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 optimize_latent_codes_source_target(args, source_latent_code):
#tf.enable_eager_execution()
tflib.init_tf()
with dnnlib.util.open_url(STYLEGAN_MODEL_URL,
cache_dir=config.cache_dir) as f:
_G, _D, Gs = pickle.load(f)
latent_code_2 = tf.get_variable(
name='latent_code_2',
shape=(1, 18, 512),
dtype='float32',
initializer=tf.initializers.constant(source_latent_code))
generated_img = Gs.components.synthesis.get_output_for(
latent_code_2, randomize_noise=False)
generated_img = tf.transpose(generated_img, [0, 2, 3, 1])
generated_img = ((generated_img + 1) / 2) * 255
original_img = tf.placeholder(
tf.float32, [None, args.input_img_size[0], args.input_img_size[1], 3])
blur_kernel = tf.placeholder(
tf.float32,
[None, args.blur_parameters[0], args.blur_parameters[0], 3])
degraded_img_resized_for_perceptual = tf.image.resize_images(
add_motion_blur(original_img, blur_kernel),
tuple(args.perceptual_img_size),
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
generated_img_resized_to_original = tf.image.resize_images(
generated_img,
tuple(args.input_img_size),
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
generated_img_resized_for_perceptual = tf.image.resize_images(
add_motion_blur(generated_img_resized_to_original, blur_kernel),
tuple(args.perceptual_img_size),
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
generated_img_for_display = tf.saturate_cast(
generated_img_resized_to_original, tf.uint8)
perceptual_model = PerceptualModel(img_size=args.perceptual_img_size)
generated_img_features = perceptual_model(
generated_img_resized_for_perceptual)
target_img_features = perceptual_model(degraded_img_resized_for_perceptual)
loss_op = tf.reduce_mean(
tf.abs(generated_img_features - target_img_features))
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=args.learning_rate)
train_op = optimizer.minimize(loss_op, var_list=[latent_code_2])
sess = tf.get_default_session()
img_names = sorted(os.listdir(args.imgs_dir_second))
for img_name in img_names:
img = imageio.imread(os.path.join(args.imgs_dir_second, img_name))
img = cv2.resize(img, dsize=tuple(args.input_img_size))
blur_kernel_3d = motion_blur_kernel(args.blur_parameters[0],
args.blur_parameters[1])
corrupted_img = add_motion_blur_single_image(img,
args.blur_parameters[0],
args.blur_parameters[1])
imageio.imwrite(os.path.join(args.corruptions_dir, img_name),
corrupted_img)
#imageio.imwrite(os.path.join(args.masks_dir, img_name), mask * 255)
sess.run(
tf.variables_initializer([latent_code_2] + optimizer.variables()))
progress_bar_iterator = tqdm(
iterable=range(args.total_iterations),
bar_format=
'{desc}: {percentage:3.0f}% |{bar}| {n_fmt}/{total_fmt}{postfix}',
desc=img_name)
for i in progress_bar_iterator:
loss, _ = sess.run(fetches=[loss_op, train_op],
feed_dict={
original_img: img[np.newaxis, ...],
blur_kernel: blur_kernel_3d[np.newaxis, ...]
})
progress_bar_iterator.set_postfix_str('loss=%.2f' % loss)
reconstructed_imgs, latent_codes = sess.run(
fetches=[generated_img_for_display, latent_code_2],
feed_dict={
original_img: img[np.newaxis, ...],
blur_kernel: blur_kernel_3d[np.newaxis, ...]
})
imageio.imwrite(os.path.join(args.restorations_dir, img_name),
reconstructed_imgs[0])
np.savez(file=os.path.join(args.latents_dir, img_name + '.npz'),
latent_code=latent_codes[0])
return latent_codes[0]
def optimize_latent_codes(args):
tflib.init_tf()
with dnnlib.util.open_url(STYLEGAN_MODEL_URL, cache_dir=config.cache_dir) as f:
_G, _D, Gs = pickle.load(f)
latent_code = tf.get_variable(
name='latent_code', shape=(1, 18, 512), dtype='float32', initializer=tf.initializers.zeros()
)
generated_img = Gs.components.synthesis.get_output_for(latent_code, randomize_noise=False)
generated_img = tf.transpose(generated_img, [0, 2, 3, 1])
generated_img = ((generated_img + 1) / 2) * 255
generated_img_for_display = tf.saturate_cast(generated_img, tf.uint8)
target_img = tf.placeholder(tf.float32, [None, args.input_img_size[0], args.input_img_size[1], 3])
target_img_resized = tf.image.resize_images(
target_img, tuple(args.perceptual_img_size), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR
)
generated_img_resized = tf.image.resize_images(
generated_img, tuple(args.perceptual_img_size), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR
)
perceptual_model = PerceptualModel(img_size=args.perceptual_img_size)
generated_img_features = perceptual_model(generated_img_resized)
target_img_features = perceptual_model(target_img_resized)
loss_op = tf.reduce_mean(tf.abs(generated_img_features - target_img_features))
if args.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
else:
optimizer = tf.train.GradientDescentOptimizer(learning_rate=args.learning_rate)
train_op = optimizer.minimize(loss_op, var_list=[latent_code])
sess = tf.get_default_session()
img_names = sorted(os.listdir(args.imgs_dir))
for img_name in img_names:
img = imageio.imread(os.path.join(args.imgs_dir, img_name))
sess.run(tf.variables_initializer([latent_code] + optimizer.variables()))
progress_bar_iterator = tqdm(
iterable=range(args.total_iterations),
bar_format='{desc}: {percentage:3.0f}% |{bar}| {n_fmt}/{total_fmt}{postfix}',
desc=img_name
)
for i in progress_bar_iterator:
loss, _ = sess.run(
fetches=[loss_op, train_op],
feed_dict={
target_img: img[np.newaxis, ...]
