search

xuexingyu24 / License_Plate_Detection_Pytorch

A two stage lightweight and high performance license plate recognition in MTCNN and LPRNet
Other
641 stars 171 forks source link

Issues on training from scratch #2

Closed JimLee1996 closed 5 years ago

JimLee1996 commented 5 years ago

Thanks for your nice work!

I've noticed that the LPRNet is initialized using a pre-trained weights before training on CCPD. Here is the problem:

When I tried to train my model (another model) from scratch on CCPD, the ctc_loss got down to <=0.01 while accuracy remained <=1%.

I wonder is it necessary to apply pre-training on high quality dataset before CCPD?

xuexingyu24 commented 5 years ago

Yes I use pre-trained weight for initialization and that will make training easily reach ~70% accuracy at first few epochs. The training might take longer and slower if you just have some random initializations.

xuexingyu24 commented 5 years ago

Believe I can close this one