search

Maclory / Deep-Iterative-Collaboration

Pytorch implementation of Deep Face Super-Resolution with Iterative Collaboration between Attentive Recovery and Landmark Estimation (CVPR 2020)
MIT License
294 stars 63 forks source link

About pretrained HourGlass model #7

Closed qibao77 closed 4 years ago

qibao77 commented 4 years ago

How to pretrain the HourGlass model? Can you provide more details? Thank you!

Steve-Tod commented 4 years ago

Hi, we pretrain the HourGlass on CelebA dataset. For training details, we follow this git repo and represent the 68 landmarks as 68-channel heatmap. Since we are using a Feedback HourGlass, we take HR Image and feedback feature of previous step as input during each recurrent step. We aggregate L2 losses between output heatmap and ground truth heatmap of all the steps and optimize the network.

qibao77 commented 4 years ago

Thank you for your reply! What's more, is the ground truth heatmap obtained by OpenFace?

Steve-Tod commented 4 years ago

Yes. We use OpenFace to get ground truth landmarks and generate the heatmaps accordingly.

JingzheLyp commented 4 years ago

Hi, I've trained the Feedback HourGlass model on CelebA according to your method, but the loss is very large about 1e+24, the init means is 'kaiming'. When I change the init means from 'kaiming' to 'normal', the loss decrease from 1e-1 to 1e-4, about 1.1W steps ,then it become to nan. @Steve-Tod , can you provide more details? such like net config, init and learning strategy, Thank you!