This repo implements CRNN in pytorch 1.8, official code can be found at https://github.com/bgshih/crnn.
Download the pretrained model at BaiduNetdisk(fetch code: 1234), put it somewhere in your disk, and change CHECKPOINT_FILE in config.py corespondingly. Then run the demo.py by:
$ python demo.py demo/display.png
Raw predicted word: d----ii-s---p---l--a---y--, predicted word: display
I use MJSynth to train and validate, use SVT to test. Use the checkpoint at epoch 19, the accuracy reaches 80.371, compared to 80.8 in original paper.
Down MJSynth and SVT, put them somewhere in your disk, and change all the xxx_DIR in config.py. Then train the model by:
cd data
python create_lmdb.py # It only needs to be executed once
cd ..
python train.py