ONNX Inference Decoding

[rafaelagrc]

Hello,

I have converted the 'PARSeq' model from the Torch Hub to the ONNX format.
I would like to ask if anyone has done inference and decoding with the ONNX model, as one cannot use the tokenizer.decode() function for this purpose.

[dankernel]

In general, onnx can transform the model, but preprocessing and postprocessing does not.
tokenizer.decode is post-processing outside of the model, so it is not converted along with the model.
In my case, I just implemented it separately because the code is simple and there is no advantage of using an accelerator.

[Shivanshmundra]

@dankernel can you share your implementation of tokenizer outside of the model please?

[WongVi]

I found the torch version of parseq which is able to convert onnx and tensorrt too
https://github.com/bharatsubedi/PARseq_torch