StyleGAN2 - Training Run No.2 - Metrics Update

This markdown contains a running update on the four training metrics computed to keep track of the styleGAN2s convergence. These include FID , PPL-W end, k-NN Precision, and k-NN Recall. Below I present the current plots.

NOTE For good training performance we want FID ↓ PPL-W end↓ (though we should be aware that as training continues it is expected for the perceptual path length to increase as the feature space becomes more entangled), k-NN Precision, and k-NN Recall ↑.

NOTE The red-dashed line represents the epoch where I re-started the training run after an AWS storage error.

All metrics are currently on-track. Below is a list of resources which explain these in greater detail:

• FID: https://nealjean.com/ml/frechet-inception-distance/
• PPL: https://arxiv.org/pdf/1812.04948.pdf
• Precision and Recall : https://arxiv.org/abs/1904.06991