def get_inception_score(test_data):
test_data = np.reshape(test_data, (-1, 28*28))
test_data = scale_value(test_data, [-0.5, 0.5])
batch_size = 100
test_size = test_data.shape[0]
total_batch = int(test_size / batch_size)
preds = []
for i in range(total_batch):
offset = (i * batch_size) % (test_size)
batch_xs = test_data[offset:(offset + batch_size), :]
y_final = sess.run(y, feed_dict={x: batch_xs, is_training: False})
pred_softmax = softmax(y_final)
preds.append(pred_softmax)
preds = np.concatenate(preds, 0)
scores = []
splits = 10
for i in range(splits):
part = preds[(i * preds.shape[0] // splits):((i + 1) * preds.shape[0] // splits), :]
kl = part * (np.log(part) - np.log(np.expand_dims(np.mean(part, 0), 0)))
kl = np.mean(np.sum(kl, 1))
scores.append(np.exp(kl))
icp = (np.mean(scores) , np.std(scores))
return icp
def compute_inception_score(n):
all_samples = []
for i in range(int(n / 100)):
all_samples.append(session.run(samples_100))
all_samples = np.concatenate(all_samples, axis=0)
all_samples = all_samples.reshape((-1, 3, 32, 32))
all_samples = scale_value(all_samples, [-1.0, 1.0])
print(all_samples.shape)
return get_inception_score(all_samples)
def compute_metric(generator_model):
global best_icp
sample_size = 20000
noise = np.random.normal(size=(sample_size, 100))
art_images = generator_model.predict(noise)
art_images = scale_value(art_images, [-1.0, 1.0])
art_images = np.transpose(art_images, (0, 3, 1, 2))
(icp_mean, icp_std) = get_inception_score(art_images)
if icp_mean > best_icp: best_icp = icp_mean
print('Inception score: ', icp_mean)
def generate_images(generator, image_path, epoch, cmap='gray'):
xsamples = generator()
xsamples = scale_value(xsamples,
[0, 1]) # convert tanh output to [0, 1] for display
image_path = image_path.format(epoch)
write_image_grid(image_path, xsamples, cmap=cmap)
def cifar10_process(x):
x = x.astype(np.float32)
x = scale_value(x, [-1, 1]) #rescale to tanh compatible output
return x
def mnist_process(x):
x = np.reshape(x, newshape=(-1, 28, 28, 1))
x = x.astype(np.float32)
x = scale_value(x, [-1.0, 1.0]) #rescale to [-1, 1] for compatible with tanh output
return x
def generator_sampler():
images = dim_ordering_unfix(generator.predict(zsamples)).transpose(
(0, 2, 3, 1))
images = scale_value(
images, [0.0, 1.0]) #rescale tanh output to [0, 1] for display
return images.reshape((10, 10, 28, 28))
def generator_sampler():
xpred = dim_ordering_unfix(generator.predict(zsamples)).transpose((0, 2, 3, 1))
xpred = scale_value(xpred, [0.0, 1.0])
return xpred.reshape((10, 10) + xpred.shape[1:])
