search

endernewton / tf-faster-rcnn

Tensorflow Faster RCNN for Object Detection
https://arxiv.org/pdf/1702.02138.pdf
MIT License
3.65k stars 1.58k forks source link

Total loss trained with 4 channels png file falls to NaN. Anyone has idea? #394

Open f27ny105t5123 opened 5 years ago

f27ny105t5123 commented 5 years ago

I concatenate a mask image to a normal RGB picture making .jpg file become .png file with alpha channel.

And I do make sure the 4 channels image loaded properly. But the total loss is always nan while the rpn_cls rpn_box loss_cls and loss_box is an ordinary float as usual.

Anyone has idea about this problem? Thanks~

f27ny105t5123 commented 5 years ago

PS: I don't know much about the channels processing in training, so maybe I left some problem need to fix in training process.