Convolutional Recurrent Neural Network + CTCLoss

I think i have fixed the ctcloss nan problem!

Now!

Please pull the latest code from master.

Please update the pytorch to >= v1.2.0

Enjoy it!

PS: Once there is ctclossnan, please

1. Change the batchSize to smaller (eg: 8, 16, 32)
2. Change the lr to smaller (eg: 0.00001, 0.0001)
3. Contact me by emailing to [email protected]

Dependence

• CentOS7
• Python3.6.5
• torch==1.2.0
• torchvision==0.4.0
• Tesla P40 - Nvidia

Run demo

• Download a pretrained model from Baidu Cloud (extraction code: si32)

• People who cannot access Baidu can download a copy from Google Drive

• Run demo

  python demo.py -m path/to/model -i data/demo.jpg

  

  Expected output

  -妳----真---的的---可---------以 => 妳真的可以

Feature

• Variable length

  It support variable length.

• Chinese support

  I change it to binary mode when reading the key and value, so you can use it to do Chinese OCR.

• Change CTCLoss from warp-ctc to torch.nn.CTCLoss

  As we know, warp-ctc need to compile and it seems that it only support PyTorch 0.4. But PyTorch support CTCLoss itself, so i change the loss function to torch.nn.CTCLoss .

• Solved PyTorch CTCLoss become nan after several epoch

  Just don't know why, but when i train the net, the loss always become nan after several epoch.

  I add a param dealwith_lossnan to params.py . If set it to True , the net will autocheck and replace all nan/inf in gradients to zero.

• DataParallel

  I add a param multi_gpu to params.py . If you want to use multi gpu to train your net, please set it to True and set the param ngpu to a proper number.

Train your data

Prepare data

Folder mode

1. Put your images in a folder and organize your images in the following format:

   label_number.jpg

   For example

   • English

   hi_0.jpg hello_1.jpg English_2.jpg English_3.jpg E n g l i s h_4.jpg...

   • Chinese

   一身转战_0.jpg 三千里_1.jpg 一剑曾当百万师_2.jpg 一剑曾当百万师_3.jpg 一 剑 曾 当 百 万 师_3.jpg ...

   So you can see, the number is used to distinguish the same label.

2. Run the create_dataset.py in tool folder by

   python tool/create_dataset.py --out lmdb/data/output/path --folder path/to/folder

3. Use the same step to create train and val data.

4. The advantage of the folder mode is that it's convenient! But due to some illegal character can't be in the path

   

   So the disadvantage of the folder mode is that it's labels are limited.

File mode

1. Your data file should like

   absolute/path/to/image/一身转战_0.jpg
   一身转战
   absolute/path/to/image/三千里_1.jpg
   三千里
   absolute/path/to/image/一剑曾当百万师_2.jpg
   一剑曾当百万师
   absolute/path/to/image/3.jpg
   一剑曾当百万师
   absolute/path/to/image/一 剑 曾 当 百 万 师_4.jpg
   一 剑 曾 当 百 万 师
   absolute/path/to/image/xxx.jpg
   label of xxx.jpg
   .
   .
   .

   DO REMEMBER:

   1. It must be the absolute path to image.
   2. The first line can't be empty.
   3. There are no blank line between two data.

2. Run the create_dataset.py in tool folder by

   python tool/create_dataset.py --out lmdb/data/output/path --file path/to/file

3. Use the same step to create train and val data.

Change parameters and alphabets

Parameters and alphabets can't always be the same in different situation.

• Change parameters

  Your can see the params.py in detail.

• Change alphabets

  Please put all the alphabets appeared in your labels to alphabets.py , or the program will throw error during training process.

Train

Run train.py by

python train.py --trainroot path/to/train/dataset --valroot path/to/val/dataset

Reference

meijieru/crnn.pytorch

Sierkinhane/crnn_chinese_characters_rec