stage2_with_object365.sh

[lhqxcc]

Hello, I encountered an issue while reproducing your work. I used your stage2_with_object365.sh file for training, but found that the batch size is very small, which leads to a very long training time. Moreover, the loss keeps increasing during the training process. Could you please tell me what might be causing this problem?

[SY-Xuan]

Do you modify the batch_size? How many GPUs do you use? Do you load the weight of stage1?

[lhqxcc]

Thank you very much for your response. Below is my training script. Please keep the script you provided unchanged:

I found the problem in the training log file:

What do you think might have caused this problem?

[SY-Xuan]

Maybe you can try a learning rate=2.5e-4. Sometimes it may have a loss spike.