multi GPU training？

[PangzeCheung]

I set the gpu_ids 2,3, but the program only runs on the GPU 2. Could you please tell me is the code support for multi GPU training? Thank you!

[RenYurui]

You can use torch.nn.DataParallel to train the model using multi-GPU. See here

Specifically, if you want to train the pose-guided person image generation task, you can modify the "__ init __" function in pose_model.py. Add

self.net_G = torch.nn.DataParallel(self.net_G, device_ids=self.gpu_ids)
self.net_D = torch.nn.DataParallel(self.net_D, device_ids=self.gpu_ids)

[RenYurui]

Currently, only the face animation model supports multi-GPU training.
We will update the code soon.
Thanks for asking.

[PangzeCheung]

@RenYurui Thank you very much!

[BhavanJ]

Hi @RenYurui,

Nice work!
It seems like even after I use the DataParallel in pose_flownet using below command, the model still uses a single GPU.
self.net_G = torch.nn.DataParallel(self.net_G, device_ids=self.gpu_ids)

It seems like all data is only loaded on first GPU in your code as show below:

            self.input_P1 = input_P1.cuda(self.gpu_ids[0], async=True)
            self.input_BP1 = input_BP1.cuda(self.gpu_ids[0], async=True)
            self.input_P2 = input_P2.cuda(self.gpu_ids[0], async=True)
            self.input_BP2 = input_BP2.cuda(self.gpu_ids[0], async=True)  

I tried to replace above with just .cuda() but still I am not able to spread batch data across multiple GPU's and first GPU is running out of memory when I uses larger batch size. Is it the case that your custom built CUDA operations don't support multiple GPUs?

Thanks,
Bhavan