}
)
progress_bar_iterator.set_postfix_str('loss=%.2f' % loss)
reconstructed_imgs, latent_codes = sess.run(
fetches=[generated_img_for_display, latent_code],
feed_dict={
target_img: img[np.newaxis, ...]
}
)
imageio.imwrite(os.path.join(args.reconstructions_dir, img_name), reconstructed_imgs[0])
np.savez(file=os.path.join(args.latents_dir, img_name + '.npz'), latent_code=latent_codes[0])
def training_loop(
submit_config,
Encoder_args = {},
E_opt_args = {},
D_opt_args = {},
E_loss_args = EasyDict(),
D_loss_args = {},
lr_args = EasyDict(),
tf_config = {},
dataset_args = EasyDict(),
decoder_pkl = EasyDict(),
drange_data = [0, 255],
drange_net = [-1,1], # Dynamic range used when feeding image data to the networks.
mirror_augment = False,
resume_run_id = config.ENCODER_PICKLE_DIR, # Run ID or network pkl to resume training from, None = start from scratch.
resume_snapshot = None, # Snapshot index to resume training from, None = autodetect.
image_snapshot_ticks = 1, # How often to export image snapshots?
network_snapshot_ticks = 4, # How often to export network snapshots?
max_iters = 150000):
tflib.init_tf(tf_config)
with tf.name_scope('input'):
real_train = tf.placeholder(tf.float32, [submit_config.batch_size, 3, submit_config.image_size, submit_config.image_size], name='real_image_train')
real_test = tf.placeholder(tf.float32, [submit_config.batch_size_test, 3, submit_config.image_size, submit_config.image_size], name='real_image_test')
real_split = tf.split(real_train, num_or_size_splits=submit_config.num_gpus, axis=0)
with tf.device('/gpu:0'):
if resume_run_id is not None:
network_pkl = misc.locate_network_pkl(resume_run_id, resume_snapshot)
print('Loading networks from "%s"...' % network_pkl)
E, G, D, Gs = misc.load_pkl(network_pkl)
start = int(network_pkl.split('-')[-1].split('.')[0]) // submit_config.batch_size
print('Start: ', start)
else:
print('Constructing networks...')
G, D, Gs = misc.load_pkl(decoder_pkl.decoder_pkl)
num_layers = Gs.components.synthesis.input_shape[1]
E = tflib.Network('E', size=submit_config.image_size, filter=64, filter_max=1024, num_layers=num_layers, phase=True, **Encoder_args)
start = 0
E.print_layers(); Gs.print_layers(); D.print_layers()
global_step0 = tf.Variable(start, trainable=False, name='learning_rate_step')
learning_rate = tf.train.exponential_decay(lr_args.learning_rate, global_step0, lr_args.decay_step,
lr_args.decay_rate, staircase=lr_args.stair)
add_global0 = global_step0.assign_add(1)
E_opt = tflib.Optimizer(name='TrainE', learning_rate=learning_rate, **E_opt_args)
D_opt = tflib.Optimizer(name='TrainD', learning_rate=learning_rate, **D_opt_args)
E_loss_rec = 0.
E_loss_adv = 0.
D_loss_real = 0.
D_loss_fake = 0.
D_loss_grad = 0.
for gpu in range(submit_config.num_gpus):
print('build graph on gpu %s' % str(gpu))
with tf.name_scope('GPU%d' % gpu), tf.device('/gpu:%d' % gpu):
E_gpu = E if gpu == 0 else E.clone(E.name + '_shadow')
D_gpu = D if gpu == 0 else D.clone(D.name + '_shadow')
G_gpu = Gs if gpu == 0 else Gs.clone(Gs.name + '_shadow')
perceptual_model = PerceptualModel(img_size=[E_loss_args.perceptual_img_size, E_loss_args.perceptual_img_size], multi_layers=False)
real_gpu = process_reals(real_split[gpu], mirror_augment, drange_data, drange_net)
with tf.name_scope('E_loss'), tf.control_dependencies(None):
E_loss, recon_loss, adv_loss = dnnlib.util.call_func_by_name(E=E_gpu, G=G_gpu, D=D_gpu, perceptual_model=perceptual_model, reals=real_gpu, **E_loss_args)
E_loss_rec += recon_loss
E_loss_adv += adv_loss
with tf.name_scope('D_loss'), tf.control_dependencies(None):
D_loss, loss_fake, loss_real, loss_gp = dnnlib.util.call_func_by_name(E=E_gpu, G=G_gpu, D=D_gpu, reals=real_gpu, **D_loss_args)
D_loss_real += loss_real
D_loss_fake += loss_fake
D_loss_grad += loss_gp
with tf.control_dependencies([add_global0]):
E_opt.register_gradients(E_loss, E_gpu.trainables)
D_opt.register_gradients(D_loss, D_gpu.trainables)
E_loss_rec /= submit_config.num_gpus
E_loss_adv /= submit_config.num_gpus
D_loss_real /= submit_config.num_gpus
D_loss_fake /= submit_config.num_gpus
D_loss_grad /= submit_config.num_gpus
E_train_op = E_opt.apply_updates()
D_train_op = D_opt.apply_updates()
print('building testing graph...')
fake_X_val = test(E, Gs, real_test, submit_config)
sess = tf.get_default_session()
print('Getting training data...')
