This is the Text Scanner paper implementation code.

About the paper unstanding, please read about it at my blog : TextScanner的一些研究

implementation list(ongoing...)

implement the network
implement the mutual-supervision mechanism
implement loss function
create the character annotation GT, and prepare none character level GT
implement train code
implement evaluation code
train the model

developing logs

2020.4.24 create the project and implement the skeleton of the project
2020.4.30 implement the network code, and finish the GT generator and loss function
2020.5.12 the network works now after hundreds of trouble-shootings,TF2.0/tk.keras is full of pit
2020.6.03 make a new branch to solave the OOM issue

Branches

b_troubleshooting_OOM: This try to fix the GPU OOM issue, Currently I work on this branch mainly.<------
b_wordform_in_model: Implement word formulation as internal layer of model, this branch did not consider the OOM issue(reduce the charsets size), focus on the elegance of code implementation.
b_multiple_gpus_train: implements the multiple gpus training

implement details

Developing detail can be tracked by my textscanner implementation issues.