def predict(self, image, description):
self.args.description = description
self.args.image_path = image
inverter = StyleGANInverter(
self.args.model_name,
mode=self.args.mode,
learning_rate=self.args.learning_rate,
iteration=self.args.num_iterations,
reconstruction_loss_weight=1.0,
perceptual_loss_weight=self.args.loss_weight_feat,
regularization_loss_weight=self.args.loss_weight_enc,
clip_loss_weight=self.args.loss_weight_clip,
description=self.args.description,
logger=None,
)
image_size = inverter.G.resolution
# Invert the given image.
image = resize_image(
load_image(str(self.args.image_path)), (image_size, image_size)
)
_, viz_results = inverter.easy_invert(image, num_viz=self.args.num_results)
out_path = Path(tempfile.mkdtemp()) / "out.png"
save_image(str(out_path), viz_results[-1])
return out_path
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
assert os.path.isfile(args.image_path)
output_dir = args.output_dir or f'results/inversion/test'
if not os.path.exists(output_dir):
os.makedirs(output_dir)
inverter = StyleGANInverter(
args.model_name,
mode=args.mode,
learning_rate=args.learning_rate,
iteration=args.num_iterations,
reconstruction_loss_weight=1.0,
perceptual_loss_weight=args.loss_weight_feat,
regularization_loss_weight=args.loss_weight_enc,
clip_loss_weight=args.loss_weight_clip,
description=args.description,
logger=None)
image_size = inverter.G.resolution
# Invert the given image.
image = resize_image(load_image(args.image_path), (image_size, image_size))
_, viz_results = inverter.easy_invert(image, num_viz=args.num_results)
if args.mode == 'man':
image_name = os.path.splitext(os.path.basename(args.image_path))[0]
else:
image_name = 'gen'
save_image(f'{output_dir}/{image_name}_enc.png', viz_results[1])
save_image(f'{output_dir}/{image_name}_inv.png', viz_results[-1])
print(f'save {image_name} in {output_dir}')
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
assert os.path.exists(args.image_list)
image_list_name = os.path.splitext(os.path.basename(args.image_list))[0]
output_dir = args.output_dir or f'results/inversion/{image_list_name}'
logger = setup_logger(output_dir, 'inversion.log', 'inversion_logger')
logger.info(f'Loading model.')
inverter = StyleGANInverter(
args.model_name,
learning_rate=args.learning_rate,
iteration=args.num_iterations,
reconstruction_loss_weight=1.0,
perceptual_loss_weight=args.loss_weight_feat,
regularization_loss_weight=args.loss_weight_enc,
logger=logger)
image_size = inverter.G.resolution
# Load image list.
logger.info(f'Loading image list.')
image_list = []
with open(args.image_list, 'r') as f:
for line in f:
image_list.append(line.strip())
# Initialize visualizer.
save_interval = args.num_iterations // args.num_results
headers = ['Name', 'Original Image', 'Encoder Output']
for step in range(1, args.num_iterations + 1):
if step == args.num_iterations or step % save_interval == 0:
headers.append(f'Step {step:06d}')
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(
num_rows=len(image_list), num_cols=len(headers), viz_size=viz_size)
visualizer.set_headers(headers)
# Invert images.
logger.info(f'Start inversion.')
latent_codes = []
for img_idx in tqdm(range(len(image_list)), leave=False):
image_path = image_list[img_idx]
image_name = os.path.splitext(os.path.basename(image_path))[0]
image = resize_image(load_image(image_path), (image_size, image_size))
code, viz_results = inverter.easy_invert(image, num_viz=args.num_results)
latent_codes.append(code)
save_image(f'{output_dir}/{image_name}_ori.png', image)
save_image(f'{output_dir}/{image_name}_enc.png', viz_results[1])
save_image(f'{output_dir}/{image_name}_inv.png', viz_results[-1])
visualizer.set_cell(img_idx, 0, text=image_name)
visualizer.set_cell(img_idx, 1, image=image)
for viz_idx, viz_img in enumerate(viz_results[1:]):
visualizer.set_cell(img_idx, viz_idx + 2, image=viz_img)
# Save results.
os.system(f'cp {args.image_list} {output_dir}/image_list.txt')
np.save(f'{output_dir}/inverted_codes.npy',
np.concatenate(latent_codes, axis=0))
visualizer.save(f'{output_dir}/inversion.html')
def load_test_images(dir, image_size = 128):
img_paths = get_test_images_paths(dir)
imgs = []
for i, image_path in enumerate(img_paths):
IMG = resize_image(load_image(image_path), (image_size, image_size))
imgs.append(IMG)
return imgs
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
src_dir = args.src_dir
src_dir_name = os.path.basename(src_dir.rstrip('/'))
assert os.path.exists(src_dir)
assert os.path.exists(f'{src_dir}/image_list.txt')
assert os.path.exists(f'{src_dir}/inverted_codes.npy')
dst_dir = args.dst_dir
dst_dir_name = os.path.basename(dst_dir.rstrip('/'))
assert os.path.exists(dst_dir)
assert os.path.exists(f'{dst_dir}/image_list.txt')
assert os.path.exists(f'{dst_dir}/inverted_codes.npy')
output_dir = args.output_dir or 'results/interpolation'
job_name = f'{src_dir_name}_TO_{dst_dir_name}'
logger = setup_logger(output_dir, f'{job_name}.log', f'{job_name}_logger')
# Load model.
logger.info(f'Loading generator.')
tflib.init_tf({'rnd.np_random_seed': 1000})
with open(args.model_path, 'rb') as f:
_, _, _, Gs = pickle.load(f)
# Build graph.
logger.info(f'Building graph.')
sess = tf.get_default_session()
num_layers, latent_dim = Gs.components.synthesis.input_shape[1:3]
wp = tf.placeholder(tf.float32, [args.batch_size, num_layers, latent_dim],
name='latent_code')
x = Gs.components.synthesis.get_output_for(wp, randomize_noise=False)
# Load image and codes.
logger.info(f'Loading images and corresponding inverted latent codes.')
src_list = []
with open(f'{src_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{src_dir}/{name}_ori.png')
src_list.append(name)
src_codes = np.load(f'{src_dir}/inverted_codes.npy')
assert src_codes.shape[0] == len(src_list)
num_src = src_codes.shape[0]
dst_list = []
with open(f'{dst_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{dst_dir}/{name}_ori.png')
dst_list.append(name)
dst_codes = np.load(f'{dst_dir}/inverted_codes.npy')
assert dst_codes.shape[0] == len(dst_list)
num_dst = dst_codes.shape[0]
# Interpolate images.
logger.info(f'Start interpolation.')
