GPU训练模型转到CPU

通常GPU训练会用到nn.Dataparallel，所以转到cpu不能仅仅利用map_location=’cpu’。
样例：

if opt.pretrain_path:
    print('loading pretrained model {}'.format(opt.pretrain_path))
    pretrain = torch.load(opt.pretrain_path, map_location='cpu')
    state_dict = pretrain['state_dict']
    from collections import OrderedDict
    state_dict_new = OrderedDict()
    for k, v in state_dict.items():
        name = k[7:]  # 去掉module
        state_dict_new[name] = v
    model.load_state_dict(state_dict_new)