deep-text-recognition-benchmark-mnn-ncnn

Brief

• Rewrite Version of clovaai/deep-text-recognition-benchmark. Make attempts of deployment on edge devices.
• Based on Pytorch(training), MNN(Inference) and NCNN(Inference)

Training

• Based on Pytorch1.3, Python3.6
• Original repo please check clovaai/deep-text-recognition-benchmark.
• Default training the TPS-ResNet-BiLSTM-Attn model.

1. Prepare datasets

• Please check the "When you need to train on your own dataset or Non-Latin language datasets." in the README.md of the original repo.
• Once when you prepare the datasets well, there should be "data.mdb" and "lock.mdb" under the output folder.
• Assign the value of "train_data" and "valid_data" with the absolute addresses of the output folders.

2. Start to train

• Modify the values in train.py:

Parameters	Position	Remarks
train_data	train.py, Line:223	lmdb trainset folder
valid_data	train.py, Line:224	lmdb valset folder
batch_max_length	train.py, Line:246	width of the resized image
imgH	train.py, Line:247	height of the resized image
imgW	train.py, Line:248	width of the resized image
character	train.py, Line:250	all your characters of your vocabulary you want to recognize

• IMPORTANT:If there're special characters in your vocabulary(e.g."-"), please modify the dataset.py Line:170.

• The output weights (end with .pth) are saved in the ./saved_models by default.

• Then run python train.py.

Visualization

• You can use netron to open your *.pth to view the model architecture.

Inference

• Please check the README.md under the ./inference.

Ideas & Issues

• Be free to open issues or pull requests.

TODOs

Not Finished yet!!!