Single card(RTX 3090) training results about maptr

[sjhfdl]

We trained it on the two A100 GPUS, and the Map result is about 0.35 in the epoch 24

Hello, I solved this problem, the reason is that the paper used 8 Gpus for training, and I trained on a single card, so I reduced the initial learning rate lr and weight_decay by 8 times, changed to lr=0.75e-4, weight_decay=0.00125, and then decreased the initial learning rate LR and weight_decay by 8 times. Also, enlarge the warmup_iters in lr_config by a factor of eight, to 4000

[adasfag]

I have also encounter the same problem

[adasfag]

We trained it on the two A100 GPUS, and the Map result is about 0.35 in the epoch 24

[adasfag]

Thanks

[adasfag]

Hello .I train the model again as your advice,but the Map is about 45.7 in the epoch 24. Could you provicd your single GPU result?Thank you very much

[sjhfdl]

Hello .I train the model again as your advice,but the Map is about 45.7 in the epoch 24. Could you provicd your single GPU result?Thank you very much

First of all, I would like to apologize to you. Due to the computing power of my graphics card, when I adjusted the learning rate, I only trained the author's code for two epochs, and I felt that the accuracy of the second epoch had reached 0.15, so I did not continue the training. Then I went to verify my method, and the accuracy of the training was similar to the results given by the author. My idea is that the results of the multi-card run will be slightly lower than those of the single card, and then I assume that the method of the author can also run on my own computer and produce similar results as in the paper.

[adasfag]

Hello .I train the model again as your advice,but the Map is about 45.7 in the epoch 24. Could you provicd your single GPU result?Thank you very much

First of all, I would like to apologize to you. Due to the computing power of my graphics card, when I adjusted the learning rate, I only trained the author's code for two epochs, and I felt that the accuracy of the second epoch had reached 0.15, so I did not continue the training. Then I went to verify my method, and the accuracy of the training was similar to the results given by the author. My idea is that the results of the multi-card run will be slightly lower than those of the single card, and then I assume that the method of the author can also run on my own computer and produce similar results as in the paper.

It seems that the single result is lower than those of the multi-card, maybe it needs a more suitable lr and It confused me.

[sjhfdl]

It seems that the single result is lower than those of the multi-card, maybe it needs a more suitable lr and It confused me.

Yes, you need a good learning rate configuration, you can try it a few times, maybe because our graphics card models are different

[sjhfdl]

Hello .I train the model again as your advice,but the Map is about 45.7 in the epoch 24. Could you provicd your single GPU result?Thank you very much

Could you show me the test results of the training？

[adasfag]

[lrx02]

I have met this problem as well