step = args.step + 2
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(num_rows=num_src * num_dst,
num_cols=step + 2,
viz_size=viz_size)
visualizer.set_headers(['Source', 'Source Inversion'] +
[f'Step {i:02d}' for i in range(1, step - 1)] +
['Target Inversion', 'Target'])
inputs = np.zeros((args.batch_size, num_layers, latent_dim), np.float32)
for src_idx in tqdm(range(num_src), leave=False):
src_code = src_codes[src_idx:src_idx + 1]
src_path = f'{src_dir}/{src_list[src_idx]}_ori.png'
codes = interpolate(src_codes=np.repeat(src_code, num_dst, axis=0),
dst_codes=dst_codes,
step=step)
for dst_idx in tqdm(range(num_dst), leave=False):
dst_path = f'{dst_dir}/{dst_list[dst_idx]}_ori.png'
output_images = []
for idx in range(0, step, args.batch_size):
batch = codes[dst_idx, idx:idx + args.batch_size]
inputs[0:len(batch)] = batch
images = sess.run(x, feed_dict={wp: inputs})
output_images.append(images[0:len(batch)])
output_images = adjust_pixel_range(
np.concatenate(output_images, axis=0))
row_idx = src_idx * num_dst + dst_idx
visualizer.set_cell(row_idx, 0, image=load_image(src_path))
visualizer.set_cell(row_idx, step + 1, image=load_image(dst_path))
for s, output_image in enumerate(output_images):
if s == 5 and row_idx == 2:
save_image(f'./results/interpolation/005_int.png',
output_image)
visualizer.set_cell(row_idx, s + 1, image=output_image)
# Save results.
visualizer.save(f'{output_dir}/{job_name}.html')
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
image_dir = args.image_dir
image_dir_name = os.path.basename(image_dir.rstrip('/'))
assert os.path.exists(image_dir)
assert os.path.exists(f'{image_dir}/image_list.txt')
assert os.path.exists(f'{image_dir}/inverted_codes.npy')
boundary_path = args.boundary_path
assert os.path.exists(boundary_path)
boundary_name = os.path.splitext(os.path.basename(boundary_path))[0]
output_dir = args.output_dir or 'results/manipulation'
job_name = f'{boundary_name.upper()}_{image_dir_name}'
logger = setup_logger(output_dir, f'{job_name}.log', f'{job_name}_logger')
# Load model.
logger.info(f'Loading generator.')
generator = build_generator(args.model_name)
# Load image, codes, and boundary.
logger.info(f'Loading images and corresponding inverted latent codes.')
image_list = []
with open(f'{image_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{image_dir}/{name}_ori.png')
assert os.path.exists(f'{image_dir}/{name}_inv.png')
image_list.append(name)
latent_codes = np.load(f'{image_dir}/inverted_codes.npy')
assert latent_codes.shape[0] == len(image_list)
num_images = latent_codes.shape[0]
logger.info(f'Loading boundary.')
boundary_file = np.load(boundary_path, allow_pickle=True)[()]
if isinstance(boundary_file, dict):
boundary = boundary_file['boundary']
manipulate_layers = boundary_file['meta_data']['manipulate_layers']
else:
boundary = boundary_file
manipulate_layers = args.manipulate_layers
if manipulate_layers:
logger.info(f' Manipulating on layers `{manipulate_layers}`.')
else:
logger.info(f' Manipulating on ALL layers.')
# Manipulate images.
logger.info(f'Start manipulation.')
step = args.step
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(num_rows=num_images,
num_cols=step + 3,
viz_size=viz_size)
visualizer.set_headers(['Name', 'Origin', 'Inverted'] +
[f'Step {i:02d}' for i in range(1, step + 1)])
for img_idx, img_name in enumerate(image_list):
ori_image = load_image(f'{image_dir}/{img_name}_ori.png')
inv_image = load_image(f'{image_dir}/{img_name}_inv.png')
visualizer.set_cell(img_idx, 0, text=img_name)
visualizer.set_cell(img_idx, 1, image=ori_image)
visualizer.set_cell(img_idx, 2, image=inv_image)
codes = manipulate(latent_codes=latent_codes,
boundary=boundary,
start_distance=args.start_distance,
end_distance=args.end_distance,
step=step,
layerwise_manipulation=True,
num_layers=generator.num_layers,
manipulate_layers=manipulate_layers,
is_code_layerwise=True,
is_boundary_layerwise=True)
for img_idx in tqdm(range(num_images), leave=False):
output_images = generator.easy_synthesize(
codes[img_idx], latent_space_type='wp')['image']
for s, output_image in enumerate(output_images):
visualizer.set_cell(img_idx, s + 3, image=output_image)
# Save results.
visualizer.save(f'{output_dir}/{job_name}.html')
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
assert os.path.exists(args.target_list)
target_list_name = os.path.splitext(os.path.basename(args.target_list))[0]
assert os.path.exists(args.context_list)
context_list_name = os.path.splitext(os.path.basename(
args.context_list))[0]
output_dir = args.output_dir or f'results/diffusion'
job_name = f'{target_list_name}_TO_{context_list_name}'
logger = setup_logger(output_dir, f'{job_name}.log', f'{job_name}_logger')
logger.info(f'Loading model.')
tflib.init_tf({'rnd.np_random_seed': 1000})
with open(args.model_path, 'rb') as f:
E, _, _, Gs = pickle.load(f)
# Get input size.
image_size = E.input_shape[2]
assert image_size == E.input_shape[3]
crop_size = args.crop_size
crop_x = args.center_x - crop_size // 2
crop_y = args.center_y - crop_size // 2
mask = np.zeros((1, image_size, image_size, 3), dtype=np.float32)
mask[:, crop_y:crop_y + crop_size, crop_x:crop_x + crop_size, :] = 1.0
# Build graph.
logger.info(f'Building graph.')
sess = tf.get_default_session()
input_shape = E.input_shape
input_shape[0] = args.batch_size
x = tf.placeholder(tf.float32, shape=input_shape, name='real_image')
x_mask = (tf.transpose(x, [0, 2, 3, 1]) + 1) * mask - 1
x_mask_255 = (x_mask + 1) / 2 * 255
latent_shape = Gs.components.synthesis.input_shape
latent_shape[0] = args.batch_size
wp = tf.get_variable(shape=latent_shape, name='latent_code')
x_rec = Gs.components.synthesis.get_output_for(wp, randomize_noise=False)
x_rec_mask = (tf.transpose(x_rec, [0, 2, 3, 1]) + 1) * mask - 1
x_rec_mask_255 = (x_rec_mask + 1) / 2 * 255
w_enc = E.get_output_for(x, phase=False)
wp_enc = tf.reshape(w_enc, latent_shape)
setter = tf.assign(wp, wp_enc)
# Settings for optimization.
logger.info(f'Setting configuration for optimization.')
