multigpu about pytorch-yolov3 HOT 9 CLOSED

you can refer to https://github.com/ujsyehao/yolov3-multigpu

from pytorch-yolov3.

Support for multigpu is on the todo-list.

from pytorch-yolov3.

@eriklindernoren Sorry to ask, do we have support for multigpu now? Thanks

from pytorch-yolov3.

just add

model =  torch.nn.DataParallel(model)

and run again.

from pytorch-yolov3.

@LeoCHANGcn It seems that multi-gpus version can not accelerate the training. I compared single GPU and multi gpus, one epoch costed similiar time. During your test, did you have this problem?

from pytorch-yolov3.

I have also noticed the same issue.

from pytorch-yolov3.

@LeoCHANGcn
I added the following at Initiate model,

@train.py
model = Darknet(opt.model_config_path)
model =  torch.nn.DataParallel(model)

but new error on loss.backward()
grad can be implicitly created only for scalar outputs
does loss function should be change too?

from pytorch-yolov3.

Hi,
this issue is coming up since loss.backward()is equivalent to loss.backward(torch.Tensor([1])).
When using DataParallel this is invalid. Instead you should use either
loss.backward(torch.Tensor([1, 1])) or loss.sum().backward().
The behaviour of both should be the exact same.

from pytorch-yolov3.

@Robmosh91 Hi,when I used your suggestion,thefollowing issue occurred:'DataParallel' object has no attribute 'losses'?

from pytorch-yolov3.