NME 0.07269 vs. 0.05837

[schmidtp1]

Thanks for sharing the code! With the code as-is I'm getting an NME of 0.07269 with PFLD at width 1 at 200 epochs with 112x112x3 input vs. 0.05837 reported in the results table (https://github.com/AnthonyF333/FaceLandmark_PFLD_UltraLight#%E7%B2%BE%E5%BA%A6). Would you have any idea why? I attached the train. log. train.log

[AnthonyF333]

@schmidtp1 Maybe you should do data augmentation for WFLW dataset like rotate, resize, flip or translation. I do 80 times augmentation for every image in WFLW.

[schmidtp1]

Thanks, I'll try. I only did the default 10 augmentations as described in the readme.