坑

指定GPU

先用nvidia-smi查看gpu负载状态，然后跑torch的时候，指定可见的gpu ID就行

在运行python blablabla前面加一句变成CUDA_VISIBLE_DEVICES=0,1,2 python blablabla，意义是可见的gpu有0,1,2号，而且torch会按照0,1,2这个顺序来使用gpu

链接

• pytorch中tensor维度操作

optimizer 和 cuda 的 bug

大概出现这种错时

  File "xxx", line 128, in update
    self.optimizer.step()
  File "/home/xalanq/python/lib/python3.6/site-packages/torch/optim/adamax.py", line 75, in step
    exp_avg.mul_(beta1).add_(1 - beta1, grad)
RuntimeError: Expected object of type torch.FloatTensor but found type torch.cuda.FloatTensor for argument #4 'other'

加上类似这份代码即可

self.optimizer.load_state_dict(state_dict['optimizer'])
if self.opt['cuda']:
    for state in self.optimizer.state.values():
        for k, v in state.items():
            if torch.is_tensor(v):
                state[k] = v.cuda()

Github issue