perceptual_model = PerceptualModel([image_size, image_size], False)
x_feat = perceptual_model(x_mask_255)
x_rec_feat = perceptual_model(x_rec_mask_255)
loss_feat = tf.reduce_mean(tf.square(x_feat - x_rec_feat), axis=[1])
loss_pix = tf.reduce_mean(tf.square(x_mask - x_rec_mask), axis=[1, 2, 3])
loss = loss_pix + args.loss_weight_feat * loss_feat
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
train_op = optimizer.minimize(loss, var_list=[wp])
tflib.init_uninitialized_vars()
# Load image list.
logger.info(f'Loading target images and context images.')
target_list = []
with open(args.target_list, 'r') as f:
for line in f:
target_list.append(line.strip())
num_targets = len(target_list)
context_list = []
with open(args.context_list, 'r') as f:
for line in f:
context_list.append(line.strip())
num_contexts = len(context_list)
num_pairs = num_targets * num_contexts
# Invert images.
logger.info(f'Start diffusion.')
save_interval = args.num_iterations // args.num_results
headers = [
'Target Image', 'Context Image', 'Stitched Image', 'Encoder Output'
]
for step in range(1, args.num_iterations + 1):
if step == args.num_iterations or step % save_interval == 0:
headers.append(f'Step {step:06d}')
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(num_rows=num_pairs,
num_cols=len(headers),
viz_size=viz_size)
visualizer.set_headers(headers)
images = np.zeros(input_shape, np.uint8)
latent_codes_enc = []
latent_codes = []
for target_idx in tqdm(range(num_targets), desc='Target ID', leave=False):
# Load target.
target_image = resize_image(load_image(target_list[target_idx]),
(image_size, image_size))
visualizer.set_cell(target_idx * num_contexts, 0, image=target_image)
for context_idx in tqdm(range(0, num_contexts, args.batch_size),
desc='Context ID',
leave=False):
row_idx = target_idx * num_contexts + context_idx
batch = context_list[context_idx:context_idx + args.batch_size]
for i, context_image_path in enumerate(batch):
context_image = resize_image(load_image(context_image_path),
(image_size, image_size))
visualizer.set_cell(row_idx + i, 1, image=context_image)
context_image[crop_y:crop_y + crop_size, crop_x:crop_x +
crop_size] = (target_image[crop_y:crop_y +
crop_size,
crop_x:crop_x +
crop_size])
visualizer.set_cell(row_idx + i, 2, image=context_image)
images[i] = np.transpose(context_image, [2, 0, 1])
inputs = images.astype(np.float32) / 255 * 2.0 - 1.0
# Run encoder.
sess.run([setter], {x: inputs})
outputs = sess.run([wp, x_rec])
latent_codes_enc.append(outputs[0][0:len(batch)])
outputs[1] = adjust_pixel_range(outputs[1])
for i, _ in enumerate(batch):
visualizer.set_cell(row_idx + i, 3, image=outputs[1][i])
# Optimize latent codes.
col_idx = 4
for step in tqdm(range(1, args.num_iterations + 1), leave=False):
sess.run(train_op, {x: inputs})
if step == args.num_iterations or step % save_interval == 0:
outputs = sess.run([wp, x_rec])
outputs[1] = adjust_pixel_range(outputs[1])
for i, _ in enumerate(batch):
visualizer.set_cell(row_idx + i,
col_idx,
image=outputs[1][i])
col_idx += 1
latent_codes.append(outputs[0][0:len(batch)])
# Save results.
code_shape = [num_targets, num_contexts] + list(latent_shape[1:])
np.save(f'{output_dir}/{job_name}_encoded_codes.npy',
np.concatenate(latent_codes_enc, axis=0).reshape(code_shape))
np.save(f'{output_dir}/{job_name}_inverted_codes.npy',
np.concatenate(latent_codes, axis=0).reshape(code_shape))
visualizer.save(f'{output_dir}/{job_name}.html')
def synthesize(self,
num,
z=None,
html_name=None,
save_raw_synthesis=False):
"""Synthesizes images.
Args:
num: Number of images to synthesize.
z: Latent codes used for generation. If not specified, this function
will sample latent codes randomly. (default: None)
html_name: Name of the output html page for visualization. If not
specified, no visualization page will be saved. (default: None)
save_raw_synthesis: Whether to save raw synthesis on the disk.
(default: False)
"""
if not html_name and not save_raw_synthesis:
return
self.set_mode('val')
temp_dir = os.path.join(self.work_dir, 'synthesize_results')
os.makedirs(temp_dir, exist_ok=True)
if z is not None:
assert isinstance(z, np.ndarray)
assert z.ndim == 2 and z.shape[1] == self.z_space_dim
num = min(num, z.shape[0])
z = torch.from_numpy(z).type(torch.FloatTensor)
if not num:
return
# TODO: Use same z during the entire training process.
self.logger.init_pbar()
task1 = self.logger.add_pbar_task('Synthesize', total=num)
indices = list(range(self.rank, num, self.world_size))
for batch_idx in range(0, len(indices), self.val_batch_size):
sub_indices = indices[batch_idx:batch_idx + self.val_batch_size]
batch_size = len(sub_indices)
if z is None:
code = torch.randn(batch_size, self.z_space_dim).cuda()
else:
code = z[sub_indices].cuda()
with torch.no_grad():
if 'generator_smooth' in self.models:
G = self.models['generator_smooth']
else:
G = self.models['generator']
images = G(code, **self.G_kwargs_val)['image']
images = self.postprocess(images)
for sub_idx, image in zip(sub_indices, images):
save_image(os.path.join(temp_dir, f'{sub_idx:06d}.jpg'), image)
self.logger.update_pbar(task1, batch_size * self.world_size)
dist.barrier()
if self.rank != 0:
return
if html_name:
task2 = self.logger.add_pbar_task('Visualize', total=num)
html = HtmlPageVisualizer(grid_size=num)
for image_idx in range(num):
image = load_image(
os.path.join(temp_dir, f'{image_idx:06d}.jpg'))
row_idx, col_idx = divmod(image_idx, html.num_cols)
html.set_cell(row_idx,
col_idx,
image=image,
text=f'Sample {image_idx:06d}')
self.logger.update_pbar(task2, 1)
html.save(os.path.join(self.work_dir, html_name))
if not save_raw_synthesis:
shutil.rmtree(temp_dir)
self.logger.close_pbar()
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
image_dir = args.image_dir
image_dir_name = os.path.basename(image_dir.rstrip('/'))
assert os.path.exists(image_dir)
assert os.path.exists(f'{image_dir}/image_list.txt')
assert os.path.exists(f'{image_dir}/inverted_codes.npy')
boundary_path = args.boundary_path
assert os.path.exists(boundary_path)
boundary_name = os.path.splitext(os.path.basename(boundary_path))[0]
output_dir = args.output_dir or 'results/manipulation'
job_name = f'{boundary_name.upper()}_{image_dir_name}'
logger = setup_logger(output_dir, f'{job_name}.log', f'{job_name}_logger')
# Load model.
logger.info(f'Loading generator.')
