Conversion from ONNX model fails with Relu

[victorromeo]

Simple models with Relu fail when converting via ONNX frontend, as Relu shape and dtype information is missing (None).

May I please confirm if ONNX is intended to be supported?

frontend/onnx.python

... # [line277]
  @staticmethod
  def _handle_relu(op_info):
    input_tensor = op_info.input_tensors[0]
    op_info.output_tensors[0].dtype = input_tensor.dtype
    op_info.output_tensors[0].shape = input_tensor.shape[:]
...

[dboyliao]

Yes. It is. And it is our plan to support PyTorch. We haven't use onnx for a while, but I'm pretty sure it was working. Maybe there is sth I miss because of some updates in PyTorch or ONNX.

Can you attach your model file here so I can take a look on what the bug is?

[victorromeo]

I've added an example here #133 using Keras2onnx to convert a basic MNIST keras model. In this example, the tests/test_keras_onnx_model fails in _handle_relu static method when assigning the shape.

I believe the _build_graph operataion in ONNX is ultimately responsible for not transferring the shape or dtype information correctly. May I get your thoughts?

My goal is to get transpose + avg pool operations supported this week if possible on my own dev branches, and our preference was to avoid tflite micro.

Thanks.

[victorromeo]

I believe the operation _build_intermediate_ops needs shape and dtype information, but its hard coded to None.

def _build_intermediate_ops(self, onnx_graph, ugraph, op_types_cnt, tensor_names_map):
    """Build all intermediate nodes, the nodes that is not in neither initialization list nor input list
    """
    # create all outupt tensors
    for node in onnx_graph.node:
      cnt = op_types_cnt[node.op_type]
      node_name = self._format_node_name(node.name, node.op_type, cnt)
      op_types_cnt[node.op_type] += 1
      for i, name in enumerate(node.output):
        tensor_names_map[name] = TensorInfo(
          name=self._format_tensor_name(name, node_name, i),
          op_name=node_name,
          dtype=None, # <== Should be set/calculated?
          shape=None, # <== Should be set/calculated?
          ugraph=ugraph
        )
    # create ops
    for node in onnx_graph.node:
      input_tensors = [
        tensor_names_map[name] for name in node.input
      ]
      output_tensors = [
        tensor_names_map[name] for name in node.output
      ]
      op_attr = {
        attrib_pb.name: _convert_op_attribute(attrib_pb)
        for attrib_pb in node.attribute
      }
      node_name = output_tensors[0].op_name
      OperationInfo(
        name=node_name,
        input_tensors=input_tensors,
        output_tensors=output_tensors,
        op_type=node.op_type,
        lib_name='onnx',
        ugraph=ugraph,
        op_attr=op_attr
      )