Fused AdamW_SGD optimizer issues
vealocia opened this issue · comments
Hi, authors! Thanks for your awesome work!
I'm confused about the usage of fused AdamW_SGD optimizer as described in paper Appendix C, paragraph implementation details.
It's said you use AdamW with 1e-3 lr and 0.05 wd for ViT vision encoder, and SGD with 0.02 lr and 1e-4 wd for text transformer.
However, in your configuration, ViT-B/32 is also optimized by SGD instead of fused AdamW_SGD. So which optimizer is your choice in experiment actually?
And, if you use fused AdamW_SGD optimizer just as said in paper, why did you use it? CLIP only uses AdamW optimizer. Is this beneficial to CLIP?
Looking forward for your reply!😁
In the DeCLIP open-source model in our paper, the AdamW_SGD optimizer was used for optimization. The settings here is slightly different from the one described in the paper because the test configuration was changed from another configuration of the same model (as you can see, the training setup is all for the YFCC dataset).
We use the AdamW_SGD optimizer in the DeCLIP paper during training because we experimentally found that the language encoder tends to crash using the AdamW optimizer when using noisy labels.
This is not a problem on the YFCC dataset, so the AdamW optimizer also works well.
Get it! Thanks for your answers.
By the way, @zlccccc, can you share any training logs of your experiments (either CLIP or DeCLIP or others)? That will be helpful and greatly appreciated.