def generate_diversity_samples(self, sess, filename, X_corrupted, X_true, horz_samples=10, vert_samples=10):
from data.dataset import noise
print('Generating samples for', filename)
predictions = [X_true[:vert_samples,:,:,:].reshape(1, vert_samples, self.height, self.width), X_corrupted[:vert_samples,:,:,:].reshape(1, vert_samples, self.height, self.width)]
X_corrupted = np.repeat(X_corrupted[:vert_samples,:,:,:], horz_samples, axis=0)
X_true = np.repeat(X_true[:vert_samples,:,:,:], horz_samples, axis=0)
start = time.perf_counter()
for i in range(self.test_iterations // 2 if filename == 'denoise' else self.test_iterations):
X_corrupted = self.update_clamp(X_corrupted, sess.run(self.predictions, feed_dict={self.X_in:X_corrupted}), filename)
if self.markov_type == 'renoise_per_two':
X_corrupted = self.update_clamp(X_corrupted, sess.run(self.predictions, feed_dict={self.X_in:X_corrupted}), filename)
if self.markov_type != 'only_denoise':
if filename == 'denoise':
noise_prop = ((self.test_iterations//2 - i - 1) / self.test_iterations)
else:
noise_prop = (self.test_iterations - i - 1) / self.test_iterations
X_corrupted = self.update_clamp(X_corrupted, np.array([noise(arr, noise_prop, noise_prop)[0] for arr in X_corrupted]), filename)
print(time.perf_counter() - start)
predictions.append(X_corrupted.reshape(vert_samples, horz_samples, self.height, self.width).transpose(1,0,2,3))
images = np.concatenate(tuple(predictions), axis=0)
images = images.transpose(1, 2, 0, 3)
images = images.reshape((self.height * vert_samples, self.width * (horz_samples + 2)))
filename = datetime.now().strftime('samples/%Y_%m_%d_%H_%M_noncausal_diversity_' + filename)+".png"
Image.fromarray((images*255).astype(np.int32)).convert('RGB').save(filename)
def generate_one_group_of_samples(self, sess, filename, X_corrupted, X_true=None, horz_samples=10):
from data.dataset import noise
print('Generating samples for', filename)
X_corrupted = X_corrupted[:horz_samples,:,:,:]
predictions = [X_true[:horz_samples,:,:,:]] if X_true is not None else []
predictions.append(X_corrupted)
start = time.perf_counter()
for i in range(self.test_iterations // 2 if filename == 'denoise' else self.test_iterations):
X_corrupted = self.update_clamp(X_corrupted, sess.run(self.predictions, feed_dict={self.X_in:X_corrupted}), filename)
if self.markov_type == 'renoise_per_two':
X_corrupted = self.update_clamp(X_corrupted, sess.run(self.predictions, feed_dict={self.X_in:X_corrupted}), filename)
predictions.append(X_corrupted)
if self.markov_type != 'only_denoise':
if filename == 'denoise':
noise_prop = ((self.test_iterations//2 - i - 1) / self.test_iterations)
else:
noise_prop = (self.test_iterations - i - 1) / self.test_iterations
X_corrupted = self.update_clamp(X_corrupted, np.array([noise(arr, noise_prop, noise_prop)[0] for arr in X_corrupted]), filename)
print(time.perf_counter() - start)
predictions = tuple(p.reshape((horz_samples, 1, self.height, self.width)) for p in predictions)
images = np.concatenate(predictions, axis=1)
images = images.transpose(1, 2, 0, 3)
images = images.reshape((self.height * ((self.test_iterations//2 if filename == 'denoise' else self.test_iterations) + 1 + (1 if X_true is not None else 0)), self.width * horz_samples))
filename = datetime.now().strftime('samples/%Y_%m_%d_%H_%M_noncausal_' + filename)+".png"
Image.fromarray((images*255).astype(np.int32)).convert('RGB').save(filename)
def generate_ten_thousand_samples(self, sess):
if self.denoising:
from data.dataset import noise
print('Generating 10000 samples for denoising')
start = time.perf_counter()
samples = []
test_data = self.data.get_noise_values()
for j in range(self.data.total_test_batches):
print('batch', j)
X_corrupted = sess.run(test_data)
for i in range(self.test_iterations):
X_corrupted = sess.run(self.predictions,
feed_dict={self.X_in: X_corrupted})
if self.markov_type == 'renoise_per_two':
X_corrupted = sess.run(
self.predictions,
feed_dict={self.X_in: X_corrupted})
if self.markov_type != 'only_denoise':
noise_prop = (self.test_iterations - i -
1) / self.test_iterations
X_corrupted = np.array([
noise(arr, noise_prop, noise_prop)[0]
for arr in X_corrupted
])
samples.append(X_corrupted)
samples = np.concatenate(samples)
print(time.perf_counter() - start)
return samples
else:
print('Generating 10000 samples for pixelcnn')
start = time.perf_counter()
sampless = []
for _ in range(10000 // 100):
samples = np.zeros(
(100, self.height, self.width, self.channels),
dtype=np.float32)
for i in range(self.height):
for j in range(self.width):
#print(i,j)
predictions = sess.run(self.predictions,
feed_dict={self.X_in: samples})
samples[:, i, j, :] = predictions[:, i, j, :]
sampless.append(samples)
samples = np.concatenate(sampless)
print(time.perf_counter() - start)
return samples
def generate_ten_thousand_samples(self, sess):
from data.dataset import noise
print('Generating 10000 samples for noncausal')
start = time.perf_counter()
samples = []
test_data = self.data.get_noise_values()
for j in range(self.data.total_test_batches):
print('batch', j)
X_corrupted = sess.run(test_data)
for i in range(self.test_iterations):
X_corrupted = sess.run(self.predictions, feed_dict={self.X_in:X_corrupted})
if self.markov_type == 'renoise_per_two':
X_corrupted = sess.run(self.predictions, feed_dict={self.X_in:X_corrupted})
if self.markov_type != 'only_denoise':
noise_prop = (self.test_iterations - i - 1) / self.test_iterations
X_corrupted = np.array([noise(arr, noise_prop, noise_prop)[0] for arr in X_corrupted])
samples.append(X_corrupted)
samples = np.concatenate(samples)
print(time.perf_counter() - start)
return samples
def generate_grid_of_samples(self, sess, X_corrupted):
from data.dataset import noise
print('Generating samples for grid')
horz_samples = 10
vert_samples = 10
X_corrupted = X_corrupted[:horz_samples*vert_samples,:,:,:]
start = time.perf_counter()
for i in range(self.test_iterations):
X_corrupted = sess.run(self.predictions, feed_dict={self.X_in:X_corrupted})
if self.markov_type == 'renoise_per_two':
X_corrupted = sess.run(self.predictions, feed_dict={self.X_in:X_corrupted})
if self.markov_type != 'only_denoise':
noise_prop = (self.test_iterations - i - 1) / self.test_iterations
X_corrupted = np.array([noise(arr, noise_prop, noise_prop)[0] for arr in X_corrupted])
print(time.perf_counter() - start)
images = X_corrupted.reshape((vert_samples, horz_samples, self.height, self.width))
images = images.transpose(0, 2, 1, 3)
images = images.reshape((self.height * vert_samples, self.width * horz_samples))
filename = datetime.now().strftime('samples/%Y_%m_%d_%H_%M_noncausal_grid')+".png"
Image.fromarray((images*255).astype(np.int32)).convert('RGB').save(filename)