search

chenhsuanlin / spatial-transformer-GAN

ST-GAN: Spatial Transformer Generative Adversarial Networks for Image Compositing :eyeglasses: (CVPR 2018)
MIT License
334 stars 72 forks source link

evaluate model #16

Closed betterze closed 5 years ago

betterze commented 5 years ago

Dear chenhsuanlin,

According to the paper, the network warp the foreground sequentially, warp1->warp2-warp3->warp4. However, the model you upload just contains the last warp. And in eval_STGAN.py, it also just uses the last model.

Should I use all warps or just the last one?

Thank you very much for your help.

Best Wishes,

Alex

chenhsuanlin commented 5 years ago

All the warps are included during graph construction, and all the pretrained weights are saved. Only the last warp is returned, but you can probe any intermediate warps if you want to.

betterze commented 5 years ago

thx a lot