SuperPowerful CRNN and SuperLight CRNN_Tiny

1. Refactoring the CRNN architecture based on the reference paper, improving ChineseLicensePlate testset by nearly 7 points compared to v0.3.0;
2. We have achieved a super lightweight CRNN_Tiny, which reduces the model size by at least 14 times and increases inference speed by about 3 times compared to the CRNN model;
3. Fixed the mixed-precision-training bug in train_plate.py.

Model	ARCH	Model Size (MB)	EMNIST Accuracy (%)	Training Data	Testing Data
CRNN	CONV+GRU	31	98.546	100,000	5,000
CRNN_Tiny	CONV+GRU	1.7	98.396	100,000	5,000

Model	ARCH	Model Size (MB)	ChineseLicensePlate Accuracy (%)	Training Data	Testing Data
CRNN	CONV+GRU	58	82.379	269,621	149,002
CRNN_Tiny	CONV+GRU	4	76.222	269,621	149,002

Refactoring CRNN and Dataset Augmentation

1. Rebuilding the CRNN model;
2. Rebuilding the EMNIST and ChineseLicelicense Plate datasets;

Model	ARCH	Model Size (MB)	EMNIST Accuracy (%)	Training Data	Testing Data
CRNN	CONV+LSTM	34	98.432	100,000	5,000
CRNN	CONV+GRU	31	98.386	100,000	5,000

Model	ARCH	Model Size (MB)	ChineseLicensePlate Accuracy (%)	Training Data	Testing Data
CRNN	CONV+LSTM	70	74.252	269,621	149,002
CRNN	CONV+GRU	58	75.649	269,621	149,002

Support License Plate Recognition

• Support training/evaluation/prediction of CRNN+CTC based on license plate;
• Support training/evaluation/prediction of CRNN+CTC based on EMNIST digital characters.

Note: I found that using mixed precision training can cause the gradient explosion of the license plate loss function.

CRNN+CTC

• Support training/evaluation/prediction of CRNN+CTC based on EMNIST digital characters.