tflib.init_tf({'rnd.np_random_seed': 1000})
with open(args.model_path, 'rb') as f:
_, _, _, Gs = pickle.load(f)
# Build graph.
logger.info(f'Building graph.')
sess = tf.get_default_session()
num_layers, latent_dim = Gs.components.synthesis.input_shape[1:3]
wp = tf.placeholder(tf.float32, [args.batch_size, num_layers, latent_dim],
name='latent_code')
x = Gs.components.synthesis.get_output_for(wp, randomize_noise=False)
# Load image, codes, and boundary.
logger.info(f'Loading images and corresponding inverted latent codes.')
image_list = []
with open(f'{image_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{image_dir}/{name}_ori.png')
assert os.path.exists(f'{image_dir}/{name}_inv.png')
image_list.append(name)
latent_codes = np.load(f'{image_dir}/inverted_codes.npy')
assert latent_codes.shape[0] == len(image_list)
num_images = latent_codes.shape[0]
logger.info(f'Loading boundary.')
boundary_file = np.load(boundary_path, allow_pickle=True)[()]
if isinstance(boundary_file, dict):
boundary = boundary_file['boundary']
manipulate_layers = boundary_file['meta_data']['manipulate_layers']
else:
boundary = boundary_file
manipulate_layers = args.manipulate_layers
if manipulate_layers:
logger.info(f' Manipulating on layers `{manipulate_layers}`.')
else:
logger.info(f' Manipulating on ALL layers.')
# Manipulate images.
logger.info(f'Start manipulation.')
step = args.step
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(num_rows=num_images,
num_cols=step + 3,
viz_size=viz_size)
visualizer.set_headers(['Name', 'Origin', 'Inverted'] +
[f'Step {i:02d}' for i in range(1, step + 1)])
for img_idx, img_name in enumerate(image_list):
ori_image = load_image(f'{image_dir}/{img_name}_ori.png')
inv_image = load_image(f'{image_dir}/{img_name}_inv.png')
visualizer.set_cell(img_idx, 0, text=img_name)
visualizer.set_cell(img_idx, 1, image=ori_image)
visualizer.set_cell(img_idx, 2, image=inv_image)
codes = manipulate(latent_codes=latent_codes,
boundary=boundary,
start_distance=args.start_distance,
end_distance=args.end_distance,
step=step,
layerwise_manipulation=True,
num_layers=num_layers,
manipulate_layers=manipulate_layers,
is_code_layerwise=True,
is_boundary_layerwise=True)
inputs = np.zeros((args.batch_size, num_layers, latent_dim), np.float32)
for img_idx in tqdm(range(num_images), leave=False):
output_images = []
for idx in range(0, step, args.batch_size):
batch = codes[img_idx, idx:idx + args.batch_size]
inputs[0:len(batch)] = batch
images = sess.run(x, feed_dict={wp: inputs})
output_images.append(images[0:len(batch)])
output_images = adjust_pixel_range(
np.concatenate(output_images, axis=0))
for s, output_image in enumerate(output_images):
visualizer.set_cell(img_idx, s + 3, image=output_image)
# Save results.
visualizer.save(f'{output_dir}/{job_name}.html')
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
style_dir = args.style_dir
style_dir_name = os.path.basename(style_dir.rstrip('/'))
assert os.path.exists(style_dir)
assert os.path.exists(f'{style_dir}/image_list.txt')
assert os.path.exists(f'{style_dir}/inverted_codes.npy')
content_dir = args.content_dir
content_dir_name = os.path.basename(content_dir.rstrip('/'))
assert os.path.exists(content_dir)
assert os.path.exists(f'{content_dir}/image_list.txt')
assert os.path.exists(f'{content_dir}/inverted_codes.npy')
output_dir = args.output_dir or 'results/style_mixing'
job_name = f'{style_dir_name}_STYLIZE_{content_dir_name}'
logger = setup_logger(output_dir, f'{job_name}.log', f'{job_name}_logger')
# Load model.
logger.info(f'Loading generator.')
generator = build_generator(args.model_name)
mix_layers = list(range(args.mix_layer_start_idx, generator.num_layers))
# Load image and codes.
logger.info(f'Loading images and corresponding inverted latent codes.')
style_list = []
with open(f'{style_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{style_dir}/{name}_ori.png')
style_list.append(name)
logger.info(f'Loading inverted latent codes.')
style_codes = np.load(f'{style_dir}/inverted_codes.npy')
assert style_codes.shape[0] == len(style_list)
num_styles = style_codes.shape[0]
content_list = []
with open(f'{content_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{content_dir}/{name}_ori.png')
content_list.append(name)
logger.info(f'Loading inverted latent codes.')
content_codes = np.load(f'{content_dir}/inverted_codes.npy')
assert content_codes.shape[0] == len(content_list)
num_contents = content_codes.shape[0]
# Mix styles.
logger.info(f'Start style mixing.')
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(num_rows=num_styles + 1,
num_cols=num_contents + 1,
viz_size=viz_size)
visualizer.set_headers(['Style'] +
[f'Content {i:03d}' for i in range(num_contents)])
for style_idx, style_name in enumerate(style_list):
style_image = load_image(f'{style_dir}/{style_name}_ori.png')
visualizer.set_cell(style_idx + 1, 0, image=style_image)
for content_idx, content_name in enumerate(content_list):
content_image = load_image(f'{content_dir}/{content_name}_ori.png')
visualizer.set_cell(0, content_idx + 1, image=content_image)
codes = mix_style(style_codes=style_codes,
content_codes=content_codes,
num_layers=generator.num_layers,
mix_layers=mix_layers)
for style_idx in tqdm(range(num_styles), leave=False):
output_images = generator.easy_synthesize(
codes[style_idx], latent_space_type='wp')['image']
for content_idx, output_image in enumerate(output_images):
visualizer.set_cell(style_idx + 1,
content_idx + 1,
image=output_image)
# Save results.
visualizer.save(f'{output_dir}/{job_name}.html')
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
assert os.path.exists(args.target_list)
target_list_name = os.path.splitext(os.path.basename(args.target_list))[0]
assert os.path.exists(args.context_list)
context_list_name = os.path.splitext(os.path.basename(args.context_list))[0]
output_dir = args.output_dir or f'results/diffusion'
job_name = f'{target_list_name}_TO_{context_list_name}'
logger = setup_logger(output_dir, f'{job_name}.log', f'{job_name}_logger')
logger.info(f'Loading model.')
