In this repository, we provide a Generative Model to adress the Face Aging problematic. We want to translate the input image to the desired class (10-19, 20-29, 30-39, 40-49, 50+) of age given as input too, the firgures bellow showcase what we want.
To do so, we train two networks in a competitive manner based on the Generative Adversarial Nets framework presented by Ian Goodfellow in 2014. The Generator (G) and the Discriminator (D) gets to see the same input, the difference is in the output. G gives an 256x256 image while D outputs a score corresponding to the probability of the image of being real.
In order to keep the same identity we add an Identity preservation module. We use a Face-Recognition model to get an embedding vector of the input and output images, we thus compute the cosine-similarity between the two and penalize the G network accordingly.
The same idea is used for the age, i.e we penalize the network when the age classifier (network) outputs a class that is different from the desired one. In fact we train on 5 classes of ages:
- 0: Ages from 10-19
- 1: Ages from 20-29
- 2: Ages from 30-39
- 3: Ages from 40-49
- 4: Ages 50+
We also made use of the perceptual loss that is computed with a pretrained VGG_19 and finaly an L2 loss.
P.S: feel free to comment and contribute to this modest work ;-)
Implement age lossImplement DataParralelUse multithreads with dataloader- Train a face age classifier (on CACD dataset?) and use it as pretrained in this project.