image_batch_train = get_train_data(sess, data_dir=dataset_args.data_train, submit_config=submit_config, mode='train')
image_batch_test = get_train_data(sess, data_dir=dataset_args.data_test, submit_config=submit_config, mode='test')
summary_log = tf.summary.FileWriter(config.GDRIVE_PATH)
cur_nimg = start * submit_config.batch_size
cur_tick = 0
tick_start_nimg = cur_nimg
start_time = time.time()
init_pascal = tf.initialize_variables(
[global_step0],
name='init_pascal'
)
sess.run(init_pascal)
print('Optimization starts!!!')
for it in range(start, max_iters):
batch_images = sess.run(image_batch_train)
feed_dict_1 = {real_train: batch_images}
_, recon_, adv_ = sess.run([E_train_op, E_loss_rec, E_loss_adv], feed_dict_1)
_, d_r_, d_f_, d_g_ = sess.run([D_train_op, D_loss_real, D_loss_fake, D_loss_grad], feed_dict_1)
cur_nimg += submit_config.batch_size
if it % 50 == 0:
print('Iter: %06d recon_loss: %-6.4f adv_loss: %-6.4f d_r_loss: %-6.4f d_f_loss: %-6.4f d_reg: %-6.4f time:%-12s' % (
it, recon_, adv_, d_r_, d_f_, d_g_, dnnlib.util.format_time(time.time() - start_time)))
sys.stdout.flush()
tflib.autosummary.save_summaries(summary_log, it)
if it % 500 == 0:
batch_images_test = sess.run(image_batch_test)
batch_images_test = misc.adjust_dynamic_range(batch_images_test.astype(np.float32), [0, 255], [-1., 1.])
samples2 = sess.run(fake_X_val, feed_dict={real_test: batch_images_test})
orin_recon = np.concatenate([batch_images_test, samples2], axis=0)
orin_recon = adjust_pixel_range(orin_recon)
orin_recon = fuse_images(orin_recon, row=2, col=submit_config.batch_size_test)
# save image results during training, first row is original images and the second row is reconstructed images
save_image('%s/iter_%08d.png' % (submit_config.run_dir, cur_nimg), orin_recon)
# save image to gdrive
img_path = os.path.join(config.GDRIVE_PATH, 'images', ('iter_%08d.png' % (cur_nimg)))
save_image(img_path, orin_recon)
if cur_nimg >= tick_start_nimg + 65000:
cur_tick += 1
tick_start_nimg = cur_nimg
if cur_tick % network_snapshot_ticks == 0:
pkl = os.path.join(submit_config.run_dir, 'network-snapshot-%08d.pkl' % (cur_nimg))
misc.save_pkl((E, G, D, Gs), pkl)
# save network snapshot to gdrive
pkl_drive = os.path.join(config.GDRIVE_PATH, 'snapshots', 'network-snapshot-%08d.pkl' % (cur_nimg))
misc.save_pkl((E, G, D, Gs), pkl_drive)
misc.save_pkl((E, G, D, Gs), os.path.join(submit_config.run_dir, 'network-final.pkl'))
summary_log.close()
def training_loop(
submit_config,
Encoder_args={},
E_opt_args={},
D_opt_args={},
E_loss_args={},
D_loss_args={},
lr_args=EasyDict(),
tf_config={},
dataset_args=EasyDict(),
decoder_pkl=EasyDict(),
drange_data=[0, 255],
drange_net=[
-1, 1
], # Dynamic range used when feeding image data to the networks.
mirror_augment=False,
resume_run_id=None, # Run ID or network pkl to resume training from, None = start from scratch.
resume_snapshot=None, # Snapshot index to resume training from, None = autodetect.
image_snapshot_ticks=1, # How often to export image snapshots?
network_snapshot_ticks=10, # How often to export network snapshots?
save_tf_graph=False, # Include full TensorFlow computation graph in the tfevents file?
save_weight_histograms=False, # Include weight histograms in the tfevents file?
max_iters=150000,
E_smoothing=0.999):
tflib.init_tf(tf_config)
with tf.name_scope('input'):
real_train = tf.placeholder(tf.float32, [
submit_config.batch_size, 3, submit_config.image_size,
submit_config.image_size
],
name='real_image_train')
real_test = tf.placeholder(tf.float32, [
submit_config.batch_size_test, 3, submit_config.image_size,
submit_config.image_size
],
name='real_image_test')
real_split = tf.split(real_train,
num_or_size_splits=submit_config.num_gpus,
axis=0)
with tf.device('/gpu:0'):
if resume_run_id is not None:
network_pkl = misc.locate_network_pkl(resume_run_id,
resume_snapshot)
print('Loading networks from "%s"...' % network_pkl)
E, G, D, Gs, NE = misc.load_pkl(network_pkl)
start = int(network_pkl.split('-')[-1].split('.')
[0]) // submit_config.batch_size
else:
print('Constructing networks...')
G, D, Gs, NE = misc.load_pkl(decoder_pkl.decoder_pkl)
E = tflib.Network('E',
size=submit_config.image_size,
filter=64,
filter_max=1024,
phase=True,
**Encoder_args)
start = 0
Gs.print_layers()
E.print_layers()
D.print_layers()
global_step = tf.Variable(start,
trainable=False,
name='learning_rate_step')
learning_rate = tf.train.exponential_decay(lr_args.learning_rate,
global_step,
lr_args.decay_step,
lr_args.decay_rate,
staircase=lr_args.stair)
add_global = global_step.assign_add(1)
E_opt = tflib.Optimizer(name='TrainE',
learning_rate=learning_rate,
**E_opt_args)
D_opt = tflib.Optimizer(name='TrainD',
learning_rate=learning_rate,
**D_opt_args)
E_loss_rec = 0.
E_loss_adv = 0.
