This project is my diploma work,realizing attribute-based human face generation using GANs.
In order to train an attribute-driven generative model, I used a method inspired by auxiliary classifier generative adversarial networks. By using ACGAN, the whole network can be trained with labeled human face images. Compared with classic conditional generative adversarial network, this model will not get stuck when the number of different labels increases because of the individual feature prediction.
There are some other tricks to improve the generative model. At first,instead of transposed concolutional layer,I used upsampling layer and convolutional layer in turns in the generator to avoid checkerboard pattern of artifacts. Secondly, I used interpolation trick inspired by Reed el.\cite{reed} to get more feature vector for the generator.This method can help avoid model collapse.
> $ git clone git@github.com:cjqw/human-face-generation.git > $ python3 main.py
Now you have to modify the python config.py to train the model.First set config[“train”]=True and set the data path.Other hyperparemeters also can be changed. I will realize more advanced interaction method as soon as possible.
- about the algorithm: algorithm.org
- about the experiments: experiment.org
- References: reference.bib