This repository contains the setup for all experiments performed in our Paper ... It is to be used in conjunction with the RL environment text-localization-environment, which is linked as a submodule. After cloning do git submodule init and git submodule update and follow the installation instructions of that repo.

The project is configured using Hydra in the cfg folder.

We use RLLib as RL framework. Train the model by executing rllib_train.py.

Every value in the cfg folder can be altered by passing it as a CLI argument, while keeping the correct file hierarchy (e.g. data.path=/data). The folder data contains templates for different dataset configurations.

Here are explanations for a few example parameters.

Training weakly supervised:

Loading a checkpoint:

Checkpoints need to be RLLib checkpoint folders. Our best three models (supervised, weakly supervised and semi-supervised) can be downloaded here.

Set the parameter restore to the checkpoint directory. Training will resume from the checkpoint. The training iterations have to be increased, as the checkpoints were made at iteration 15k.

Execute evaluate.py.

python evaluate.py <checkpoint_dir> <eval_data_dir> <eval_gt_file> --dataset icdar2013 [--framestacking grayscale]

For IDE debugging change ray.init() in rllib_train.py to ray.init(local_mode=True).