D_loss_real = 0.
D_loss_fake = 0.
D_loss_grad = 0.
for gpu in range(submit_config.num_gpus):
print('build graph on gpu %s' % str(gpu))
with tf.name_scope('GPU%d' % gpu), tf.device('/gpu:%d' % gpu):
E_gpu = E if gpu == 0 else E.clone(E.name + '_shadow')
D_gpu = D if gpu == 0 else D.clone(D.name + '_shadow')
G_gpu = Gs if gpu == 0 else Gs.clone(Gs.name + '_shadow')
perceptual_model = PerceptualModel(
img_size=[submit_config.image_size, submit_config.image_size],
multi_layers=False)
real_gpu = process_reals(real_split[gpu], mirror_augment,
drange_data, drange_net)
with tf.name_scope('E_loss'), tf.control_dependencies(None):
E_loss, recon_loss, adv_loss = dnnlib.util.call_func_by_name(
E=E_gpu,
G=G_gpu,
D=D_gpu,
perceptual_model=perceptual_model,
reals=real_gpu,
**E_loss_args)
E_loss_rec += recon_loss
E_loss_adv += adv_loss
with tf.name_scope('D_loss'), tf.control_dependencies(None):
D_loss, loss_fake, loss_real, loss_gp = dnnlib.util.call_func_by_name(
E=E_gpu, G=G_gpu, D=D_gpu, reals=real_gpu, **D_loss_args)
D_loss_real += loss_real
D_loss_fake += loss_fake
D_loss_grad += loss_gp
with tf.control_dependencies([add_global]):
E_opt.register_gradients(E_loss, E_gpu.trainables)
D_opt.register_gradients(D_loss, D_gpu.trainables)
E_loss_rec /= submit_config.num_gpus
E_loss_adv /= submit_config.num_gpus
D_loss_real /= submit_config.num_gpus
D_loss_fake /= submit_config.num_gpus
D_loss_grad /= submit_config.num_gpus
E_train_op = E_opt.apply_updates()
D_train_op = D_opt.apply_updates()
#Es_update_op = Es.setup_as_moving_average_of(E, beta=E_smoothing)
print('building testing graph...')
fake_X_val = test(E, Gs, real_test, submit_config)
sess = tf.get_default_session()
print('Getting training data...')
image_batch_train = get_train_data(sess,
data_dir=dataset_args.data_train,
submit_config=submit_config,
mode='train')
image_batch_test = get_train_data(sess,
data_dir=dataset_args.data_test,
submit_config=submit_config,
mode='test')
summary_log = tf.summary.FileWriter(submit_config.run_dir)
if save_tf_graph:
summary_log.add_graph(tf.get_default_graph())
if save_weight_histograms:
E.setup_weight_histograms()
D.setup_weight_histograms()
cur_nimg = start * submit_config.batch_size
cur_tick = 0
tick_start_nimg = cur_nimg
start_time = time.time()
print('Optimization starts!!!')
for it in range(start, max_iters):
feed_dict = {real_train: sess.run(image_batch_train)}
sess.run([E_train_op, E_loss_rec, E_loss_adv], feed_dict)
sess.run([D_train_op, D_loss_real, D_loss_fake, D_loss_grad],
feed_dict)
cur_nimg += submit_config.batch_size
if it % 100 == 0:
print("Iter: %06d kimg: %-8.1f time: %-12s" %
(it, cur_nimg / 1000,
dnnlib.util.format_time(time.time() - start_time)))
sys.stdout.flush()
tflib.autosummary.save_summaries(summary_log, it)
if cur_nimg >= tick_start_nimg + 65000:
cur_tick += 1
tick_start_nimg = cur_nimg
if cur_tick % image_snapshot_ticks == 0:
batch_images_test = sess.run(image_batch_test)
batch_images_test = misc.adjust_dynamic_range(
batch_images_test.astype(np.float32), [0, 255], [-1., 1.])
samples2 = sess.run(fake_X_val,
feed_dict={real_test: batch_images_test})
samples2 = samples2.transpose(0, 2, 3, 1)
batch_images_test = batch_images_test.transpose(0, 2, 3, 1)
orin_recon = np.concatenate([batch_images_test, samples2],
axis=0)
imwrite(immerge(orin_recon, 2, submit_config.batch_size_test),
'%s/iter_%08d.png' % (submit_config.run_dir, cur_nimg))
if cur_tick % network_snapshot_ticks == 0:
pkl = os.path.join(submit_config.run_dir,
'network-snapshot-%08d.pkl' % (cur_nimg))
misc.save_pkl((E, G, D, Gs, NE), pkl)
misc.save_pkl((E, G, D, Gs, NE),
os.path.join(submit_config.run_dir, 'network-final.pkl'))
summary_log.close()
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 training_loop(
submit_config,
Encoder_args = {},
E_opt_args = {},
D_opt_args = {},
E_loss_args = EasyDict(),
D_loss_args = {},
lr_args = EasyDict(),
tf_config = {},
dataset_args = EasyDict(),
decoder_pkl = EasyDict(),
inversion_pkl = EasyDict(),
drange_data = [0, 255],
drange_net = [-1,1], # Dynamic range used when feeding image data to the networks.
mirror_augment = False,
resume_run_id = config.ENCODER_PICKLE_DIR, # Run ID or network pkl to resume training from, None = start from scratch.
resume_snapshot = None, # Snapshot index to resume training from, None = autodetect.
image_snapshot_ticks = 1, # How often to export image snapshots?
network_snapshot_ticks = 4, # How often to export network snapshots?
