Unable to reproduce the results on Pascal VOC datasets about defrcn HOT 5 CLOSED

If you change the number of GPUs, I think both the batch-size and the number of iterations should be changed accordingly to ensure that the model is fully trained (to ensure fair comparison with this repo), see issue #19 for more details.

from defrcn.

@er-muyue What if the batch size is same, but the total number of GPU is changed. DO we need to modify the setting to reproduce the result?

from defrcn.

@er-muyue If I understand correctly, the stated batch size (which is 16) means total batch size instead of batch size per GPU. In the above experiments, I kept the total batch size unchanged (16) even though I changed the GPU number from 8 to 2.

In your last comment, do you mean to suggest changing the total batch size and number of iterations if I changed the GPU number? If yes, could you please suggest a direction to change (e.g. in 10-shot) so that the result gap between reported and reproduced can be reduced? Thanks.

from defrcn.

Just realized the result reported of Pascal voc is AP50 and result reported of coco is mAP. The result is able to be reproduced. Thanks.

from defrcn.

Hello, I would like to ask if your versions of python, cuda and pytorch are the same as the author's, because I use a 3090 graphics card, and the versions of cuda and pytorch cannot be consistent with the author's. Only the number of GPUs has been modified in the code, and the others have not been modified. The reproduced experimental results are far from the author's.Thank you very much.

from defrcn.