about training 4M dataset and the loss converge slowly about recognize-anything

[Darylmeng]

I check the provided 4M data json files and find that only 4441 tags exists, which is smaller than the 4585 tags as claimed in the paper. May be the other 144 tags are in the 12M dataset? Will the missing labels cause bad effects on the training of 4M dataset training? As mentioned above, i relabel the json file and the model can't perform well.
I would be truly grateful if the author could offer any insights or responses to the questions I've presented.

[Darylmeng]

I also raise two more questions. In Appendix A in paper, batch_size is 720. However, the batch_size in config file "pretrain" is set to 52. According to the paper, the model is trained by 8 A100, so the batch_size should be 52X8=416? Also, the loss is computed as the sum of (loss_tag+loss_align+loss_diss). However, i observe the magnitudes of the three losses are in the hundreds, tens, and less than one, respectively. I'm curious if simply adding losses with such significant differences in magnitude without balancing them with any weights will affect the training? Sincerely awaiting for your reply.

[xinyu1205]

Did you load the ImageNet-pretrained image encoder weights? It is fixed by subsequent PR.