My Huggingface model gives an error when trying "Hosted Inference API"

[CanKeles5]

I trained a ConvTasNet model on a custom dataset and I uploaded the model to HuggingFace. The model works fine when I try to use it by loading it from huggingface with "ConvTasNet.from_pretrained("/link/to/model")", but when I try the hosted inference api it gives me the following error: "No module named 'torch.torch_version'".

The model is available here: https://huggingface.co/cankeles/ConvTasNet_WHAMR_enhsingle_16k

I am wondering if I am doing something wrong.

[JorisCos]

Weird, I have checked several models uploaded by our community on HF some work some don't with an unknown error @mpariente is it from our side or HF's ?

[mpariente]

Thanks for the quick reaction !
Could it be a problem of reading files/resampling related to sampling rate? (both non-working models are 16k)

[CanKeles5]

Hi, thank you all for the quick responses. I uploaded the models manually, could that cause a problem? Maybe the tags could be problematic too? I first used a different dataset tag that wasnt already in use and I changed it later to LibriMix.

[osanseviero]

Given the error is related to pytorch imports I don't think it's an issue with sampling rate. The tags don't have any impact in the Inference code. In case you're curious, the code is in https://github.com/huggingface/huggingface_hub/blob/main/api-inference-community/docker_images/asteroid/app/pipelines/audio_to_audio.py.

[Narsil]

This is now fixed. we merely rebuilt the docker file.

It's probably linked to the fact that your code runs on torch==1.10 where the docker image was running an older version of torch (torch_version.py seems new: https://github.com/pytorch/pytorch/blame/master/torch/torch_version.py, only 7 months old)

Since the model is serialized through pickle, it contained the torch_version in your file, and it couldn't unpickle it on an older lib.

Since asteroid does not define a hard number of torch version, the docker building took the latest version 1.10 and everything worked.

[mpariente]

Thanks everybody !