About the DataParallel and the assignment to anonther device of the model

[zdy023]

I tried to build the model on another gpu device instead of cuda:0. I tried the with statement in torch and tried to pass the device to the gpu parameter of Reader as well. In both cases, the init is ok.

However, it will reports an exception like

RuntimeError: module must have its parameters and buffers on device cuda:0 (device_ids[0]) but found one of them on device: cuda:1

This seems to be caused by missing specifying the device_ids parameter in DataParallel in the 62nd line in get_detector function in detection.py and the 157th line in get_recognizer function in recognition.py.

I modified the original code such as

net = torch.nn.DataParallel(net).to(device)

to

net = torch.nn.DataParallel(net, device_ids=[torch.device(device)]).to(device)

And the exception disappeared.

Maybe you could make a little modification to fix this minor problem. Thanks very much!

[rkcosmos]

thanks, I'll look into this.

[lorenzo-stacchio]

Good morning everyone, i solved the problem by simply loading the model for multi-gpu as usual and then saving weights for single-gpu with
torch.save(net.state_dict(), "weights_single_gpu.pth") and then modifyng the code removing the usage of nn.DataParallel with the following:
net.load_state_dict(copyStateDict(torch.load( "weights_single_gpu.pth", map_location=device))).to(device)
In this way i think you could uncouple the training part from the inference one.