Custom Training

[fadi212]

How can we train on custom data and can we train the same model on Number Plates with multiple lines like two or three

[sirius-ai]

1. if it is not chinese license plate and license plate lenth equal to 8, you should delete check(label) funtcion in LPRDataLoader.getitem;
2. you should modify the list of CHARS base on your license plate characters, and the filename of train/test images should be its label character;
3. T_length should be bigger or equal than twice max license plate lenth;
4. our model not support multiple lines at currenct, maybe you can merge multiple lines into one line.

[2017TJM]

2 'targets' to have scalar type Int; but got torch.cuda.FloatTensor instead (while checking arguments for ctc_loss_gpu)

[sirius-ai]

@2017TJM please let me know how to reproduce this issue, thanks!

[gunjan5489]

Hi , im getting below error while train when i change T_length = 24, since my dataset include Indian License Plate , with single and double line license plate.

File "LPRNet_Train.py", line 94, in
loss = ctc_loss(log_probs, labels, input_lengths=input_lengths, target_lengths=target_lengths)
File "/home/arohi/Gunjan/Python Programming/venv/lib/python3.5/site-packages/torch/nn/modules/module.py", line 532, in call
result = self.forward(*input, **kwargs)
File "/home/arohi/Gunjan/Python Programming/venv/lib/python3.5/site-packages/torch/nn/modules/loss.py", line 1295, in forward
self.zero_infinity)
File "/home/arohi/Gunjan/Python Programming/venv/lib/python3.5/site-packages/torch/nn/functional.py", line 1781, in ctc_loss
zero_infinity)
RuntimeError: Expected input_lengths to have value at most 18, but got value 24 (while checking arguments for ctc_loss_cpu)

[anhthoai]

Anybody to help with 2 lines and 1 line license plates?

[duorouputao]

Have u solved the problem?

[gunjan5489]

Have u solved the problem?

the author has not done nothing relating to this issue

[jeungmin717]

@sirius-ai
Thanks for such a great work!

what does T_length mean?
i am confused with args.lpr_max_len

if I change lpr_max_len
then output shape could change?

for now my output shape is [1,50,18] Korean Character 50 is class score nums.
I want to make longer sequence than 18? is it available?