Tensorflow implementation of Discriminator Rejection Sampling
SAGAN with hinge loss is used as a base model. The network is first trained for 50 epochs, and as it's descriped in the paper, the model is further trained using a smaller learning rate (5 epochs). The model uses hinge loss and a sigmoid output is requried to perform DRS. Therefore fc layers are added on top of the discriminator and is further trained using a cross entropy loss.
The rejection sampling algorithm is implemented in the sample.py file.
Once finished training, about 50k samples are generated to estimate M (BurnIn phase).
After estimating M, new samples are generated and their F_hat values are calculated.
For CelebA dataset, 80th percentile value is used for gamma (80th percentile value of F is used as gamma).
- numpy
- opencv2
- Tensorflow
