Easy-Recognizer

Introduce

This is a framework that is user-friendly for text recognition task.

Feature

• automatic resize all images to the minimum shape in a batch data/dataloader.py:resize_collector
• three level configuration, from default to user-defined: config/default.py -> experiment/*.yaml -> command line args
• tensorboard support for visualize acc/loss curve and inference result
• fast lmdb packing data/scripts/dataset_to_lmdb.py
• lmdb visualize tools with Flask data/scripts/lmdb_vis.py
• reimplement of bezier align and bezier control point fit in ABCNet tools/bezier.py
• default lr scheduler with warmup and cosine learning rate decay lr_scheduler/build_lr_scheduler.py

Usage

git clone ...
cd easy_recognizer
pip install -r requirements.txt
python train.py --config-file experiment/base/crnn.yaml num_gpus 4 dataloader.batch_size 12