MichalBusta
commented
6 years ago Hi, we have trained modes from scratch. It is about the speed. (even in cpp implementation pooling of proposals is slow - GPU is just waiting for job).
For the fine tuning OCR network the optimizer state is important. (since provided script creates small batches - expected thing is that it will destroy even "good" model in first few iterations)
if you are interested in simple working solution, just train 2 stand alone networks - the extra augmentation with inperfect proposals give you just 2-5% on ICDAR tasks (good for competition mode, wasting of resources otherwise).
@MichalBusta Hope this message finds you well.
I have been reading your PhD thesis and you have mentioned about pre-training your models using SynthText and Synthetic Word datasets. There's no script for performing that. I can only see training in
train.py. Is it necessary to pre-train further? does it make any difference?Please help.