inception_score_numpy.py

# Inception Score Calculator
#
# A Brock, 2017
#
# This snippet assumes you have two functions defined:
# 1. sample_net, which takes in a batch x num_latents random vector and returns batch samples,
# 2. eval_net, which takes in batch samples and returns a batch x #classes prediction vector. 

num_latents = 100

# Get predicted class given z, p(c|z), for 50,000 samples
p_C_Z = []
for i in range(1000):
  p_C_Z.append(eval_net(sample_net(np.float32(np.random.randn(50,num_latents))))) # Append score to p_C_Z
p_C_Z = np.asarray(p_C_Z)

# Now take ten chunks of p(c|z) and get their scores
score = []
for i in range(10):
    p = p_C_Z[(i * p_C_Z.shape[0] // 10):((i + 1) * p_C_Z.shape[0] // 10), :]
    kl = p * (np.log(p) - np.log(np.expand_dims(np.mean(p, 0), 0)))
    kl = np.mean(np.mean(kl, 1))
    score.append(np.exp(kl))

# Calculate mean and std across ten chunks, and report
incept = np.mean(np.asarray(score))
incept_std = np.std(np.asarray(score))
print('Inception score is ' +str(incept)+'+/-'+str(incept_std))