max_iters = 150000):
tflib.init_tf(tf_config)
with tf.name_scope('input'):
placeholder_real_portraits_train = tf.placeholder(tf.float32, [submit_config.batch_size, 3, submit_config.image_size, submit_config.image_size], name='placeholder_real_portraits_train')
placeholder_real_landmarks_train = tf.placeholder(tf.float32, [submit_config.batch_size, 3, submit_config.image_size, submit_config.image_size], name='placeholder_real_landmarks_train')
placeholder_real_shuffled_train = tf.placeholder(tf.float32, [submit_config.batch_size, 3, submit_config.image_size, submit_config.image_size], name='placeholder_real_shuffled_train')
placeholder_landmarks_shuffled_train = tf.placeholder(tf.float32, [submit_config.batch_size, 3, submit_config.image_size, submit_config.image_size], name='placeholder_landmarks_shuffled_train')
placeholder_real_portraits_test = tf.placeholder(tf.float32, [submit_config.batch_size_test, 3, submit_config.image_size, submit_config.image_size], name='placeholder_real_portraits_test')
placeholder_real_landmarks_test = tf.placeholder(tf.float32, [submit_config.batch_size_test, 3, submit_config.image_size, submit_config.image_size], name='placeholder_real_landmarks_test')
placeholder_real_shuffled_test = tf.placeholder(tf.float32, [submit_config.batch_size_test, 3, submit_config.image_size, submit_config.image_size], name='placeholder_real_shuffled_test')
placeholder_real_landmarks_shuffled_test = tf.placeholder(tf.float32, [submit_config.batch_size_test, 3, submit_config.image_size, submit_config.image_size], name='placeholder_real_landmarks_shuffled_test')
real_split_landmarks = tf.split(placeholder_real_landmarks_train, num_or_size_splits=submit_config.num_gpus, axis=0)
real_split_portraits = tf.split(placeholder_real_portraits_train, num_or_size_splits=submit_config.num_gpus, axis=0)
real_split_shuffled = tf.split(placeholder_real_shuffled_train, num_or_size_splits=submit_config.num_gpus, axis=0)
real_split_lm_shuffled = tf.split(placeholder_landmarks_shuffled_train, num_or_size_splits=submit_config.num_gpus, axis=0)
placeholder_training_flag = tf.placeholder(tf.string, name='placeholder_training_flag')
with tf.device('/gpu:0'):
if resume_run_id is not None:
network_pkl = misc.locate_network_pkl(resume_run_id, resume_snapshot)
print('Loading networks from "%s"...' % network_pkl)
E, G, D, Gs = misc.load_pkl(network_pkl)
start = int(network_pkl.split('-')[-1].split('.')[0]) // submit_config.batch_size
print('Start: ', start)
else:
print('Constructing networks...')
G, _, Gs = misc.load_pkl(decoder_pkl.decoder_pkl) # don't use pre-trained discriminator!
num_layers = Gs.components.synthesis.input_shape[1]
# here we add a new discriminator!
D = tflib.Network('D', # name of the network how we call it
num_channels=3, resolution=128, label_size=0, #some needed for this build function
func_name="training.networks_stylegan.D_basic") # function of that network. more was not passed in d_args!
# input is not passed here (just construction - note that we do not call the actual function!). Instead, network will inspect build function and require it for the get_output_for function.
print("Created new Discriminator!")
E = tflib.Network('E', size=submit_config.image_size, filter=64, filter_max=1024, num_layers=num_layers, phase=True, **Encoder_args)
start = 0
Inv, _, _, _ = misc.load_pkl(inversion_pkl.inversion_pkl)
E.print_layers(); Gs.print_layers(); D.print_layers()
global_step0 = tf.Variable(start, trainable=False, name='learning_rate_step')
learning_rate = tf.train.exponential_decay(lr_args.learning_rate, global_step0, lr_args.decay_step,
lr_args.decay_rate, staircase=lr_args.stair)
add_global0 = global_step0.assign_add(1)
E_opt = tflib.Optimizer(name='TrainE', learning_rate=learning_rate, **E_opt_args)
D_opt = tflib.Optimizer(name='TrainD', learning_rate=learning_rate, **D_opt_args)
E_loss_rec = 0.
E_loss_adv = 0.
D_loss_real = 0.
D_loss_fake = 0.
D_loss_grad = 0.
for gpu in range(submit_config.num_gpus):
print('build graph on gpu %s' % str(gpu))
with tf.name_scope('GPU%d' % gpu), tf.device('/gpu:%d' % gpu):
E_gpu = E if gpu == 0 else E.clone(E.name + '_shadow')
D_gpu = D if gpu == 0 else D.clone(D.name + '_shadow')
G_gpu = Gs if gpu == 0 else Gs.clone(Gs.name + '_shadow')
Inv_gpu = Inv if gpu == 0 else Inv.clone(Inv.name + '_shadow')
perceptual_model = PerceptualModel(img_size=[E_loss_args.perceptual_img_size, E_loss_args.perceptual_img_size], multi_layers=False)
real_portraits_gpu = process_reals(real_split_portraits[gpu], mirror_augment, drange_data, drange_net)
shuffled_portraits_gpu = process_reals(real_split_shuffled[gpu], mirror_augment, drange_data, drange_net)
real_landmarks_gpu = process_reals(real_split_landmarks[gpu], mirror_augment, drange_data, drange_net)
shuffled_landmarks_gpu = process_reals(real_split_lm_shuffled[gpu], mirror_augment, drange_data, drange_net)
with tf.name_scope('E_loss'), tf.control_dependencies(None):
E_loss, recon_loss, adv_loss = dnnlib.util.call_func_by_name(E=E_gpu, G=G_gpu, D=D_gpu, Inv=Inv_gpu, perceptual_model=perceptual_model, real_portraits=real_portraits_gpu, shuffled_portraits=shuffled_portraits_gpu, real_landmarks=real_landmarks_gpu, shuffled_landmarks=shuffled_landmarks_gpu, training_flag=placeholder_training_flag, **E_loss_args)
E_loss_rec += recon_loss
E_loss_adv += adv_loss
with tf.name_scope('D_loss'), tf.control_dependencies(None):
D_loss, loss_fake, loss_real, loss_gp = dnnlib.util.call_func_by_name(E=E_gpu, G=G_gpu, D=D_gpu, Inv=Inv_gpu, real_portraits=real_portraits_gpu, shuffled_portraits=shuffled_portraits_gpu, real_landmarks=real_landmarks_gpu, training_flag=placeholder_training_flag, **D_loss_args) # change signature in ...
