RuntimeError：尝试反序列化CUDA设备2上的对象，但torch.cuda.device_count（）为1

Question

我有一段用于训练模型的python代码。问题是跑完后：

loaded_state = torch.load(model_path+seq_to_seq_test_model_fname)

加载预训练模型，我得到：

  Traceback (most recent call last):
  File "img_to_text.py", line 480, in <module>
    main()
  File "img_to_text.py", line 475, in main
    r = setup_test()
  File "img_to_text.py", line 259, in setup_test
    s2s_data = s2s.setup_test()
  File "/media/ahrzb/datasets/notebooks/mzh/SemStyle/semstyle/code/seq2seq_pytorch.py", line 220, in setup_test
    loaded_state= torch.load(model_path+seq_to_seq_test_model_fname)
  File "/home/ahrzb/.pyenv/versions/2.7.15/envs/mzh2.7/lib/python2.7/site-packages/torch/serialization.py", line 358, in load
    return _load(f, map_location, pickle_module)
  File "/home/ahrzb/.pyenv/versions/2.7.15/envs/mzh2.7/lib/python2.7/site-packages/torch/serialization.py", line 542, in _load
    result = unpickler.load()
  File "/home/ahrzb/.pyenv/versions/2.7.15/envs/mzh2.7/lib/python2.7/site-packages/torch/serialization.py", line 508, in persistent_load
    data_type(size), location)
  File "/home/ahrzb/.pyenv/versions/2.7.15/envs/mzh2.7/lib/python2.7/site-packages/torch/serialization.py", line 372, in restore_location
    return default_restore_location(storage, location)
  File "/home/ahrzb/.pyenv/versions/2.7.15/envs/mzh2.7/lib/python2.7/site-packages/torch/serialization.py", line 104, in default_restore_location
    result = fn(storage, location)
  File "/home/ahrzb/.pyenv/versions/2.7.15/envs/mzh2.7/lib/python2.7/site-packages/torch/serialization.py", line 85, in _cuda_deserialize
    device, torch.cuda.device_count()))

我认为这是因为他们已经在两个GPU上训练了模型，我需要在一个GPU中加载它。我更改了这一行：

loaded_state = torch.load(model_path+seq_to_seq_test_model_fname) 

至

loaded_state = torch.load(model_path+seq_to_seq_test_model_fname, map_location={'cuda:1': 'cuda:0'} ) 

为了将cuda 1的数据映射到cuda 0，但它没有用。

Answer 1

我只是想通了：

 loaded_state = torch.load(model_path+seq_to_seq_test_model_fname,map_location='cuda:0')

是解决方案