lambda_mse, lambda_adv, lambda_perc = args.lmse, args.ladv, args.lperc
itr_out_dir = args.expName + '-itrOut'
if os.path.isdir(itr_out_dir):
shutil.rmtree(itr_out_dir)
os.mkdir(itr_out_dir) # to save temp output
# redirect print to a file
sys.stdout = open('%s/%s' % (itr_out_dir, 'iter-prints.log'), 'w')
print('X train: {}\nY train: {}\nX test: {}\nY test: {}'.format(
args.xtrain, args.ytrain, args.xtest, args.ytest))
# build minibatch data generator with prefetch
mb_data_iter = bkgdGen(data_generator=gen_train_batch_bg(x_fn=args.xtrain, \
y_fn=args.ytrain, mb_size=mb_size, \
in_depth=in_depth, img_size=img_size), \
max_prefetch=16)
generator = make_generator_model(input_shape=(None, None, in_depth),
nlayers=args.lunet)
discriminator = make_discriminator_model(input_shape=(img_size, img_size, 1))
feature_extractor_vgg = tf.keras.applications.VGG19(\
weights='vgg19_weights_notop.h5', \
include_top=False)
# This method returns a helper function to compute cross entropy loss
cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)
def discriminator_loss(real_output, fake_output):
Methodname = 'CPCE_3D'
lambda_p = 0.1
in_depth = 1
Networkfolder = Methodname
is_transfer_learning = False
###################################################
xtrain = 'npy_frames_split_1024_train/input_*'
ytrain = 'npy_frames_split_1024_train/target_*'
img_size = 1024
# build minibatch data generator with prefetch
mb_data_iter = bkgdGen(data_generator=gen_train_batch_bg(x_fn=xtrain, \
y_fn=ytrain, mb_size=input_width, \
in_depth=in_depth, img_size=img_size), \
max_prefetch=16)
# Generator
X = tf.placeholder(dtype=tf.float32,
shape=[batch_size, in_depth, input_width, input_height, 1])
with tf.variable_scope('generator_model') as scope:
Y_ = CPCE_3D(X, padding='valid')
real_data = tf.placeholder(dtype=tf.float32,
shape=[batch_size, output_width, output_height, 1])
alpha = tf.random_uniform(shape=[batch_size, 1], minval=0., maxval=1.)
# Discriminator
with tf.variable_scope('discriminator_model') as scope: