Clone CPM and checkpoints from https://github.com/VinAIResearch/CPM
1. Using Makeup Transfer model to transform non-makeup to makeup image.Then use these 2 images to do data de-makeup (input: photo makeup and the output is an image without makeup). (train=1000, val=100, test=100)
2. Build model : Unet
Unet architecture
- Pretrained Netwrok for Encoder (Resnet 50, ...)
- Decoder
3. Loss function: Perceptual Loss
Makeup & non-makeup : Link
De-makeup data : Link
https://www.facebook.com/aivietnam.edu.vn/