args = train_utils.parse_params(config_file_path)
L = args.num_classes
args.image_dim = [128, 128, 3]
#%% load saved network parameters and open new session
tf.reset_default_graph()
in_placeholder = tf.placeholder(
tf.float32,
shape=[None, None, None, L + args.image_dim[2]],
name="in_placeholder")
out_placeholder = tf.placeholder(tf.float32,
shape=[None, None, None, L],
name='out_placeholder')
phase = tf.placeholder(tf.bool, name='phase')
net_class = Models(args)
net_class.build_model(in_placeholder, phase)
sess = tf.Session()
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(checkpoints_path)
if ckpt and ckpt.model_checkpoint_path:
ckpt_path = checkpoints_path + 'my_model-' + str(checkpoint[0])
saver.restore(sess, ckpt_path)
#%%
if not os.path.exists(output_path):
os.makedirs(output_path)
image_list = sorted(os.listdir(pascal_path + '/images'))
elif args.app == 'dehaze':
(x_test, gt_test) = (f_HDF5["test"], f_HDF5["test_gt"])
app_loss = Dehazing_Loss(args)
elif args.app == 'matte':
(x_test, x_test_1, x_test_2,
gt_test) = (f_HDF5["val"], f_HDF5["test_t2"], f_HDF5["test_t1"],
f_HDF5["test_gt"])
app_loss = Matting_Loss(args)
print("Constructing network...")
net_class = Models(args)
tf.reset_default_graph()
input_placeholder = tf.placeholder(tf.float32, (None, ) + x_train.shape[1:],
'images_with_seeds')
phase = tf.placeholder(tf.bool, name='phase')
output_from_network = net_class.build_model(input_placeholder, phase)
loss = app_loss._loss(input_placeholder, output_from_network)
samples_in_epoch = int(
np.ceil(np.shape(x_train)[0] / np.float32(args.batch_size)))
num_iter = args.num_epochs * samples_in_epoch + 1
global_step = tf.Variable(0, trainable=False)
if args.train_schedule == 'exp': # decrease learning exponentially
learning_rate = tf.train.exponential_decay(args.init_lr, global_step,samples_in_epoch*args.decay_every,\
args.decay_factor, staircase=True)
elif args.train_schedule == 'plateau': # decrease learning rate when plateau is reached
lr_placeholder = tf.placeholder(tf.float32, [], name='learning_rate')
learning_rate = lr_placeholder
