Model runs faster with Hugging Face Transformers

[apeguero1]

Hi everyone,

I thought I'd share this colab which ports the Image GPT weights into a Hugging Face transformers model. Runs about 10x faster for me.

[ssss1029]

Hi @apeguero1, Thanks for the colab! It seems there is a small bug in the notebook:

I'm thinking it should be

return img/127.5 - 1

[apeguero1]

Oh yes, thanks @ssss1029! Just fixed it 👍

[licoriceleaf]

Hello @apeguro1, while attempting to run this colab, I've been getting various 'name is not defined' errors. Any tips on how to deal with this? Thanks in advance.

[apeguero1]

Seems like you're missing a couple of imports. run

import matplotlib.pyplot as plt
import torch

[licoriceleaf]

Still getting 'name 'samples' is not defined' and 'name 'clusters' is not defined'. I ran the aforementioned code in line 7. Perhaps I ran the code in the wrong location?

[apeguero1]

Hmm.. could you try running all the cells by clicking 'Runtime' and then 'Run all'?

[licoriceleaf]

Seems to work now, thanks!

[vitormm44]

Hey @apeguero1 , I'm trying to use Image GPT for a classification task, but I'm not very sure about how to do it using your Colab. How exactly should I feed the model in the training phase?

[MayuO2]

Hey @apeguero1, can you please assist with this error?

[apeguero1]

Hi @MayuO2, just pushed up a fix! BTW the huggingface repo has now added image-gpt if you wanna check that out (:

[MayuO2]

Thank you @apeguero1! It is working perfectly now:) I appreciate you sharing the huggingface repo with us