CRNN

An implementation of CRNN algorithm using Pytorch framework

The most typical CTC algorithm is CRNN (Convolutional Recurrent Neural Network), which introduces the bidirectional LSTM (Long Short-Term Memory) to enhance the context modeling. Experiments have proved that the bidirectional LSTM module can effectively extract the contextual information of the picture, finally enter the output feature sequence into the CTC module, and decode the sequence result.

Environment The dependencies are listed in `requirements.txt`. Please install and follow the command below:

pip install -r requirements.txt

Data Preparation

You can download and use ICDAR2015 or MJSynth dataset for the training model.

After that,

• Put the data folder under the datasets directory
• Setup config file for each dataset in src/__init__.py: replace CFG_PATH augment by src/config/rec_lmdb.yml or src/config/rec_icdar15.yml

Training

Before training, please modify configurations in src/config/rec_lmdb.yml or src/config/rec_icdar15.yml

python -m src.train

To train the model with PyTorch lightning

python -m src.pl_modules.pl_train

Evaluation

python -m src.eval

Export model

python -m src.exports.exporter --export_format <torchscript/onnx/paddle/tensorrt> --device cuda

Prediction

python -m src.predict --img_path <path_to_image_file> --export_format <torchscript/onnx/paddle/tensorrt> --device cuda

Example:

python -m src.predict --image_path images/tuongan.png --export_format pt --device cuda
>> tuongan

Tensorboard

tensorboard --logdir debugs/tensorboard

Train Loss	Valid Loss	Valid Acc

Experiments

Export format	Inference time	ACC with CTCLoss	ACC with Focal CTCLoss
Pytorch	0.0008s	95.47	96
TorchScript	0.0007s	-	-
ONNX	0.001s	-	-
TensorRT	0.0001s	-	-
Paddle	0.01s	-	-