could you provide the results and configuration only using one or two GPUs(A6000)?

[codwest]

I am interested in your project. But I only have two A6000 GPUs, and I cannot reproduce the results reported in your paper. If you have time, could you try to train your model on the SBD and COCOLVIS datasets using one or two GPUs. If you could, could you let me know what the final results are and their configurations, especially the batch size value?
Another question is: What are the configurations of your Xtiny model if it is also trained using one GPU or two GPUs? what results?
I need your help.
thank you.

[qinliuliuqin]

@codwest Hi, thanks for your question. Training with fewer GPUs would not change the results significantly if the training epochs were similar. Could you provide your results? Are they substantially worse than those in the paper? As for the batch size, you just need to decrease it accordingly. For example, new_batch_size = new_#GPUs * (old_batch_size // old_#GPUs). Training the xTiny model is the same as training other models except for the batch size.

Feel free to let me know if you still have issues.

[codwest]

Thank you very much and give me a detailed explanation.

I trained your Xtiny model using one A6000 GPU with default parameters in your code. So, batch_size=32, num_epochs=1001. The final test result is:

Did you have a similar result? Could you provide your results and training configuration?

[qinliuliuqin]

@codwest The training curve looks normal, but the SBD results in the table are worse than I obtained. As reported in the paper, I obtained 4.71 and 7.09 on the SBD dataset for xTiny trained on C+L. However, I didn't record all the other results. BTW, can you reproduce the results in the paper using the released xTiny model?