Evaluate sample qualities of GANs in browser
This is the demo of evaluation metric from “ Quantitatively Evaluating GANs with Divergences proposed for Training”. This paper shows how to measure the similarity between the data distribution and model distribution by constructing a critic network based on certain divergence or distance metrics used in training GANs, e.g., least square divergence, Jenson-Shannon divergence, Wasserstein distance, etc. Therefore the evaluation process does not need external model and data labels.
Instructions:
Click " Show Image" to inference samples from data distribution and model distribution.
Click " Evaluate" on each card to evaluate sample qualities.
Real data distribution, as a baseline, has the best quality.
