lornatang / crnn-pytorch

PyTorch implemnts `An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition` paper.

Hi, my dear friend, I am committed to promoting the development of GAN.

crnn-pytorch's Issues

CER calculation

Hi, thanks for the great work first and foremost!

I noticed that the accuracy calculation is not quite the CER needed. I was wondering if this is just a temporary fix or what does this: "".join(prediction_chars[0]) == target[0].lower() mean?

大佬您好，可以重新给一下谷歌云盘的链接吗，现在的失效了

关于使用自己数据集进行训练出错

使用自己训练集并且将config文件修改为train模式后 使用python train.py 命令出现 如下报错:
Load all datasets successfully.
Build CRNN model successfully.
Define all loss functions successfully.
Define all optimizer functions successfully.
Check whether the pretrained model is restored...

• Acc: 0.00%
  Traceback (most recent call last):
  File "train.py", line 364, in
  main()
  File "train.py", line 85, in main
  acc = validate(model, test_dataloader, epoch, writer, "Valid")
  File "train.py", line 291, in validate
  raise ValueError("Unsupported mode, please use valid or test.")
  ValueError: Unsupported mode, please use valid or test.

Questions on training hyperparameters

Hi Lorna,
i was hoping you could shed light on some of the training hyperparmeter choices. I have tried using the ICDAR2013 dataset which is significantly smaller than MJSynth. For training and validation, i use its train and test set respectively. However the model seems to be converging to outputting complete blanks. I tried making the learning rate to 1e-3 for example and adding a cosine-annealing lr to it. But it doesn't seem to matter. So my questions are (also for when i move onto MJSynth training later):

1. From your experience, where could the problem lie? Is the amount of data from ICDAR2013 too few without data augmentation for this model?
2. I'm not versed in mixed-precision training, but could this be messing with the model's convergence at the start, since the scaler's scale is somewhere around 1e4?
3. What kind of preprocessing are used for characters that are not supported by the model, i.e. "-"? Do you omit samples containing such characters or do we just ignore it in the ground truth labeling?
4. If you can share, what learning rate schedule did you use for your pretrained model?

Thanks for sharing your experience as you have been! Very appreciated! Thanks in advance!

Regards,
Zheng

CRNN-MJSynth-e9341ede.pth.tar predict quite bad

I try some images,find CRNN-MJSynth-e9341ede.pth.tar predict quite bad.