MaskOCR: Text Recognition with Masked Encoder-Decoder Pretraining

General purpose state of the art OCR network based on ViT. This is an unofficial implementation, however I tried to be as close as possible to the original.

Train

I'm currently using Chinese City Parking Dataset for my training since it is freely available.

Test

TODO

Changes from Paper

I've made some changes to make the training faster:

1. Adam from The Road Less Scheduled, with original implementation from https://github.com/facebookresearch/schedule_free
2. One Cycle Policy from Super-Convergence: Very Fast Training of Neural Networks Using Large Learning Rates (implemented on pytorch)
3. Learning rate finder Cyclical Learning Rates for Training Neural Networks through fastai