报错记录(nlp
)
文章目录
1. ‘lengths‘ argument should be a 1D CPU int64 tensor, but got 1D cuda:0 Long tensor
解决:Torch
版本问题,Pytorch1.5
以上版本升级了Bi-LSTM
导致,直接的方法可以直接修改库代码(如果没有读懂当前代码可以一试)
修改的库文件:
D:\software\anaconda3\envs\nlp\Lib\site-packages\torch\nn\utils\rnn.py
其中的函数:
pack_padded_sequence
修改的代码部分:
data, batch_sizes = _VF._pack_padded_sequence(input, lengths.cpu(), batch_first)
中的lengths.cpu()
,原代码:lengths
# 修改前
data, batch_sizes = \
_VF._pack_padded_sequence(input, lengths, batch_first)
# 修改后
data, batch_sizes = \
_VF._pack_padded_sequence(input, lengths.cpu(), batch_first)
更好的方法自然是修改自身代码。
2. RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper__index_select)
解决:依旧是Torch
版本问题,可以看出高版本的Torch
会给出警告,new_tensor()
函数已经不推荐用于构建新张量。
UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than tensor.new_tensor(sourceTensor).
sequences_lengths.new_tensor(torch.arange(0, len(sequences_lengths)))
基于new_tensor()
构建的张量设备与原张量不一致,因此可以进一步指定予以解决。或者依据sourceTensor.clone().detach()
进行创建。
# 修改前
idx_range =\
sequences_lengths.new_tensor(torch.arange(0, len(sequences_lengths)))
# 修改后
idx_range =\
sequences_lengths.new_tensor(torch.arange(0, len(sequences_lengths)))
idx_range = idx_range.to(sequences_lengths.device)