can't train planrcnn with multiple gpu when I modify the batchsize>1 in dataloader?

[ywcmaike]

can't train planrcnn with multiple gpu when I modify the batchsize>1 in dataloader?
hello, I have 4 titanxp, and I want to train planercnn, in train_planercnn.py
dataloader = DataLoader(dataset, batch_size=1, shuffle=True, num_workers=16) and it only use one gpu. when I modify the batch_size>1(such as 4), it failed to train. the error is as followed.

return torch.stack(batch, 0, out=out)
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0,
I found the similar issue: https://discuss.pytorch.org/t/problem-with-custom-dataset-invalid-argument-0-sizes-of-tensors-must-match-except-in-dimension-0/34792 , but can't solve it.

can you give me some tips? or how you train planercnn with multiple gpu and more batchsize? @tmbdev @vinodgro @mjgarland @dumerrill @sdalton1

[ywcmaike]

@art-programmer @kihwan23

[art-programmer]

Sorry for the late response. I'm sorry but I have no experience with multi-gpu training. Maybe you can find more information on the original mask r-cnn repo, https://github.com/multimodallearning/pytorch-mask-rcnn

[shsjxzh]

I find that in the original mask-rcnn repo, the author mentions in README that he uses the ROIAlign implemented by others, and it only supports one gpu training: https://github.com/longcw/RoIAlign.pytorch. Maybe this information can help you.