search

jhb86253817 / PIPNet

Efficient facial landmark detector
MIT License
411 stars 82 forks source link

Problem about UDA training with curriculum #8

Closed backkon closed 3 years ago

backkon commented 3 years ago

when I tried to train with unlabeled data, I got results like this: image

I followed your guide in README.md and run python lib/train_gssl.py experiments/data_300W_COFW_WFLW/pip_32_16_60_r18_l2_l1_10_1_nb10_wcc.py

I had no idea what the problem was.

backkon commented 3 years ago

After each iteration, the NME is almost constant, as shown in the figure above

backkon commented 3 years ago

log.TXT This is the output file. Look forward to your reply.

jhb86253817 commented 3 years ago

Hi, it seems to be the problem during inference, not training. I have modified line 223 and 228 of _functionsgssl.py. You may try again.

backkon commented 3 years ago

Thanks a lot!