D_loss_real += loss_real
D_loss_fake += loss_fake
D_loss_grad += loss_gp
with tf.control_dependencies([add_global0]):
E_opt.register_gradients(E_loss, E_gpu.trainables)
D_opt.register_gradients(D_loss, D_gpu.trainables)
E_loss_rec /= submit_config.num_gpus
E_loss_adv /= submit_config.num_gpus
D_loss_real /= submit_config.num_gpus
D_loss_fake /= submit_config.num_gpus
D_loss_grad /= submit_config.num_gpus
E_train_op = E_opt.apply_updates()
D_train_op = D_opt.apply_updates()
print('building testing graph...')
fake_X_val = test(E, Gs, Inv, placeholder_real_portraits_test, placeholder_real_landmarks_test, placeholder_real_shuffled_test, submit_config)
inv_X_val = test_inversion(E, Gs, Inv, placeholder_real_portraits_test, placeholder_real_landmarks_test, placeholder_real_shuffled_test, submit_config)
#sampled_portraits_val = sample_random_portraits(Gs, submit_config.batch_size)
#sampled_portraits_val_test = sample_random_portraits(Gs, submit_config.batch_size_test)
sess = tf.get_default_session()
print('Getting training data...')
# x_batch is a batch of (2, ..., ..., ...) records!
stack_batch_train = get_train_data(sess, data_dir=dataset_args.data_train, submit_config=submit_config, mode='train')
stack_batch_test = get_train_data(sess, data_dir=dataset_args.data_test, submit_config=submit_config, mode='test')
stack_batch_train_secondary = get_train_data(sess, data_dir=dataset_args.data_train, submit_config=submit_config, mode='train_secondary')
stack_batch_test_secondary = get_train_data(sess, data_dir=dataset_args.data_test, submit_config=submit_config, mode='test_secondary')
summary_log = tf.summary.FileWriter(config.getGdrivePath())
cur_nimg = start * submit_config.batch_size
cur_tick = 0
tick_start_nimg = cur_nimg
start_time = time.time()
init_fix = tf.initialize_variables(
[global_step0],
name='init_fix'
)
sess.run(init_fix)
print('Optimization starts!!!')
# here is the actual training loop: all iterations
for it in range(start, max_iters):
batch_stacks = sess.run(stack_batch_train)
batch_portraits = batch_stacks[:,0,:,:,:]
batch_landmarks = batch_stacks[:,1,:,:,:]
batch_stacks_secondary = sess.run(stack_batch_train_secondary)
batch_shuffled = batch_stacks_secondary[:,0,:,:,:]
batch_lm_shuffled = batch_stacks_secondary[:,1,:,:,:]
training_flag = "pose"
feed_dict_1 = {placeholder_real_portraits_train: batch_portraits, placeholder_real_landmarks_train: batch_landmarks, placeholder_real_shuffled_train:batch_shuffled, placeholder_landmarks_shuffled_train:batch_lm_shuffled, placeholder_training_flag: training_flag}
# here we query these encoder- and discriminator losses. as input we provide: batch_stacks = batch of images + landmarks.
_, recon_, adv_ = sess.run([E_train_op, E_loss_rec, E_loss_adv], feed_dict_1)
_, d_r_, d_f_, d_g_= sess.run([D_train_op, D_loss_real, D_loss_fake, D_loss_grad], feed_dict_1)
cur_nimg += submit_config.batch_size
if it % 50 == 0:
print('Iter: %06d recon_loss: %-6.4f adv_loss: %-6.4f d_r_loss: %-6.4f d_f_loss: %-6.4f d_reg: %-6.4f time:%-12s' % (
it, recon_, adv_, d_r_, d_f_, d_g_, dnnlib.util.format_time(time.time() - start_time)))
sys.stdout.flush()
tflib.autosummary.save_summaries(summary_log, it)
if it % 500 == 0:
batch_stacks_test = sess.run(stack_batch_test)
batch_portraits_test = batch_stacks_test[:,0,:,:,:]
batch_landmarks_test = batch_stacks_test[:,1,:,:,:]
batch_stacks_test_secondary = sess.run(stack_batch_test_secondary)
batch_shuffled_test = batch_stacks_test_secondary[:,0,:,:,:]
batch_shuffled_lm_test = batch_stacks_test_secondary[:,1,:,:,:]
batch_portraits_test = misc.adjust_dynamic_range(batch_portraits_test.astype(np.float32), [0, 255], [-1., 1.])
batch_landmarks_test = misc.adjust_dynamic_range(batch_landmarks_test.astype(np.float32), [0, 255], [-1., 1.])
batch_shuffled_test = misc.adjust_dynamic_range(batch_shuffled_test.astype(np.float32), [0, 255], [-1., 1.])
batch_shuffled_lm_test = misc.adjust_dynamic_range(batch_shuffled_lm_test.astype(np.float32), [0, 255], [-1., 1.])
