我用自己的训练集转lmdb，CRNN训练时loss都是20多，不收敛。我只识别数字和字母

[ZhangZoffy]

No description provided.

[ZhangZoffy]

这是数据集，自己造的

[ShangLe0607]

大神你好，能不能指教一下请问我训练的时候出现这问题是咋回事啊，
[root@localhost crnn]# python3 train.py
Traceback (most recent call last):
File "train.py", line 2, in
from torch.nn.modules.loss import CTCLoss
ImportError: cannot import name 'CTCLoss' from 'torch.nn.modules.loss' (/usr/local/lib/python3.7/site-packages/torch/nn/modules/loss.py)
[root@localhost crnn]#

[ZhangZoffy]

大神你好，能不能指教一下请问我训练的时候出现这问题是咋回事啊，
[root@localhost crnn]# python3 train.py
Traceback (most recent call last):
File "train.py", line 2, in
from torch.nn.modules.loss import CTCLoss
ImportError: cannot import name 'CTCLoss' from 'torch.nn.modules.loss' (/usr/local/lib/python3.7/site-packages/torch/nn/modules/loss.py)
[root@localhost crnn]#

安装warpctc吧，先降一下torch版本，我被这个问题坑了2天，降版本搞定

[ShangLe0607]

@ZhangZoffy 装了warpctc,torch版本降到了1.3.0，但是还报错

段错误(吐核)，该怎么办呢

[ZhangZoffy]

@ZhangZoffy 装了warpctc,torch版本降到了1.3.0，但是还报错
段错误(吐核)，该怎么办呢
必须安装torch0.4.0版本，超过1.0的都不行

[ShangLe0607]

@ZhangZoffy 谢谢你，都装好了，还是会报错，是咋回事啊[0/20][20/282] Loss: 42.145580
[0/20][40/282] Loss: 19.428555
[0/20][60/282] Loss: 17.523825
[0/20][80/282] Loss: 16.701580
[0/20][100/282] Loss: 16.425825
[0/20][120/282] Loss: 16.065802
[0/20][140/282] Loss: 15.806343
[0/20][160/282] Loss: 15.621722
[0/20][180/282] Loss: 15.718269
[0/20][200/282] Loss: 15.642500
[0/20][220/282] Loss: 15.453972
[0/20][240/282] Loss: 15.550603
[0/20][260/282] Loss: 15.239004
Traceback (most recent call last):
File "train.py", line 165, in
cost = trainBatch(net, criterion, optimizer, train_iter)
File "train.py", line 76, in trainBatch
t, l = converter.encode(cpu_texts)
File "/opt/crnn/lib/convert.py", line 47, in encode
index = self.dict[char]
KeyError: '\ufeff'

[ZhangZoffy]

@ZhangZoffy 谢谢你，都装好了，还是会报错，是咋回事啊[0/20][20/282] Loss: 42.145580
[0/20][40/282] Loss: 19.428555
[0/20][60/282] Loss: 17.523825
[0/20][80/282] Loss: 16.701580
[0/20][100/282] Loss: 16.425825
[0/20][120/282] Loss: 16.065802
[0/20][140/282] Loss: 15.806343
[0/20][160/282] Loss: 15.621722
[0/20][180/282] Loss: 15.718269
[0/20][200/282] Loss: 15.642500
[0/20][220/282] Loss: 15.453972
[0/20][240/282] Loss: 15.550603
[0/20][260/282] Loss: 15.239004
Traceback (most recent call last):
File "train.py", line 165, in
cost = trainBatch(net, criterion, optimizer, train_iter)
File "train.py", line 76, in trainBatch
t, l = converter.encode(cpu_texts)
File "/opt/crnn/lib/convert.py", line 47, in encode
index = self.dict[char]
KeyError: '\ufeff'

你这个应该是编码问题，py2和py3在编码上的区别