|
|
import os |
|
|
import pdb |
|
|
import torch |
|
|
import torch.nn as nn |
|
|
|
|
|
|
|
|
class GpuDataParallel(object): |
|
|
def __init__(self): |
|
|
self.gpu_list = [] |
|
|
self.output_device = None |
|
|
|
|
|
def set_device(self, device): |
|
|
device = str(device) |
|
|
if device != 'None': |
|
|
self.gpu_list = [i for i in range(len(device.split(',')))] |
|
|
os.environ["CUDA_VISIBLE_DEVICES"] = device |
|
|
output_device = self.gpu_list[0] |
|
|
self.occupy_gpu(self.gpu_list) |
|
|
self.output_device = output_device if len(self.gpu_list) > 0 else "cpu" |
|
|
|
|
|
def model_to_device(self, model): |
|
|
|
|
|
model = model.to(self.output_device) |
|
|
if len(self.gpu_list) > 1: |
|
|
model = nn.DataParallel( |
|
|
model, |
|
|
device_ids=self.gpu_list, |
|
|
output_device=self.output_device) |
|
|
return model |
|
|
|
|
|
def data_to_device(self, data): |
|
|
if isinstance(data, torch.FloatTensor): |
|
|
return data.to(self.output_device) |
|
|
elif isinstance(data, torch.DoubleTensor): |
|
|
return data.float().to(self.output_device) |
|
|
elif isinstance(data, torch.ByteTensor): |
|
|
return data.long().to(self.output_device) |
|
|
elif isinstance(data, torch.LongTensor): |
|
|
return data.to(self.output_device) |
|
|
elif isinstance(data, list) or isinstance(data, tuple): |
|
|
return [self.data_to_device(d) for d in data] |
|
|
else: |
|
|
raise ValueError(data.shape, "Unknown Dtype: {}".format(data.dtype)) |
|
|
|
|
|
def criterion_to_device(self, loss): |
|
|
return loss.to(self.output_device) |
|
|
|
|
|
def occupy_gpu(self, gpus=None): |
|
|
""" |
|
|
make program appear on nvidia-smi. |
|
|
""" |
|
|
if len(gpus) == 0: |
|
|
torch.zeros(1).cuda() |
|
|
else: |
|
|
gpus = [gpus] if isinstance(gpus, int) else list(gpus) |
|
|
for g in gpus: |
|
|
torch.zeros(1).cuda(g) |
|
|
|
