def grid_placed(images, size=None):
assert len(images.shape) == 4, f'images should be 4D, but get shape {images.shape}'
B, H, W, C = images.shape
if size is None:
size = squarest_grid_size(B)
image_type = images.dtype
if (images.shape[3] in (3,4)):
img = np.zeros((H * size[0], W * size[1], C), dtype=image_type)
for idx, image in enumerate(images):
i = idx % size[1]
j = idx // size[1]
img[j * H:j * H + H, i * W:i * W + W, :] = image
if np.issubdtype(image_type, np.uint8):
return img
if np.min(img) < -.5:
# for images in range [-1, 1], make it in range [0, 1]
img = (img + 1) / 2
elif np.min(img) < 0:
# for images in range [-.5, .5]
img = img + .5
assert np.min(img) >= 0, np.min(img)
assert np.max(img) <= 1, np.max(img)
img = np.clip(255 * img, 0, 255).astype(np.uint8)
return img
elif images.shape[3]==1:
img = np.zeros((H * size[0], W * size[1]), dtype=image_type)
for idx, image in enumerate(images):
i = idx % size[1]
j = idx // size[1]
img[j * H:j * H + H, i * W:i * W + W] = image[:,:,0]
return img
else:
NotImplementedError
def video_summary(name, video, size=None, fps=30, step=None):
name = name if isinstance(name, str) else name.decode('utf-8')
if np.issubdtype(video.dtype, np.floating):
video = np.clip(255 * video, 0, 255).astype(np.uint8)
while len(video.shape) < 5:
video = np.expand_dims(video, 0)
B, T, H, W, C = video.shape
if size is None and B != 1:
bh, bw = squarest_grid_size(B)
frames = video.reshape((bh, bw, T, H, W, C))
frames = frames.transpose((2, 0, 3, 1, 4, 5))
frames = frames.reshape((T, bh*H, bw*W, C))
else:
if size is None:
size = (1, 1)
assert size[0] * size[1] == B, f'Size({size}) does not match the batch dimension({B})'
frames = video.transpose((1, 2, 0, 3, 4)).reshape((T, size[0]*H, size[1]*W, C))
try:
summary = tf1.Summary()
image = tf1.Summary.Image(height=B * H, width=T * W, colorspace=C)
image.encoded_image_string = encode_gif(frames, fps)
summary.value.add(tag=name, image=image)
tf.summary.experimental.write_raw_pb(summary.SerializeToString(), step)
except (IOError, OSError) as e:
print('GIF summaries require ffmpeg in $PATH.', e)
frames = video.transpose((0, 2, 1, 3, 4)).reshape((1, B * H, T * W, C))
tf.summary.image(name + '/image', frames, step)
def save_image(images, path, size=None):
assert images.shape.ndims == 4, f'images should be 4D, but get shape {images.shape}'
num_images = images.shape[0]
if size is None:
size = utils.squarest_grid_size(num_images)
images = grid_placed(images, size)
utils.check_make_dir(path)
imsave(path, images)
def save_image(images, path, size=None):
assert_colorize(
len(images.shape) == 4,
f'images should be 4D, but get shape {images.shape}')
num_images = images.shape[0]
if size is None:
size = utils.squarest_grid_size(num_images)
images = merge(images, size)
utils.check_make_dir(path)
imsave(path, images)
def _log_train_info(self):
num_images = min(self.batch_size, 16)
image_shape = self.image_shape[:-1]
image_grid = lambda vis_images: tfgan.eval.image_grid(
vis_images[:num_images],
grid_shape=squarest_grid_size(num_images),
image_shape=image_shape)
def image_stats(image):
means = tf.reduce_mean(image, 0, keep_dims=True)
vars = tf.reduce_mean(tf.squared_difference(image, means), 0, keep_dims=True)
mean, var = tf.reduce_mean(means), tf.reduce_mean(vars)
return mean, var
if self.log_tensorboard:
with tf.name_scope('train_info'):
tf.summary.histogram('z_', self.generator.z)
with tf.name_scope('loss'):
tf.summary.scalar('generator_loss_', self.gen_loss)
tf.summary.scalar('discriminator_loss_', self.dis_loss)
tf.summary.scalar('loss_', self.gen_loss + self.dis_loss)
with tf.name_scope('image'):
tf.summary.image('generated_image_', image_grid(self.gen_image), max_outputs=1)
tf.summary.histogram('generated_image_hist_', self.gen_image)
gen_mean, gen_var = image_stats(self.gen_image)
real_mean, real_var = image_stats(self.image)
tf.summary.scalar('gen_mean_', gen_mean)
tf.summary.scalar('gen_var_', gen_var)
tf.summary.scalar('real_mean_', real_mean)
tf.summary.scalar('real_var_', real_var)
with tf.name_scope('prob'):
tf.summary.histogram('real_prob_hist_', self.real_discriminator.prob)
tf.summary.histogram('fake_prob_hist_', self.fake_discriminator.prob)
tf.summary.scalar('real_prob_', tf.reduce_mean(self.real_discriminator.prob))
tf.summary.scalar('fake_prob_', tf.reduce_mean(self.fake_discriminator.prob))
with tf.name_scope('logit'):
tf.summary.histogram('real_logits_hist_', self.real_discriminator.logits)
tf.summary.histogram('fake_logits_hist_', self.fake_discriminator.logits)
tf.summary.histogram('real_logits_', tf.reduce_mean(self.real_discriminator.logits))
tf.summary.histogram('fake_logits_', tf.reduce_mean(self.fake_discriminator.logits))
def save_video(name, video, fps=30):
name = name if isinstance(name, str) else name.decode('utf-8')
video = np.array(video, copy=False)
if np.issubdtype(video.dtype, np.floating):
video = np.clip(255 * video, 0, 255).astype(np.uint8)
while len(video.shape) < 5:
video = np.expand_dims(video, 0)
B, T, H, W, C = video.shape
if B != 1:
bh, bw = squarest_grid_size(B)
frames = video.reshape((bh, bw, T, H, W, C))
frames = frames.transpose((2, 0, 3, 1, 4, 5))
frames = frames.reshape((T, bh*H, bw*W, C))
else:
frames = video.transpose((1, 2, 0, 3, 4)).reshape((T, H, W, C))
f1, *frames = [Image.fromarray(f) for f in frames]
if not os.path.isdir('results'):
os.mkdir('results')
path = f'results/{name}.gif'
f1.save(fp=path, format='GIF', append_images=frames,
save_all=True, duration=1000//fps, loop=0)
print(f"video is saved to '{path}'")