Question

已经训练了我的编码器-解码器模型，并使用：

保存了

model_state = {
             'encoder': encoder,
             'encoder_optimizer': encoder_optimizer,
             'decoder': decoder,
             'decoder_optimizer': decoder_optimizer
             }
torch.save(model_state, "best_model.pth.tar")

当我独立使用模型时，这可以很好地工作，但是当我尝试在另一个应用程序中使用模型时，这会给我错误。因此，我尝试加载模型并将编码器和解码器另存为state_dicts。这适用于我的编码器，但是当我尝试时：

checkpoint = torch.load(path_to_model, map_location=torch.device("cpu"))
decoder = checkpoint['decoder']
decoder = decoder.to(device)
encoder = checkpoint['encoder']
encoder = encoder.to(device)
torch.save(encoder.state_dict(), 'encoder.dict')
torch.save(decoder.state_dict(), 'decoder.dict')

它在torch.save(decoder.state_dict(), 'decoder.dict')上失败，并且出现错误：

 File "<stdin>", line 1, in <module>
  File "caption.py", line 31, in load_maps
    torch.save(decoder.state_dict(), 'decoder.dict')
  File "/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 695, in state_dict
    module.state_dict(destination, prefix + name + '.', keep_vars=keep_vars)
  File "/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 695, in state_dict
    module.state_dict(destination, prefix + name + '.', keep_vars=keep_vars)
  File "/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 696, in state_dict
    for hook in self._state_dict_hooks.values():
  File "/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 585, in __getattr__
    type(self).__name__, name))
AttributeError: 'Softmax' object has no attribute '_state_dict_hooks'

有没有一种方法可以绕过此错误或重新创建state_dict而不必重新训练我的模型？我不明白为什么要保存整个模型却无法从中获得state_dict，这大概是模型的一部分。

这是调用for m in decoder.modules(): print(m)的输出：

DecoderWithAttention(
  (attention): Attention(
    (encoder_att): Linear(in_features=2048, out_features=512, bias=True)
    (decoder_att): Linear(in_features=512, out_features=512, bias=True)
    (full_att): Linear(in_features=512, out_features=1, bias=True)
    (relu): ReLU()
    (softmax): Softmax(dim=1)
  )
  (embedding): Embedding(9490, 512)
  (dropout): Dropout(p=0.5, inplace=False)
  (decode_step): LSTMCell(2560, 512, bias=1)
  (init_h): Linear(in_features=2048, out_features=512, bias=True)
  (init_c): Linear(in_features=2048, out_features=512, bias=True)
  (f_beta): Linear(in_features=512, out_features=2048, bias=True)
  (sigmoid): Sigmoid()
  (fc): Linear(in_features=512, out_features=9490, bias=True)
)
Attention(
  (encoder_att): Linear(in_features=2048, out_features=512, bias=True)
  (decoder_att): Linear(in_features=512, out_features=512, bias=True)
  (full_att): Linear(in_features=512, out_features=1, bias=True)
  (relu): ReLU()
  (softmax): Softmax(dim=1)
)
Linear(in_features=2048, out_features=512, bias=True)
Linear(in_features=512, out_features=512, bias=True)
Linear(in_features=512, out_features=1, bias=True)
ReLU()
Softmax(dim=1)
Embedding(9490, 512)
Dropout(p=0.5, inplace=False)
LSTMCell(2560, 512, bias=1)
Linear(in_features=2048, out_features=512, bias=True)
Linear(in_features=2048, out_features=512, bias=True)
Linear(in_features=512, out_features=2048, bias=True)
Sigmoid()
Linear(in_features=512, out_features=9490, bias=True)

Answer 1

尝试这样保存模型。

torch.save({'state_dict': decoder.state_dict()}, 'decoder.pth.tar')

从Pytorch中的序列化模型获取state_dict时出错

1 个答案: