Pre-trained models - Githubissues

yash0307 / RecallatK_surrogate

Code for Recall@k Surrogate Loss with Large Batches and Similarity Mixup, CVPR 2022.

MIT License

56 stars 8 forks source link

Pre-trained models #4

Closed elias-ramzi closed 2 years ago

elias-ramzi commented 2 years ago

Hi,

Unfortunately I was not able to run the repository (I tried on iNaturalist). I had CPU memory issues when trying to load a batch, so I tried to modify the repository:

I used torch.utils.data.Subset to load sub-batches;
I used mixed-precision to speed-up training. But the results were way bellow the ones reported in your paper. Were you planing to release the weights of your models ?

Thank you!

yash0307 commented 2 years ago

Hi, I am not able to reproduce the issue that you mentioned. What is the CPU memory?

yash0307 commented 2 years ago

closing due to inactivity, feel free to reopen.