All humans are fake, atleast the ones generated by this GAN!
Humans-Are-Fake is a Deep Convolutional Generative Adversarial Network (DCGAN) to generate human-like faces which was trained for over 8 hours using 25,000 images on NVIDIA Tesla P100 GPU. In the stable model, losses for the Discriminator Network were less than 0.018 while losses for the Generator Network were less than 6.282.
View notebook on Kaggle
- Both, the generator models and the discriminator models were stored along with their indexes after every 5 epoch of training.
- Unzip the zip file containing all the checkpoints.
- Download and use the stable version (ckpt-6).
- Overall GAN workflow -
- Generator Network Architecture -
- Discriminator Network Architecture -
After training for around 35 Epochs, the GAN produced the results listed below. The following pictures were created by the GAN entirely out of its own imagination; they were never observed by it during its training phase.
View dataset (Credits: GREATGAMEDOTA)
If you have any feedback or suggestions please reach out to the project admin sanidhyak or you can create an issue and mention there which new features can be added to make Humans-Are-Fake better.