How long do I have to train?

[mowkee]

I've been training for almost a week my solverstate files are yolo_mobile_iter_221000.solverstate and recog_iter_112196.solverstate respectively but it detects nothing. Do I have to train more or something is missing? The training is done on SynthText dataset. I know it's already trained on this dataset but I just wanted to retrace the steps to learn. any comments, suggestions or tips are welcome. If any more info is required please don't hesitate to ask. great work by the way @MichalBusta .

[MichalBusta]

there are quite a lot of meta-parameters which you can play with:

• spatial size of batch images: see line 35 in train.py (the bigger size you can feed = better) - current value is just test settings able to run on laptop computer
• batch size
• optimizer policy (the step works best so far)

if spatial size of batch images is too small, almost all of the regions will be marked as do-not-care since they are partially out of image

batch size should have reasonable size ~ at least 8

[MichalBusta]

In one day of training localization branch, you should reach ~ 60% of recall