GANs-Diversity

Code for Birthday Paradox test for diversity of GANs, proposed in Do GANs actually learn the distribution? An empirical study

based on DCGAN-Tensorflow

Requirement

• Python, Numpy, Tensorflow

Get collision candidates of generated faces from a DCGAN

Download the model checkpoint and unzip it

$ wget https://www.dropbox.com/s/sfdoyvfl5eozfa5/checkpoint.zip
$ unzip checkpoint.zip

Generate a pool of $topK=20$ candidate collision pairs from a batch of 6*64=384

$ python main.py --topK=20 --num_batches=6 --sample_dir=samples

The top 20 potential collisions will be in ./samples

Example code for collision candiate selection

import numpy as np
import heapq as hq
import copy as cp



images = sample_from_gan()	# a batch of generated samples
queue = []	# a priority queue maintaining top K most similar pairs
n_image = images.shape[0]
topK = 20 # keep top

for i in range(n_image):
for j in range(i+1, n_image):
  # measure similarity in pixel space (could be done in some embedding space too)
  dist = np.sum((images[i] - images[j])**2)
  if len(queue) == 0 or -1*dist > queue[0][0]:
    hq.heappush(queue, (-1*cp.deepcopy(dist), cp.deepcopy(images[i]), cp.deepcopy(images[j])))
    if len(queue) > topK:
      hq.heappop(queue)

for idx in range(topK):
	neg_dist, img1, img2 = hq.heappop(queue)
	scipy.misc.imsave(config.sample_dir + '/pair#%d_%f_%d.png'%(idx, -1*neg_dist, 1), (img1+1.)/2)
	scipy.misc.imsave(config.sample_dir + '/pair#%d_%f_%d.png'%(idx, -1*neg_dist, 2), (img2+1.)/2)

To estimate the diversity of a GAN model:

• Sample a batch of size s from the generator
• Use the code above to select topK (say 20) potential collisions
• Visually inspect the flagged pairs and check for duplicates
• Repeat

If this test reveals that samples of size $s$ have duplicate images with good probability (say >= 50%), then suspect that the distribution has support size about $s^2$.