This repo is a demo for how to use warpctc in PyTorch to recognize captcha in images.
Baidu's warpctc project is a famous library to compute CTC loss. It's writen by C++11, which makes it easy to bind warpctc with other deep learning framework, such as MXNet and PyTorch.
There is a warpctc binding for Caffe. Here, we use warpctc binding for PyTorch.
Please refer to this page: deepspeech.pytorch to install WarpCTC binding for PyTorch first.
The python package captcha is used to generate dataset for training and testing.
You can use python generate_captcha.py to generate the captcha images. The default data folder is $PROJECT_DIR/data. If you want to specific another one, you need to modify the variable DATASET_PATH in utils.py.
The neural network is a stacked LSTM. See model.py for detail.
To train the model, run the following command:
python main.py
If tensorboard has been installed, a logging event file will be generated in the folder $PROJECT_DIR/log, which can be used to visualize the training process using tensorboard command of TensorFlow.
The testing accuracy may be weird because a global step is used, which leads to a larger logging interval for testing, comparing with training.
During the training process, it will save checkpoint every 10 epochs. The default folder to save the checkpoints is $PROJECT_DIR/pretrained
When the training is done, you can check the trained model using this command:
python test.py test_image_path trained_model_path
The output are two figures, showing the probabilities of different digits vs column.
