Convergence and force output about lavis

[dxli94]

Hi @mcjinye, thanks for your interest.

You can change max_epoch to limit the number of epochs, see https://github.com/salesforce/LAVIS/blob/main/lavis/projects/blip/train/caption_coco_ft.yaml#L34.

We don't track the convergence of the model.

You may force saving the last model after the max_epochs by adding self._save_checkpoint(cur_epoch, is_best=False) to https://github.com/salesforce/LAVIS/blob/main/lavis/runners/runner_base.py#L387.

Thanks.

[mcjinye]

Hi @dxli94, thank you for the quick response! I think I misunderstand the tools. I can use the generated .pth file and register this model for my own use case. Is this the correct routine?

I have a follow-up question on this topic. I want to do a zero-shot classification with my own dataset. If I don't want to use the existing model, should I pretrain the base model (blip_pretrain) or tune this blip_feature_extractor? I notice that in the example you are using blip_feature_extractor. But when I retrain it, it seems like the repo is not supporting retraining blip_feature_extractor. (Although I feel like the final layer of blip_pretrain is what we have from blip_feature_extractor).

Sorry for a lot of questions. This is a great work. Thanks!