Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Warning: Encountered known unsupported method torch.instance_norm

See original GitHub issue

Hi, thanks for the great repo!!! I tried with both installations: with and without plugins from master and got the same warnings. The conversion ended and speedup was large but the output was very different. For reference that’s the model I was trying to convert: https://github.com/KaiyangZhou/deep-person-reid/blob/master/torchreid/models/osnet_ain.py

Warning: Encountered known unsupported method torch.instance_norm
Warning: Encountered known unsupported method torch.nn.functional.instance_norm

Are you planning to support this? Or should I just switch to batch_norm?

Issue Analytics

State:
Created 4 years ago
Comments:7 (2 by maintainers)

Top GitHub Comments

1reaction

jaybdubcommented, Dec 6, 2019

Hi anuar12,

Thanks for your feedback!

I agree, it seems like it would be possible to implement with some combination of native TensorRT layers. Usually this is easier and more distributable than integrating a plugin.

I’ll look into it if I can find time. If you can easily use batch norm, that may be your safest bet for now.

Best, John

0reactions

anuar12commented, Dec 13, 2019

I just tested it on my network and it has worked! Thanks a lot John!

Tested my network (https://github.com/KaiyangZhou/deep-person-reid/blob/master/torchreid/models/osnet_ain.py) batch_size=8: [8, 3, 256, 128] images on Xavier: Plain Pytorch eval Forward-pass time: 0.048 s TRT FP32 Forward-pass time: 0.031 s TRT FP16 Forward-pass time: 0.018 s

Had to change flatten(1) to view() because flatten was not supported.