g_vars, d_vars, losses = model.build_graph_with_losses(images,
config=config)
# validation images
if config.VAL:
with open(config.DATA_FLIST[config.DATASET][1]) as f:
val_fnames = f.read().splitlines()
# progress monitor by visualizing static images
for i in range(config.STATIC_VIEW_SIZE):
static_fnames = val_fnames[i:i + 1]
static_images = ng.data.DataFromFNames(
static_fnames,
config.IMG_SHAPES,
nthreads=1,
random_crop=config.RANDOM_CROP).data_pipeline(1)
static_inpainted_images = model.build_static_infer_graph(
static_images, config, name='static_view/%d' % i)
# training settings
lr = tf.get_variable('lr',
shape=[],
trainable=False,
initializer=tf.constant_initializer(1e-4))
d_optimizer = tf.train.AdamOptimizer(lr, beta1=0.5, beta2=0.9)
g_optimizer = d_optimizer
# gradient processor
if config.GRADIENT_CLIP:
gradient_processor = lambda grad_var: (tf.clip_by_average_norm(
grad_var[0], config.GRADIENT_CLIP_VALUE), grad_var[1])
else:
gradient_processor = None
# validation images
if FLAGS.val:
with open(FLAGS.data_flist[FLAGS.dataset][1]) as f:
val_fnames = f.read().splitlines()
if FLAGS.guided:
val_fnames = [(fname, fname[:-4] + '_edge.jpg')
for fname in val_fnames]
# progress monitor by visualizing static images
for i in range(FLAGS.static_view_size):
static_fnames = val_fnames[i:i + 1]
static_images = ng.data.DataFromFNames(
static_fnames,
img_shapes,
nthreads=1,
random_crop=FLAGS.random_crop).data_pipeline(1)
static_inpainted_images = model.build_static_infer_graph(
FLAGS, static_images, name='static_view/%d' % i)
# training settings
lr = tf.get_variable('lr',
shape=[],
trainable=False,
initializer=tf.constant_initializer(1e-4))
d_optimizer = tf.train.AdamOptimizer(lr, beta1=0.5, beta2=0.999)
g_optimizer = d_optimizer
# train discriminator with secondary trainer, should initialize before
# primary trainer.
# discriminator_training_callback = ng.callbacks.SecondaryTrainer(
discriminator_training_callback = ng.callbacks.SecondaryMultiGPUTrainer(
num_gpus=FLAGS.num_gpus_per_job,
pstep=1,
optimizer=d_optimizer,
var_list=d_vars,
images = data.data_pipeline(config.BATCH_SIZE)
# main model
model = InpaintCAModel()
g_vars, d_vars, losses = model.build_graph_with_losses(
images[0], config=config, exclusionmask=images[exclusionmask_index] if config.EXC_MASKS else None, mask=None if config.GEN_MASKS else images[mask_index])
# validation images
if config.VAL:
with open(config.DATA_FLIST[config.DATASET][1]) as f:
val_fnames = [(l.split(' ')[0], l.split(' ')[1]) for l in f.read().splitlines()]
# progress monitor by visualizing static images
for i in range(config.STATIC_VIEW_SIZE):
static_fnames = val_fnames[i:i+1]
static_images = ng.data.DataFromFNames(
static_fnames, config.IMG_SHAPES, nthreads=1,
random_crop=config.RANDOM_CROP, random_flip=config.RANDOM_FLIP).data_pipeline(1)
static_inpainted_images = model.build_static_infer_graph(
static_images[0], config, name='static_view/%d' % i, exclusionmask=images[exclusionmask_index] if config.EXC_MASKS else None)
# training settings
lr = tf.get_variable(
'lr', shape=[], trainable=False,
initializer=tf.constant_initializer(1e-4))
d_optimizer = tf.train.AdamOptimizer(lr, beta1=0.5, beta2=0.9)
g_optimizer = d_optimizer
# gradient processor
if config.GRADIENT_CLIP:
gradient_processor = lambda grad_var: (
tf.clip_by_average_norm(grad_var[0], config.GRADIENT_CLIP_VALUE),
grad_var[1])
else:
gradient_processor = None
# log dir
log_prefix = 'model_logs/' + '_'.join([
