1. 初始化device

    if torch.cuda.is_available():
        if not opt.gpuid:#opt为命令行传入的参数对象
            opt.gpuid = 0
        opt.device = torch.device("cuda:%d" % opt.gpuid)
    else:
        opt.device = torch.device("cpu")
        opt.gpuid = -1
        print("CUDA is not available, fall back to CPU.")

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8

2. 将criterion，model，tensor转移到GPU上

	# model = Seq2SeqModel(opt)
	model = model.to(opt.device)
	criterion = torch.nn.CrossEntropyLoss().to(opt.device)
	#src: a LongTensor
	src = src.to(opt.device)

• 1
• 2
• 3
• 4
• 5

3. 从GPU取出数据

    src = src.cpu()

• 1