Resume training from a saved checkpoint about clip4clip HOT 4 CLOSED

Hi @zara0m, setting the --init_model with the saved checkpoint may be a choice, but it is imperfect because the checkpoint has not saved the optimization state. So, the best way is to save and resume the optimization state along with the checkpoint. Thanks.

from clip4clip.

@ArrowLuo Thanks.

from clip4clip.

we can use --init_model to resume the model from pytorch_model.bin.2, but we resume the pytorch_opt.bin.2, could you please give an specific command instance ?

from clip4clip.

Use both of them simultaneously to resume training:

--init_model ckpts/ckpt_msrvtt_retrieval_looseType/pytorch_model.bin.2
--resume_model ckpts/ckpt_msrvtt_retrieval_looseType/pytorch_opt.bin.2

from clip4clip.