search

mitmul / deeppose

DeepPose implementation in Chainer
http://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/42237.pdf
GNU General Public License v2.0
408 stars 129 forks source link

Changed __call__ func when model.train is False #24

Closed shiba24 closed 8 years ago

shiba24 commented 8 years ago

Hi, @mitmul thank you for the codes.

I faced an error when I am running train of ResNet50 on my environment: python scripts/train.py \ --model models/ResNet50.py \ --gpu 0 \ --datadir data/FLIC-full \ (....... and other settings)

Error subscription is not saved unfortunately, but occurred in https://github.com/mitmul/deeppose/blob/master/scripts/train.py#L203 . The message said sth like "Nonetype doesn't have .data". This suggests model.data is None temporally.

I think the function "one_epoch" in scripts/train.py DOES need to calculate model.loss in the TEST session. So I changed and now working. Although I didn't face this error when running AlexNet because only it doesn't have function "clear()" like ResNet50, this bug may be also in the other model file? Thanks.

mitmul commented 8 years ago

@shiba24 Oops, Thank you very much. There might still be many bugs in the current code, so I'm thinking I have to spend some time to fix them.