Samir Yitzhak Gadre

Results 8 comments of Samir Yitzhak Gadre

Hi @ypwang61! If I understand correctly, you are currently using `dataset2metadata` to compute your `post_process_feature`, which is a mask that will affect CLIP score computation. In this step you can...

Yeah! Check out this notebook from Hila Chefer: https://github.com/hila-chefer/Transformer-MM-Explainability/blob/main/CLIP_explainability.ipynb Found this to qualitatively work pretty well on OAI ViT B/32 model!

Here is a preliminary investigation on incomplete streams. Seems that this does not explain the hash mismatch, at least not at clean byte boundaries. Will now try to track down...

For my use case I realized that I did not need `tf_agents`, so I ended up re-writing parts of the codebase to remove this dependency. It might be the case...

Hi @isidentical, thanks for the questions! In our scaling experiments we scaled both the IN1k and IN21k strategies up to the large pool (filtering 1.28B samples). Looking at Table 27...

Hi @alexjc thanks for the issue! I’ll try to reproduce on my end, but in the meantime can your please share the command lines you are running?

One thing that jumps out is not sure 3090s have bf16 support. May also be worth trying fp32 and fp16 for debugging