def run_network(downscaled_image, checkpoint):
# network for original image
config = tf.ConfigProto(
device_count={'GPU': 1}
)
graph_or = tf.Graph()
sess_or = tf.Session(graph=graph_or, config=config)
with graph_or.as_default():
input_or = tf.placeholder(tf.float32, (1, downscaled_image.shape[1], downscaled_image.shape[2], params.num_channels), name='input_or')
_, output_or = params.network_architecture(input_or)
saver = tf.train.Saver()
saver.restore(sess_or, checkpoint)
graph_tr = tf.Graph()
sess_tr = tf.Session(graph=graph_tr, config=config)
with graph_tr.as_default():
input_tr = tf.placeholder(tf.float32, (1, downscaled_image.shape[2], downscaled_image.shape[1], params.num_channels), name='input_tr')
_, output_tr = params.network_architecture(input_tr, reuse=False)
saver = tf.train.Saver()
saver.restore(sess_tr, checkpoint)
num_images = downscaled_image.shape[0]
cnn_output = []
for image in downscaled_image:
out_images = []
# original 0
res = trim_image(sess_or.run(output_or, {input_or: [image]})[0])
out_images.append(res)
# flip 0
res = trim_image(sess_or.run(output_or, {input_or: [flip_image(image)]})[0])
out_images.append(reverse_flip_image(res))
# original 180
rot180_image = rotate_image_180(image)
res = trim_image(sess_or.run(output_or, {input_or: [rot180_image]})[0])
out_images.append(reverse_rotate_image_180(res))
# flip 180
res = trim_image(sess_or.run(output_or, {input_or: [flip_image(rot180_image)]})[0])
out_images.append(reverse_rotate_image_180(reverse_flip_image(res)))
if use_mean:
cnn_output.append(np.round(np.mean(np.array(out_images), axis=0)))
else:
cnn_output.append(np.round(np.median(np.array(out_images), axis=0)))
cnn_output = np.array(cnn_output)
return cnn_output
def upscale(downscaled_image, checkpoint):
scale_factor = params.scale
# cnn resize
input = tf.placeholder(tf.float32,
(1, downscaled_image.shape[1],
downscaled_image.shape[2], params.num_channels),
name='input')
_, output = params.network_architecture(input)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
print('restoring from ' + checkpoint)
saver.restore(sess, checkpoint)
# step 1 - apply cnn on each resized image, maybe as a batch
cnn_output = []
for image in downscaled_image:
cnn_output.append(sess.run(output, feed_dict={input: [image]})[0])
cnn_output = np.array(cnn_output)
cnn_output = np.round(cnn_output)
cnn_output[cnn_output > 255] = 255
return cnn_output
'./data/validation',
'./data/test',
SHOW_IMAGES=False)
# training
batch_size = 128
input = tf.placeholder(tf.float32,
(batch_size, data_reader.dim_patch_in_rows,
data_reader.dim_patch_in_cols, params.num_channels),
name='input')
target = tf.placeholder(tf.float32,
(batch_size, data_reader.dim_patch_gt_rows,
data_reader.dim_patch_gt_cols, params.num_channels),
name='target')
output_PS, output = params.network_architecture(input)
print('output shape is ', output.shape, target.shape)
if params.LOSS == params.L1_LOSS:
loss = tf.reduce_mean(
tf.reduce_mean(tf.abs(output - target)) +
tf.reduce_mean(tf.abs(output_PS - target)))
if params.LOSS == params.L2_LOSS:
loss = tf.reduce_mean(
tf.reduce_mean(tf.square(output - target)) +
tf.reduce_mean(tf.square(output_PS - target)))
# alpha = 0.7
# loss = alpha * tf.reduce_mean(tf.reduce_mean(tf.abs(output - target)) + tf.reduce_mean(tf.abs(output_PS - target))) + (1 - alpha) * tf.reduce_mean(tf.reduce_mean(tf.square(output - target)) + tf.reduce_mean(tf.square(output_PS - target)))
global_step = tf.Variable(0, trainable=False)
def upscale(downscaled_image, checkpoint):
# network for original image
config = tf.ConfigProto(
device_count={'GPU': 1}
)
sess_or = tf.Session(config=config)
with sess_or:
input_or = tf.placeholder(tf.float32, (1, downscaled_image.shape[1], downscaled_image.shape[2], params.num_channels), name='input_or')
_, output_or = params.network_architecture(input_or)
sess_tr = tf.Session(config=config)
with sess_tr:
input_tr = tf.placeholder(tf.float32, (1, downscaled_image.shape[2], downscaled_image.shape[1], params.num_channels), name='input_tr')
_, output_tr = params.network_architecture(input_tr, reuse=True)
saver = tf.train.Saver()
print('restoring from ' + checkpoint)
saver.restore(sess_or, checkpoint)
saver.restore(sess_tr, checkpoint)
num_images = downscaled_image.shape[0]
cnn_output = []
for image in downscaled_image:
out_images = []
# original 0
res = trim_image(sess_or.run(output_or, {input_or: [image]})[0])
out_images.append(res)
# flip 0
res = trim_image(sess_or.run(output_or, {input_or: [flip_image(image)]})[0])
out_images.append(reverse_flip_image(res))
# original 90
rot90_image = rotate_image_90(image)
res = trim_image(sess_tr.run(output_tr, {input_tr: [rot90_image]})[0])
out_images.append(reverse_rotate_image_90(res))
# flip 90
res = trim_image(sess_tr.run(output_tr, {input_tr: [flip_image(rot90_image)]})[0])
out_images.append(reverse_rotate_image_90(reverse_flip_image(res)))
# original 180
rot180_image = rotate_image_180(image)
res = trim_image(sess_or.run(output_or, {input_or: [rot180_image]})[0])
out_images.append(reverse_rotate_image_180(res))
# flip 180
res = trim_image(sess_or.run(output_or, {input_or: [flip_image(rot180_image)]})[0])
out_images.append(reverse_rotate_image_180(reverse_flip_image(res)))
# original 270
rot270_image = rotate_image_270(image)
res = trim_image(sess_tr.run(output_tr, {input_tr: [rot270_image]})[0])
out_images.append(reverse_rotate_image_270(res))
# flip 270
res = trim_image(sess_tr.run(output_tr, {input_tr: [flip_image(rot270_image)]})[0])
out_images.append(reverse_rotate_image_270(reverse_flip_image(res)))
# pdb.set_trace()
if use_mean:
cnn_output.append(np.round(np.mean(np.array(out_images), axis=0)))
else:
cnn_output.append(np.round(np.median(np.array(out_images), axis=0)))
cnn_output = np.array(cnn_output)
return cnn_output