround,
+    make_tf_partial_model_inputs,
+    dummy_tf_inference,
 )
+from typing import Any, Dict, List
 
 
 @print_node_info
@@ -65,12 +68,27 @@ def make_node(
         'dtype': dtype,
     }
 
-    # Generation of TF OP
     input_tensor_1 = tf_layers_dict[graph_node_input_1.name]['tf_node'] \
         if isinstance(graph_node_input_1, gs.Variable) else graph_node_input_1
     input_tensor_2 = tf_layers_dict[graph_node_input_2.name]['tf_node'] \
         if isinstance(graph_node_input_2, gs.Variable) else graph_node_input_2
 
+    # Acquisition of test data for validation
+    if not isinstance(graph_node_input_1, np.ndarray) \
+        and 'verification_data' in tf_layers_dict[graph_node_input_1.name].keys():
+        test_data1: np.ndarray = tf_layers_dict[graph_node_input_1.name]['verification_data']
+    elif isinstance(graph_node_input_1, np.ndarray):
+        test_data1: np.ndarray = graph_node_input_1
+    else:
+        test_data1 = None
+    if not isinstance(graph_node_input_2, np.ndarray) \
+        and 'verification_data' in tf_layers_dict[graph_node_input_2.name].keys():
+        test_data2: np.ndarray = tf_layers_dict[graph_node_input_2.name]['verification_data']
+    elif isinstance(graph_node_input_2, np.ndarray):
+        test_data2: np.ndarray = graph_node_input_2
+    else:
+        test_data2 = None
+
     # Disable unnecessary Transpose
     #   1. If both x and y are gs.Variable
     #   2. If only one of the two is the output of Transpose
@@ -122,12 +140,59 @@ def make_node(
         **kwargs,
     )
 
+    # Generate input OPs for TensorFlow subgraphs
+    # For inference testing on OP stand-alone
+    tf_partial_model_inputs: List[tf.keras.Input] = \
+        make_tf_partial_model_inputs(
+            input_tensors=[
+                input_tensor_1,
+                input_tensor_2,
+            ]
+        )
+    tf_partial_model_outputs = None
+
+    # Generation of TF OP
+    ### Overall model
     tf_layers_dict[graph_node_output.name]['tf_node'] = \
         tf.math.add(
             x=input_tensor_1,
             y=input_tensor_2,
             name=graph_node.name,
         )
+    ### Partial model
+    if tf_partial_model_inputs is not None:
+        tf_partial_model_outputs = \
+            [
+                tf.math.add(
+                    x=tf_partial_model_inputs[0],
+                    y=tf_partial_model_inputs[1],
+                )
+            ]
+        tf_partial_model = tf.keras.Model(
+            inputs=tf_partial_model_inputs,
+            outputs=tf_partial_model_outputs,
+        )
+        test_data1 = None
+        if isinstance(input_tensor_1, np.ndarray):
+            test_data1 = input_tensor_1
+        test_data2 = None
+        if isinstance(input_tensor_2, np.ndarray):
+            test_data2 = input_tensor_2
+        tf_partial_model_result_infos: Dict[Any] = dummy_tf_inference(
+            model=tf_partial_model,
+            inputs=tf_partial_model_inputs,
+            verification_datas=[
+                test_data1,
+                test_data2,
+            ]
+        )
+        tf_layers_dict[graph_node_output.name]['verification_data'] = \
+            list(tf_partial_model_result_infos.values())[0]
+        del tf_partial_model
+        del tf_partial_model_inputs
+        del tf_partial_model_outputs
+        del test_data1
+        del test_data2
 
     # Post-process transpose
     tf_layers_dict[graph_node_output.name]['tf_node'] = post_process_transpose(

onnx2tf/ops/Concat.py

@@ -14,7 +14,10 @@
     make_tf_node_info,
     pre_process_transpose,
     post_process_transpose,
+    make_tf_partial_model_inputs,
+    dummy_tf_inference,
 )
+from typing import Any, Dict, List
 
 
 @print_node_info
@@ -136,12 +139,77 @@ def make_node(
         value if len(value.shape) > 0 else tf.reshape(value, [1]) for value in values
     ]
 
+    # Generate input OPs for TensorFlow subgraphs
+    # For inference testing on OP stand-alone
+    tf_partial_model_inputs: List[tf.keras.Input] = \
+        make_tf_partial_model_inputs(
+            input_tensors=values
+        )
+    tf_partial_model_outputs = None
+
+    # Generation of TF OP
+    ### Overall model
     tf_layers_dict[graph_node_output.name]['tf_node'] = \
         tf.concat(
             values=values,
             axis=axis,
             name=graph_node.name,
         )
+    ### Partial model
+    if tf_partial_model_inputs is not None:
+        test_datas = []
+        for graph_node_input, value in zip(graph_node.inputs, values):
+            test_data = None
+            if not isinstance(value, np.ndarray):
+                if not isinstance(graph_node_input, np.ndarray) \
+                    and 'verification_data' in tf_layers_dict[graph_node_input.name].keys():
+                    test_data: np.ndarray = tf_layers_dict[graph_node_input.name]['verification_data']
+                elif isinstance(graph_node_input, np.ndarray):
+                    test_data: np.ndarray = graph_node_input
+                else:
+                    test_data = None
+            else:
+                test_data = value
+            test_datas.append(test_data)
+        new_test_datas = []
+        is_expanded = False
+        for tf_partial_model_input, test_data in zip(tf_partial_model_inputs, test_datas):
+            if isinstance(test_data, np.ndarray) \
+                and len(test_data.shape) == 1 \
+                and len(tf_partial_model_input.shape) == 2:
+                test_data = np.expand_dims(test_data, 0)
+                is_expanded = True
+            new_test_datas.append(test_data)
+        test_datas = new_test_datas
+        tf_partial_model_outputs = \
+            [
+                tf.concat(
+                    values=tf_partial_model_inputs,
+                    axis=axis,
+                ) if not is_expanded else \
+                tf.squeeze(
+                    input=tf.concat(
+                        values=tf_partial_model_inputs,
+                        axis=axis if not is_expanded else axis+1,
+                    ),
+                    axis=0,
+                )
+            ]
+        tf_partial_model = tf.keras.Model(
+            inputs=tf_partial_model_inputs,
+            outputs=tf_partial_model_outputs,
+        )
+        tf_partial_model_result_infos: Dict[Any] = dummy_tf_inference(
+            model=tf_partial_model,
+            inputs=tf_partial_model_inputs,
+            verification_datas=test_datas
+        )
+        tf_layers_dict[graph_node_output.name]['verification_data'] = \
+            list(tf_partial_model_result_infos.values())[0]
+        del tf_partial_model
+        del tf_partial_model_inputs
+        del tf_partial_model_outputs
+        del test_data
 
     # Post-process transpose
     tf_layers_dict[graph_node_output.name]['tf_node'] = post_process_transpose(