# first: input + target landmarks = manipulated image
samples_manipulated = sess.run(fake_X_val, feed_dict={placeholder_real_portraits_test: batch_portraits_test, placeholder_real_landmarks_test: batch_shuffled_lm_test})
# 2nd: manipulated + original landmarks
samples_reconstructed = sess.run(fake_X_val, feed_dict={placeholder_real_portraits_test: samples_manipulated, placeholder_real_landmarks_test: batch_landmarks_test})
# also: show direct reconstruction
samples_direct_rec = sess.run(fake_X_val, feed_dict={placeholder_real_portraits_test: batch_portraits_test, placeholder_real_landmarks_test: batch_landmarks_test})
# show results of the inverison
portraits_inverted = sess.run(inv_X_val, feed_dict={placeholder_real_portraits_test: batch_portraits_test, placeholder_real_landmarks_test: batch_landmarks_test})
# show: original portrait, original landmark, diret reconstruction, fake landmark, manipulated, rec.
debug_img = np.concatenate([
batch_landmarks_test, # original landmarks
batch_portraits_test, # original portraits,
samples_direct_rec, # direct
batch_shuffled_lm_test, # shuffled landmarks
samples_manipulated, # manipulated images
samples_reconstructed,
portraits_inverted# cycle reconstructed images
], axis=0)
debug_img = adjust_pixel_range(debug_img)
debug_img = fuse_images(debug_img, row=6, col=submit_config.batch_size_test)
# save image results during training, first row is original images and the second row is reconstructed images
save_image('%s/iter_%08d.png' % (submit_config.run_dir, cur_nimg), debug_img)
# save image to gdrive
img_path = os.path.join(config.getGdrivePath(), 'images', ('iter_%08d.png' % (cur_nimg)))
save_image(img_path, debug_img)
if cur_nimg >= tick_start_nimg + 65000:
cur_tick += 1
tick_start_nimg = cur_nimg
if cur_tick % network_snapshot_ticks == 0:
pkl = os.path.join(submit_config.run_dir, 'network-snapshot-%08d.pkl' % (cur_nimg))
misc.save_pkl((E, G, D, Gs), pkl)
# save network snapshot to gdrive
pkl_drive = os.path.join(config.getGdrivePath(), 'snapshots', 'network-snapshot-%08d.pkl' % (cur_nimg))
misc.save_pkl((E, G, D, Gs), pkl_drive)
misc.save_pkl((E, G, D, Gs), os.path.join(submit_config.run_dir, 'network-final.pkl'))
summary_log.close()
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 training_loop(
submit_config,
Encoder_args={},
D_args={},
G_args={},
E_opt_args={},
D_opt_args={},
E_loss_args=EasyDict(),
D_loss_args={},
lr_args=EasyDict(),
tf_config={},
dataset_args=EasyDict(),
decoder_pkl=EasyDict(),
drange_data=[0, 255],
drange_net=[
-1, 1
], # Dynamic range used when feeding image data to the networks.
mirror_augment=False,
filter=64, # Minimum number of feature maps in any layer.
filter_max=512, # Maximum number of feature maps in any layer.
resume_run_id=None, # Run ID or network pkl to resume training from, None = start from scratch.
resume_snapshot=None, # Snapshot index to resume training from, None = autodetect.
image_snapshot_ticks=1, # How often to export image snapshots?
network_snapshot_ticks=10, # How often to export network snapshots?
d_scale=0.1, # Decide whether to update discriminator.
pretrained_D=True, # Whether to use pre trained Discriminator.
max_iters=150000):
tflib.init_tf(tf_config)
with tf.name_scope('Input'):
real_train = tf.placeholder(tf.float32, [
submit_config.batch_size, 3, submit_config.image_size,
submit_config.image_size
],
name='real_image_train')
real_test = tf.placeholder(tf.float32, [
submit_config.batch_size_test, 3, submit_config.image_size,
submit_config.image_size
],
name='real_image_test')
real_split = tf.split(real_train,
num_or_size_splits=submit_config.num_gpus,
axis=0)
with tf.device('/gpu:0'):
if resume_run_id is not None:
network_pkl = misc.locate_network_pkl(resume_run_id,
resume_snapshot)
print('Loading networks from "%s"...' % network_pkl)
E, G, D, Gs = misc.load_pkl(network_pkl)
G_style_mod = tflib.Network('G_StyleMod',
resolution=submit_config.image_size,
label_size=0,
**G_args)
start = int(network_pkl.split('-')[-1].split('.')
[0]) // submit_config.batch_size
print('Start: ', start)
else:
print('Constructing networks...')
G, PreD, Gs = misc.load_pkl(decoder_pkl.decoder_pkl)
num_layers = Gs.components.synthesis.input_shape[1]
E = tflib.Network('E_gpu0',
size=submit_config.image_size,
filter=filter,
filter_max=filter_max,
num_layers=num_layers,
is_training=True,
num_gpus=submit_config.num_gpus,
**Encoder_args)
OriD = tflib.Network('D_ori',
resolution=submit_config.image_size,
label_size=0,
**D_args)
G_style_mod = tflib.Network('G_StyleMod',
resolution=submit_config.image_size,
label_size=0,
**G_args)
if pretrained_D:
D = PreD
else:
D = OriD
start = 0
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]})
E.print_layers()
Gs.print_layers()
D.print_layers()
global_step0 = tf.Variable(start,
trainable=False,
name='learning_rate_step')
learning_rate = tf.train.exponential_decay(lr_args.learning_rate,
global_step0,
lr_args.decay_step,
lr_args.decay_rate,
staircase=lr_args.stair)
add_global0 = global_step0.assign_add(1)
E_opt = tflib.Optimizer(name='TrainE',
learning_rate=learning_rate,
**E_opt_args)
if d_scale > 0:
D_opt = tflib.Optimizer(name='TrainD',
learning_rate=learning_rate,
**D_opt_args)
E_loss_rec = 0.
