Help: Convert ICNET_0.5 to onnx file about fast-semantic-segmentation

[oandrienko]

Hey @SiR0N, thanks for your interest in the project. Unfortunately, I haven't played around with ONNX or their conversion script. However, I think your problem might be related to using a Frozen Graph Def as a SavedModel (since you mention you just rename). Although they are similar, the formats are different.

Taking a quick look - the conversion tool seems to allow a Frozen Graph Def as an input, you just need to modify the flags you're using. Or, you can always convert the the model to a SavedModel.

Hope this helps! Will close this issue for now but let me know if you have any issues with the checkpoints.

[SiR0N]

Hello!
thanks for your reply, it was really useful, I thought that a SavedModel and Frozen were the same, I should have checked deeply the tool before asking.

I could convert it with this:
python -m tf2onnx.convert --graphdef .\0818_icnet_0.5_1025_resnet_v1\frozen_inference_graph.pb --output frozen.onnx --fold_const --opset 10 --inputs inputs:0 --outputs predictions:0

I attach the file in case you want to have a look. (Have not checked it yet)
ICNET_0.5.onnx.zip

[oandrienko]

Awesome I'm glad that it was helpful! The input and output tensors look correct so sounds like it would work. Good luck!

[sctrueew]

@SiR0N Hi,

How can I use the onnx model in C++ or python? Can you share the evolution code?

[SiR0N]

Hi @zpmmehrdad ,
unfortunately, I have no access to the code anymore.
I could make it running in python, check this for pre/postprocessing: https://modeldepot.io/oandrienko/icnet-for-fast-segmentation
The only part to change is the inference, to do that I would recommend having a look here (step 3): https://github.com/microsoft/onnxruntime/blob/master/docs/python/tutorial.rst in the sess.run you should include your inputs and output names, to do so open the onnx file with netron: netron. It is really useful when working with onnx, it shows all the info about the computational graph.

Hope it helps you!