search

clovaai / deep-text-recognition-benchmark

Text recognition (optical character recognition) with deep learning methods, ICCV 2019
Apache License 2.0
3.77k stars 1.11k forks source link

Suggestions for Fine-Tuning? #270

Open PascalHbr opened 3 years ago

PascalHbr commented 3 years ago

I want to fine-tune the model on my own dataset. The default lr is set to 1., which seems to be quite high. What lr is suggested for fine-tuning?

Moreover, is it recommended to fine-tune on cut out single digits/characters or on complete images?

My last question is about the input parameters concerning the image width/height. The default parameters are 100/32. Is it recommended to adjust this during the fine-tuning step? My fine-tuning dataset is of size 64x64, should i adjust the parameters to 64x64 accordingly?

Thanks a lot in advance!

henryle97 commented 3 years ago
  1. Learning rate for fine tuning can be 1e-4
  2. None
  3. If you use pretrained model, you must use image size 32x32. If training from scratch, you can use 64x64
mjack3 commented 3 years ago

@hisiter97 I think he can fine tunning with the widthxheight that he desire, no matter from scratch or fine tunning.