inverter = StyleGANInverter(
args.model_name,
learning_rate=args.learning_rate,
iteration=args.num_iterations,
reconstruction_loss_weight=1.0,
perceptual_loss_weight=args.loss_weight_feat,
regularization_loss_weight=0.0,
logger=logger)
image_size = inverter.G.resolution
# Load image list.
logger.info(f'Loading target images and context images.')
target_list = []
with open(args.target_list, 'r') as f:
for line in f:
target_list.append(line.strip())
num_targets = len(target_list)
context_list = []
with open(args.context_list, 'r') as f:
for line in f:
context_list.append(line.strip())
num_contexts = len(context_list)
num_pairs = num_targets * num_contexts
# Initialize visualizer.
save_interval = args.num_iterations // args.num_results
headers = ['Target Image', 'Context Image', 'Stitched Image',
'Encoder Output']
for step in range(1, args.num_iterations + 1):
if step == args.num_iterations or step % save_interval == 0:
headers.append(f'Step {step:06d}')
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(
num_rows=num_pairs, num_cols=len(headers), viz_size=viz_size)
visualizer.set_headers(headers)
# Diffuse images.
logger.info(f'Start diffusion.')
latent_codes = []
for target_idx in tqdm(range(num_targets), desc='Target ID', leave=False):
# Load target.
target_image = resize_image(load_image(target_list[target_idx]),
(image_size, image_size))
visualizer.set_cell(target_idx * num_contexts, 0, image=target_image)
for context_idx in tqdm(range(num_contexts), desc='Context ID',
leave=False):
row_idx = target_idx * num_contexts + context_idx
context_image = resize_image(load_image(context_list[context_idx]),
(image_size, image_size))
visualizer.set_cell(row_idx, 1, image=context_image)
code, viz_results = inverter.easy_diffuse(target=target_image,
context=context_image,
center_x=args.center_x,
center_y=args.center_y,
crop_x=args.crop_size,
crop_y=args.crop_size,
num_viz=args.num_results)
for viz_idx, viz_img in enumerate(viz_results):
visualizer.set_cell(row_idx, viz_idx + 2, image=viz_img)
latent_codes.append(code)
# Save results.
os.system(f'cp {args.target_list} {output_dir}/target_list.txt')
os.system(f'cp {args.context_list} {output_dir}/context_list.txt')
np.save(f'{output_dir}/{job_name}_inverted_codes.npy',
np.concatenate(latent_codes, axis=0))
visualizer.save(f'{output_dir}/{job_name}.html')
def encode(_target_image, _context_image, _output_dir):
gpu_id = '0'
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_id
print(_target_image)
assert os.path.exists('./static/' + _target_image)
_output_dir = _output_dir[:-4]
output_dir = './static/' + _output_dir
tflib.init_tf({'rnd.np_random_seed': 1000})
model_path = './styleganinv_face_256.pkl'
with open(model_path, 'rb') as f:
E, _, _, Gs = pickle.load(f)
# Get input size.
image_size = E.input_shape[2]
assert image_size == E.input_shape[3]
crop_size = 110 # default crop size.
center_x = 125
center_y = 145
crop_x = center_x - crop_size // 2 # default coordinate-X
crop_y = center_y - crop_size // 2 # default coordinate-Y
mask = np.zeros((1, image_size, image_size, 3), dtype=np.float32)
mask[:, crop_y:crop_y + crop_size, crop_x:crop_x + crop_size, :] = 1.0
# Build graph.
sess = tf.get_default_session()
batch_size = 4
input_shape = E.input_shape
input_shape[0] = batch_size # default batch size
x = tf.placeholder(tf.float32, shape=input_shape, name='real_image')
x_mask = (tf.transpose(x, [0, 2, 3, 1]) + 1) * mask - 1
x_mask_255 = (x_mask + 1) / 2 * 255
latent_shape = Gs.components.synthesis.input_shape
latent_shape[0] = batch_size # default batch size
wp = tf.get_variable(shape=latent_shape, name='latent_code')
x_rec = Gs.components.synthesis.get_output_for(wp, randomize_noise=False)
x_rec_mask = (tf.transpose(x_rec, [0, 2, 3, 1]) + 1) * mask - 1
x_rec_mask_255 = (x_rec_mask + 1) / 2 * 255
w_enc = E.get_output_for(x, phase=False)
wp_enc = tf.reshape(w_enc, latent_shape)
setter = tf.assign(wp, wp_enc)
# Settings for optimization.
print("Diffusion : Settings for Optimization.")
perceptual_model = PerceptualModel([image_size, image_size], False)
x_feat = perceptual_model(x_mask_255)
x_rec_feat = perceptual_model(x_rec_mask_255)
loss_feat = tf.reduce_mean(tf.square(x_feat - x_rec_feat), axis=[1])
loss_pix = tf.reduce_mean(tf.square(x_mask - x_rec_mask), axis=[1, 2, 3])
loss_weight_feat = 5e-5
learning_rate = 0.01
loss = loss_pix + loss_weight_feat * loss_feat # default The perceptual loss scale for optimization. (default 5e-5)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(loss, var_list=[wp])
tflib.init_uninitialized_vars()
# Invert image
num_iterations = 100
num_results = 5
save_interval = num_iterations // num_results
images = np.zeros(input_shape, np.uint8)
print("Load target image.")
_target_image = './static/' + _target_image
target_image = resize_image(load_image(_target_image),
(image_size, image_size))
save_image('./' + output_dir + '_tar.png', target_image)
print("Load context image.")
context_image = getContextImage(_context_image)
context_image = resize_image(load_image(context_image),
(image_size, image_size))
save_image('./' + output_dir + '_cont.png', context_image)
# Inverting Context Image.
# context_image = invert(model_path, getContextImage(_context_image), wp, latent_shape)
save_image('./' + output_dir + '_cont_inv.png', context_image)
# Create Stitched Image
# context_image[crop_y:crop_y + crop_size, crop_x:crop_x + crop_size] = (
# target_image[crop_y:crop_y + crop_size, crop_x:crop_x + crop_size]
# )
# context_image[crop_y:crop_y + 170, crop_x - 70:crop_x + crop_size + 190] = (
# target_image[crop_y:crop_y + 170, crop_x - 70:crop_x + crop_size + 190]
# )
print("Cropping Image...")
