Onnx to Tensorflow, model becomes too complicated.

[Maximilianxu]

Hi to all. As indicated by the title, I have some questions.
The following is the onnx graph.

The following is the tensorflow graph converted from onnx.

I have some questions on the latter graph.

1. The second maxpool operator is very complicated and a SparseToDense operator was added to the graph. This is the most confusing part for me. Why does a dense model have a SparseToDense operation? As I want to use TVM to compile this model which does not support SparseToDense, I am stucking at this position...
2. The output node was missing.

[arnaudbrejeon]

I have similar issue with missing output node and I don't know how to solve it

[andreydung]

Second this too. It seems that they're trying to handle conversion from NCHW in ONNX to NHWC in tensorflow?

[jspyth]

I have same problem with it,how to solve it?

[JiayiFu]

About the problem: the output graph is too complicated. I guess that use the onnx node name as TensorFlow name scope may make the graph more readable... #524

[fumihwh]

The reason is ONNX support 0 value in shape (second input) of Reshape which means unchanged.
https://github.com/onnx/onnx/blob/master/docs/Operators.md#reshape
In tf, no mechanism like that, so we add some codes to support such case.

For an ad-hoc fix, you can comment out onnx_tf/handlers/backend/reshape.py L25 - L36, and modify L39 to node, inputs=[tensor, shape], attrs=attrs, **kwargs).

[jakubondrus]

@Maximilianxu did you manage to solve 2nd problem? Seems like there are multiple reported issues on this topic but can't find any answer. (trying to convert arcface model from onnx model zoo)

[Maximilianxu]

@jakubondrus No. I gave up onnx that time.

[fumihwh]

@jakubondrus
BTW,

1. which tf version you are using?
2. What is the name of const node at the bottom of the graph?

[jakubondrus]

@fumihwh

1. TF 2.X (I don't know what was the newest tf version at time I was dealing with this). Onnx-tf was installed from git project.
2. If you mean that last black AddV2 node, it's name is fc1 (the same name as node that is last before output node in onnx model). If you mean that const nodes, they have names like Const_1, Const_2, ... and concat/concat_dim, concat_1/concat_dim, ...

[fumihwh]

@jakubondrus Maybe we could add following code as post processing. Before that, you could try this locally.

def optimize_graph_def(graph_def, input_names, output_names):
  from tensorflow.tools.graph_transforms import TransformGraph
  transforms = [
    'remove_nodes(op=Identity)',
    'strip_unused_nodes',
    'fold_batch_norms',
    'fold_constants(ignore_errors=true)',
  ]
  optimized_graph_def = TransformGraph(graph_def, input_names, output_names,
                                       transforms)
  return optimized_graph_def

[jakubondrus]

@fumihwh I am sorry for late answer, didn't get to that immediately adn then forgot about it. Where am I supposed to put that code?

[dimitraka]

Is there any solution of this issue finally?