E_loss_adv = 0.
D_loss_real = 0.
D_loss_fake = 0.
D_loss_grad = 0.
for gpu in range(submit_config.num_gpus):
print('Building Graph on GPU %s' % str(gpu))
with tf.name_scope('GPU%d' % gpu), tf.device('/gpu:%d' % gpu):
E_gpu = E if gpu == 0 else E.clone(E.name[:-1] + str(gpu))
D_gpu = D if gpu == 0 else D.clone(D.name + '_shadow')
G_gpu = G_style_mod if gpu == 0 else G_style_mod.clone(
G_style_mod.name + '_shadow')
feature_model = PerceptualModel(img_size=[
E_loss_args.perceptual_img_size,
E_loss_args.perceptual_img_size
],
multi_layers=False)
real_gpu = process_reals(real_split[gpu], mirror_augment,
drange_data, drange_net)
with tf.name_scope('E_loss'), tf.control_dependencies(None):
E_loss, recon_loss, adv_loss = dnnlib.util.call_func_by_name(
E=E_gpu,
G=G_gpu,
D=D_gpu,
feature_model=feature_model,
reals=real_gpu,
**E_loss_args)
E_loss_rec += recon_loss
E_loss_adv += adv_loss
with tf.name_scope('D_loss'), tf.control_dependencies(None):
D_loss, loss_fake, loss_real, loss_gp = dnnlib.util.call_func_by_name(
E=E_gpu, G=G_gpu, D=D_gpu, reals=real_gpu, **D_loss_args)
D_loss_real += loss_real
D_loss_fake += loss_fake
D_loss_grad += loss_gp
with tf.control_dependencies([add_global0]):
E_opt.register_gradients(E_loss, E_gpu.trainables)
if d_scale > 0:
D_opt.register_gradients(D_loss, D_gpu.trainables)
E_loss_rec /= submit_config.num_gpus
E_loss_adv /= submit_config.num_gpus
D_loss_real /= submit_config.num_gpus
D_loss_fake /= submit_config.num_gpus
D_loss_grad /= submit_config.num_gpus
E_train_op = E_opt.apply_updates()
if d_scale > 0:
D_train_op = D_opt.apply_updates()
print('Building testing graph...')
fake_X_val = test(E, G_style_mod, real_test, submit_config)
sess = tf.get_default_session()
print('Getting training data...')
image_batch_train = get_train_data(sess,
data_dir=dataset_args.data_train,
submit_config=submit_config,
mode='train')
image_batch_test = get_train_data(sess,
data_dir=dataset_args.data_test,
submit_config=submit_config,
mode='test')
summary_log = tf.summary.FileWriter(submit_config.run_dir)
cur_nimg = start * submit_config.batch_size
cur_tick = 0
tick_start_nimg = cur_nimg
start_time = time.time()
print('Optimization starts!!!')
for it in range(start, max_iters):
batch_images = sess.run(image_batch_train)
feed_dict = {real_train: batch_images}
_, recon_, adv_, lr = sess.run(
[E_train_op, E_loss_rec, E_loss_adv, learning_rate], feed_dict)
if d_scale > 0:
_, d_r_, d_f_, d_g_ = sess.run(
[D_train_op, D_loss_real, D_loss_fake, D_loss_grad], feed_dict)
cur_nimg += submit_config.batch_size
run_time = time.time() - start_time
iter_time = run_time / (it - start + 1)
eta_time = iter_time * (max_iters - it - 1)
if it % 50 == 0:
print(
'Iter: %06d/%d, lr: %-.8f recon_loss: %-6.4f adv_loss: %-6.4f run_time:%-12s eta_time:%-12s'
% (it, max_iters, lr, recon_, adv_,
dnnlib.util.format_time(time.time() - start_time),
dnnlib.util.format_time(eta_time)))
if d_scale > 0:
print('d_r_loss: %-6.4f d_f_loss: %-6.4f d_reg: %-6.4f ' %
(d_r_, d_f_, d_g_))
sys.stdout.flush()
tflib.autosummary.save_summaries(summary_log, it)
if cur_nimg >= tick_start_nimg + 65000:
cur_tick += 1
tick_start_nimg = cur_nimg
if cur_tick % image_snapshot_ticks == 0:
batch_images_test = sess.run(image_batch_test)
batch_images_test = misc.adjust_dynamic_range(
batch_images_test.astype(np.float32), [0, 255], [-1., 1.])
recon = sess.run(fake_X_val,
feed_dict={real_test: batch_images_test})
orin_recon = np.concatenate([batch_images_test, recon], axis=0)
orin_recon = adjust_pixel_range(orin_recon)
orin_recon = fuse_images(orin_recon,
row=2,
col=submit_config.batch_size_test)
# save image results during training, first row is original images and the second row is reconstructed images
save_image(
'%s/iter_%09d.png' % (submit_config.run_dir, cur_nimg),
orin_recon)
if cur_tick % network_snapshot_ticks == 0:
pkl = os.path.join(submit_config.run_dir,
'network-snapshot-%09d.pkl' % (cur_nimg))
misc.save_pkl((E, G, D, Gs), pkl)
misc.save_pkl((E, G, D, Gs),
os.path.join(submit_config.run_dir, 'network-final.pkl'))
summary_log.close()