# context_image = cropImage(target_image, context_image)
target_image, rect = cropWithWhite(target_image)
target_image = fourChannels(target_image)
target_image = cut(target_image)
target_image = transBg(target_image)
context_image = createStitchedImage(context_image, target_image, rect)
save_image('./' + output_dir + '_sti.png', context_image)
images[0] = np.transpose(context_image, [2, 0, 1])
input = images.astype(np.float32) / 255 * 2.0 - 1.0
# Run encoder
print("Start Diffusion.")
sess.run([setter], {x: input})
output = sess.run([wp, x_rec])
output[1] = adjust_pixel_range(output[1])
col_idx = 4
for step in tqdm(range(1, num_iterations + 1), leave=False):
sess.run(train_op, {x: input})
if step == num_iterations or step % save_interval == 0:
output = sess.run([wp, x_rec])
output[1] = adjust_pixel_range(output[1])
if step == num_iterations:
save_image(f'{output_dir}.png', output[1][0])
col_idx += 1
exit()
def invert(model_path, _image, _wp, _latent_shape):
print("Inverting")
tflib.init_tf({'rnd.np_random_seed': 1000})
with open(model_path, 'rb') as f:
E, _, _, Gs = pickle.load(f)
# Get input size.
image_size = E.input_shape[2]
assert image_size == E.input_shape[3]
# Build graph.
print("Inverting : Build Graph.")
sess = tf.get_default_session()
batch_size = 4
input_shape = E.input_shape
input_shape[0] = batch_size # default batch size
x = tf.placeholder(tf.float32, shape=input_shape, name='real_image')
x_255 = (tf.transpose(x, [0, 2, 3, 1]) + 1) / 2 * 255
wp = _wp
x_rec = Gs.components.synthesis.get_output_for(wp, randomize_noise=False)
x_rec_255 = (tf.transpose(x_rec, [0, 2, 3, 1]) + 1) / 2 * 255
w_enc = E.get_output_for(x, phase=False)
wp_enc = tf.reshape(w_enc, _latent_shape)
setter = tf.assign(wp, wp_enc)
# Settings for optimization.
print("Inverting : Settings for Optimization.")
perceptual_model = PerceptualModel([image_size, image_size], False)
x_feat = perceptual_model(x_255)
x_rec_feat = perceptual_model(x_rec_255)
loss_feat = tf.reduce_mean(tf.square(x_feat - x_rec_feat), axis=[1])
loss_pix = tf.reduce_mean(tf.square(x - x_rec), axis=[1, 2, 3])
w_enc_new = E.get_output_for(x_rec, phase=False)
wp_enc_new = tf.reshape(w_enc_new, _latent_shape)
loss_enc = tf.reduce_mean(tf.square(wp - wp_enc_new), axis=[1, 2])
loss = (loss_pix + 5e-5 * loss_feat + 2.0 * loss_enc)
optimizer = tf.train.AdamOptimizer(learning_rate=0.01)
train_op = optimizer.minimize(loss, var_list=[wp])
tflib.init_uninitialized_vars()
# Invert image
print("Start Inverting.")
num_iterations = 40
num_results = 2
save_interval = num_iterations // num_results
context_images = np.zeros(input_shape, np.uint8)
context_image = resize_image(load_image(_image), (image_size, image_size))
# Inverting Context Image.
context_images[0] = np.transpose(context_image, [2, 0, 1])
context_input = context_images.astype(np.float32) / 255 * 2.0 - 1.0
sess.run([setter], {x: context_input})
context_output = sess.run([wp, x_rec])
context_output[1] = adjust_pixel_range(context_output[1])
context_image = np.transpose(context_images[0], [1, 2, 0])
for step in tqdm(range(1, num_iterations + 1), leave=False):
sess.run(train_op, {x: context_input})
if step == num_iterations or step % save_interval == 0:
context_output = sess.run([wp, x_rec])
context_output[1] = adjust_pixel_range(context_output[1])
if step == num_iterations: context_image = context_output[1][0]
return context_image
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
assert os.path.exists(args.image_list)
image_list_name = os.path.splitext(os.path.basename(args.image_list))[0]
output_dir = args.output_dir or f'results/inversion/{image_list_name}'
logger = setup_logger(output_dir, 'inversion.log', 'inversion_logger')
logger.info(f'Loading model.')
tflib.init_tf({'rnd.np_random_seed': 1000})
with open(args.model_path, 'rb') as f:
E, _, _, Gs = pickle.load(f)
# Get input size.
image_size = E.input_shape[2]
assert image_size == E.input_shape[3]
# Build graph.
logger.info(f'Building graph.')
sess = tf.get_default_session()
input_shape = E.input_shape
input_shape[0] = args.batch_size
x = tf.placeholder(tf.float32, shape=input_shape, name='real_image')
x_255 = (tf.transpose(x, [0, 2, 3, 1]) + 1) / 2 * 255
latent_shape = Gs.components.synthesis.input_shape
latent_shape[0] = args.batch_size
wp = tf.get_variable(shape=latent_shape, name='latent_code')
x_rec = Gs.components.synthesis.get_output_for(wp, randomize_noise=False)
x_rec_255 = (tf.transpose(x_rec, [0, 2, 3, 1]) + 1) / 2 * 255
if args.random_init:
logger.info(f' Use random initialization for optimization.')
wp_rnd = tf.random.normal(shape=latent_shape, name='latent_code_init')
setter = tf.assign(wp, wp_rnd)
else:
logger.info(
f' Use encoder output as the initialization for optimization.')
w_enc = E.get_output_for(x, is_training=False)
wp_enc = tf.reshape(w_enc, latent_shape)
setter = tf.assign(wp, wp_enc)
# Settings for optimization.
logger.info(f'Setting configuration for optimization.')
perceptual_model = PerceptualModel([image_size, image_size], False)
x_feat = perceptual_model(x_255)
x_rec_feat = perceptual_model(x_rec_255)
loss_feat = tf.reduce_mean(tf.square(x_feat - x_rec_feat), axis=[1])
loss_pix = tf.reduce_mean(tf.square(x - x_rec), axis=[1, 2, 3])
if args.domain_regularizer:
logger.info(f' Involve encoder for optimization.')
w_enc_new = E.get_output_for(x_rec, is_training=False)
wp_enc_new = tf.reshape(w_enc_new, latent_shape)
loss_enc = tf.reduce_mean(tf.square(wp - wp_enc_new), axis=[1, 2])
else:
logger.info(f' Do NOT involve encoder for optimization.')
