Question - regarding pretraining

Open harirajeev opened this issue 6 years ago • 1 comments

Hi,

Thank you for sharing details about your work.

I have this question ?. Why are you initializing using the model pre-trained on MSCOCO ?

thanks Hari

Apr 30 '19 02:04 harirajeev

So I could just use the same image mean and variance as MSCOCO. It converges faster but maybe training the network from scratch will get a better result. After all, document images are quite different than natural images in MSCOCO.

Jul 29 '19 10:07 leonlulu