loss_enc = 0
loss = (loss_pix + args.loss_weight_feat * loss_feat +
args.loss_weight_enc * loss_enc)
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
train_op = optimizer.minimize(loss, var_list=[wp])
tflib.init_uninitialized_vars()
# Load image list.
logger.info(f'Loading image list.')
image_list = []
with open(args.image_list, 'r') as f:
for line in f:
image_list.append(line.strip())
# Invert images.
logger.info(f'Start inversion.')
save_interval = args.num_iterations // args.num_results
headers = ['Name', 'Original Image', 'Encoder Output']
for step in range(1, args.num_iterations + 1):
if step == args.num_iterations or step % save_interval == 0:
headers.append(f'Step {step:06d}')
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(num_rows=len(image_list),
num_cols=len(headers),
viz_size=viz_size)
visualizer.set_headers(headers)
images = np.zeros(input_shape, np.uint8)
names = ['' for _ in range(args.batch_size)]
latent_codes_enc = []
latent_codes = []
for img_idx in tqdm(range(0, len(image_list), args.batch_size),
leave=False):
# Load inputs.
batch = image_list[img_idx:img_idx + args.batch_size]
for i, image_path in enumerate(batch):
image = resize_image(load_image(image_path),
(image_size, image_size))
images[i] = np.transpose(image, [2, 0, 1])
names[i] = os.path.splitext(os.path.basename(image_path))[0]
inputs = images.astype(np.float32) / 255 * 2.0 - 1.0
# Run encoder.
sess.run([setter], {x: inputs})
outputs = sess.run([wp, x_rec])
latent_codes_enc.append(outputs[0][0:len(batch)])
outputs[1] = adjust_pixel_range(outputs[1])
for i, _ in enumerate(batch):
image = np.transpose(images[i], [1, 2, 0])
save_image(f'{output_dir}/{names[i]}_ori.png', image)
save_image(f'{output_dir}/{names[i]}_enc.png', outputs[1][i])
visualizer.set_cell(i + img_idx, 0, text=names[i])
visualizer.set_cell(i + img_idx, 1, image=image)
visualizer.set_cell(i + img_idx, 2, image=outputs[1][i])
# Optimize latent codes.
col_idx = 3
for step in tqdm(range(1, args.num_iterations + 1), leave=False):
sess.run(train_op, {x: inputs})
if step == args.num_iterations or step % save_interval == 0:
outputs = sess.run([wp, x_rec])
outputs[1] = adjust_pixel_range(outputs[1])
for i, _ in enumerate(batch):
if step == args.num_iterations:
save_image(f'{output_dir}/{names[i]}_inv.png',
outputs[1][i])
visualizer.set_cell(i + img_idx,
col_idx,
image=outputs[1][i])
col_idx += 1
latent_codes.append(outputs[0][0:len(batch)])
# Save results.
os.system(f'cp {args.image_list} {output_dir}/image_list.txt')
np.save(f'{output_dir}/encoded_codes.npy',
np.concatenate(latent_codes_enc, axis=0))
np.save(f'{output_dir}/inverted_codes.npy',
np.concatenate(latent_codes, axis=0))
visualizer.save(f'{output_dir}/inversion.html')
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
src_dir = args.src_dir
src_dir_name = os.path.basename(src_dir.rstrip('/'))
assert os.path.exists(src_dir)
assert os.path.exists(f'{src_dir}/image_list.txt')
assert os.path.exists(f'{src_dir}/inverted_codes.npy')
dst_dir = args.dst_dir
dst_dir_name = os.path.basename(dst_dir.rstrip('/'))
assert os.path.exists(dst_dir)
assert os.path.exists(f'{dst_dir}/image_list.txt')
assert os.path.exists(f'{dst_dir}/inverted_codes.npy')
output_dir = args.output_dir or 'results/interpolation'
job_name = f'{src_dir_name}_TO_{dst_dir_name}'
logger = setup_logger(output_dir, f'{job_name}.log', f'{job_name}_logger')
# Load model.
logger.info(f'Loading generator.')
generator = build_generator(args.model_name)
# Load image and codes.
logger.info(f'Loading images and corresponding inverted latent codes.')
src_list = []
with open(f'{src_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{src_dir}/{name}_ori.png')
src_list.append(name)
src_codes = np.load(f'{src_dir}/inverted_codes.npy')
assert src_codes.shape[0] == len(src_list)
num_src = src_codes.shape[0]
dst_list = []
with open(f'{dst_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{dst_dir}/{name}_ori.png')
dst_list.append(name)
dst_codes = np.load(f'{dst_dir}/inverted_codes.npy')
assert dst_codes.shape[0] == len(dst_list)
num_dst = dst_codes.shape[0]
# Interpolate images.
logger.info(f'Start interpolation.')
step = args.step + 2
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(num_rows=num_src * num_dst,
num_cols=step + 2,
viz_size=viz_size)
visualizer.set_headers(['Source', 'Source Inversion'] +
[f'Step {i:02d}' for i in range(1, step - 1)] +
['Target Inversion', 'Target'])
for src_idx in tqdm(range(num_src), leave=False):
src_code = src_codes[src_idx:src_idx + 1]
src_path = f'{src_dir}/{src_list[src_idx]}_ori.png'
codes = interpolate(src_codes=np.repeat(src_code, num_dst, axis=0),
dst_codes=dst_codes,
step=step)
for dst_idx in tqdm(range(num_dst), leave=False):
dst_path = f'{dst_dir}/{dst_list[dst_idx]}_ori.png'
output_images = generator.easy_synthesize(
codes[dst_idx], latent_space_type='wp')['image']
row_idx = src_idx * num_dst + dst_idx
visualizer.set_cell(row_idx, 0, image=load_image(src_path))
visualizer.set_cell(row_idx, step + 1, image=load_image(dst_path))
for s, output_image in enumerate(output_images):
visualizer.set_cell(row_idx, s + 1, image=output_image)
# Save results.
visualizer.save(f'{output_dir}/{job_name}.html')
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
style_dir = args.style_dir
style_dir_name = os.path.basename(style_dir.rstrip('/'))
assert os.path.exists(style_dir)
assert os.path.exists(f'{style_dir}/image_list.txt')
assert os.path.exists(f'{style_dir}/inverted_codes.npy')
content_dir = args.content_dir
content_dir_name = os.path.basename(content_dir.rstrip('/'))
assert os.path.exists(content_dir)
assert os.path.exists(f'{content_dir}/image_list.txt')
assert os.path.exists(f'{content_dir}/inverted_codes.npy')
output_dir = args.output_dir or 'results/style_mixing'
job_name = f'{style_dir_name}_STYLIZE_{content_dir_name}'
logger = setup_logger(output_dir, f'{job_name}.log', f'{job_name}_logger')
# Load model.
logger.info(f'Loading generator.')
tflib.init_tf({'rnd.np_random_seed': 1000})
with open(args.model_path, 'rb') as f:
_, _, _, Gs = pickle.load(f)
# Build graph.
logger.info(f'Building graph.')
sess = tf.get_default_session()
num_layers, latent_dim = Gs.components.synthesis.input_shape[1:3]
wp = tf.placeholder(
tf.float32, [args.batch_size, num_layers, latent_dim], name='latent_code')
x = Gs.components.synthesis.get_output_for(wp, randomize_noise=False)
mix_layers = list(range(args.mix_layer_start_idx, num_layers))
# Load image and codes.
logger.info(f'Loading images and corresponding inverted latent codes.')
style_list = []
with open(f'{style_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{style_dir}/{name}_ori.png')
style_list.append(name)
logger.info(f'Loading inverted latent codes.')
style_codes = np.load(f'{style_dir}/inverted_codes.npy')
assert style_codes.shape[0] == len(style_list)
num_styles = style_codes.shape[0]
content_list = []
with open(f'{content_dir}/image_list.txt', 'r') as f:
for line in f:
name = os.path.splitext(os.path.basename(line.strip()))[0]
assert os.path.exists(f'{content_dir}/{name}_ori.png')
content_list.append(name)
logger.info(f'Loading inverted latent codes.')
content_codes = np.load(f'{content_dir}/inverted_codes.npy')
assert content_codes.shape[0] == len(content_list)
num_contents = content_codes.shape[0]
# Mix styles.
logger.info(f'Start style mixing.')
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(
num_rows=num_styles + 1, num_cols=num_contents + 1, viz_size=viz_size)
visualizer.set_headers(
['Style'] +
[f'Content {i:03d}' for i in range(num_contents)]
)
for style_idx, style_name in enumerate(style_list):
style_image = load_image(f'{style_dir}/{style_name}_ori.png')
visualizer.set_cell(style_idx + 1, 0, image=style_image)
for content_idx, content_name in enumerate(content_list):
content_image = load_image(f'{content_dir}/{content_name}_ori.png')
visualizer.set_cell(0, content_idx + 1, image=content_image)
codes = mix_style(style_codes=style_codes,
content_codes=content_codes,
num_layers=num_layers,
mix_layers=mix_layers)
inputs = np.zeros((args.batch_size, num_layers, latent_dim), np.float32)
for style_idx in tqdm(range(num_styles), leave=False):
output_images = []
for idx in range(0, num_contents, args.batch_size):
batch = codes[style_idx, idx:idx + args.batch_size]
inputs[0:len(batch)] = batch
images = sess.run(x, feed_dict={wp: inputs})
output_images.append(images[0:len(batch)])
output_images = adjust_pixel_range(np.concatenate(output_images, axis=0))
for content_idx, output_image in enumerate(output_images):
visualizer.set_cell(style_idx + 1, content_idx + 1, image=output_image)
# Save results.
visualizer.save(f'{output_dir}/{job_name}.html')
def main():
"""Main function."""
args = parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
assert os.path.exists(args.image_list)
image_list_name = os.path.splitext(os.path.basename(args.image_list))[0]
output_dir = args.output_dir or f'results/ghfeat/{image_list_name}'
logger = setup_logger(output_dir, 'extract_feature.log',
'inversion_logger')
logger.info(f'Loading model.')
tflib.init_tf({'rnd.np_random_seed': 1000})
with open(args.model_path, 'rb') as f:
E, _, _, Gs = pickle.load(f)
# Get input size.
image_size = E.input_shape[2]
assert image_size == E.input_shape[3]
G_args = EasyDict(func_name='training.networks_stylegan.G_synthesis')
G_style_mod = tflib.Network('G_StyleMod',
resolution=image_size,
label_size=0,
**G_args)
Gs_vars_pairs = {
name: tflib.run(val)
for name, val in Gs.components.synthesis.vars.items()
}
for g_name, g_val in G_style_mod.vars.items():
tflib.set_vars({g_val: Gs_vars_pairs[g_name]})
# Build graph.
logger.info(f'Building graph.')
sess = tf.get_default_session()
input_shape = E.input_shape
input_shape[0] = args.batch_size
x = tf.placeholder(tf.float32, shape=input_shape, name='real_image')
ghfeat = E.get_output_for(x, is_training=False)
x_rec = G_style_mod.get_output_for(ghfeat, randomize_noise=False)
# Load image list.
logger.info(f'Loading image list.')
image_list = []
with open(args.image_list, 'r') as f:
for line in f:
image_list.append(line.strip())
# Extract GH-Feat from images.
logger.info(f'Start feature extraction.')
headers = ['Name', 'Original Image', 'Encoder Output']
viz_size = None if args.viz_size == 0 else args.viz_size
visualizer = HtmlPageVisualizer(num_rows=len(image_list),
num_cols=len(headers),
viz_size=viz_size)
visualizer.set_headers(headers)
images = np.zeros(input_shape, np.uint8)
names = ['' for _ in range(args.batch_size)]
features = []
for img_idx in tqdm(range(0, len(image_list), args.batch_size),
leave=False):
# Load inputs.
batch = image_list[img_idx:img_idx + args.batch_size]
for i, image_path in enumerate(batch):
image = resize_image(load_image(image_path),
(image_size, image_size))
images[i] = np.transpose(image, [2, 0, 1])
names[i] = os.path.splitext(os.path.basename(image_path))[0]
inputs = images.astype(np.float32) / 255 * 2.0 - 1.0
# Run encoder.
outputs = sess.run([ghfeat, x_rec], {x: inputs})
features.append(outputs[0][0:len(batch)])
outputs[1] = adjust_pixel_range(outputs[1])
for i, _ in enumerate(batch):
image = np.transpose(images[i], [1, 2, 0])
save_image(f'{output_dir}/{names[i]}_ori.png', image)
save_image(f'{output_dir}/{names[i]}_enc.png', outputs[1][i])
visualizer.set_cell(i + img_idx, 0, text=names[i])
visualizer.set_cell(i + img_idx, 1, image=image)
visualizer.set_cell(i + img_idx, 2, image=outputs[1][i])
# Save results.
os.system(f'cp {args.image_list} {output_dir}/image_list.txt')
np.save(f'{output_dir}/ghfeat.npy', np.concatenate(features, axis=0))
visualizer.save(f'{output_dir}/reconstruction.html')
