tl.files.assign_params(sess, load_params, net)
print_("\tdone\n")
if not os.path.exists(FLAGS.out_dir):
os.makedirs(FLAGS.out_dir)
print_("\nStarting prediction...\n\n")
k = 0
for i in range(len(frames)):
print("Frame %d: '%s'" % (i, frames[i]))
try:
# Read frame
print_("\tReading...")
# 读取LDR图像，resize到对应的尺度，并进行亮度缩放与截取，返回一个4维张量
x_buffer = img_io.readLDR(frames[i], (sy, sx), True, FLAGS.scaling)
print_("\tdone")
print_("\t(Saturation: %0.2f%%)\n" % (100.0 * (x_buffer >= 1).sum() / x_buffer.size), 'm')
# Run prediction.
# The gamma value is used to allow for boosting/reducing the intensity of
# the reconstructed highlights. If y = f(x) is the reconstruction, the gamma
# g alters this according to y = f(x^(1/g))^g
print_("\tInference...")
feed_dict = {x: np.power(np.maximum(x_buffer, 0.0), 1.0 / FLAGS.gamma)}
y_predict = sess.run([y], feed_dict=feed_dict)
y_predict = np.power(np.maximum(y_predict, 0.0), FLAGS.gamma)
print_("\tdone\n")
# Gamma corrected output 用伽马校正代替色调映射
img_path_list.sort()
img_list_H = list()
img_list = list()
if is_upexposure_trained:
img_order = range(len(img_path_list))
else:
img_order = range(len(img_path_list) - 1, -1, -1)
img_H_ = img_io.readHDR(img_path_list_H[0], (sy, sx))
img_list_H.append(np.squeeze(img_H_))
img_list_H = np.array(img_list_H)
for j in img_order:
img_path = img_path_list[j]
img = img_io.readLDR(img_path, (sy, sx), True, FLAGS.scaling)
img_list.append(np.squeeze(img))
img_list = np.array(img_list)
for input_frame_id in range(len(img_list) - 1):
start_frame_id = input_frame_id + 2
end_frame_id = min(start_frame_id + predicted_window_len,
len(img_list))
x_batch = np.array([img_list[input_frame_id, :, :, :]])
y_batch_0 = img_list_H.reshape((1, ) + x_batch.shape[:]).astype(
np.float32)
y_batch_1 = np.array(
[img_list[start_frame_id:end_frame_id, :, :, :]])
y_batch = np.concatenate([y_batch_0, y_batch_1], axis=1)
dummy_len = predicted_window_len - y_batch.shape[1]
zero_dummy = np.zeros(
start_time = time.clock()
dir_path = dir_path_list[i]
# frams = [glob.glob(dir_path + '/LDR/1.png')[0], glob.glob(dir_path + '/LDR/4.png')[0], glob.glob(dir_path + '/LDR/7.png')[0]] # used for DML, Fairchild, NewHDR datasets
frams = [
glob.glob(dir_path + '/LDR/002.png')[0],
glob.glob(dir_path + '/LDR/005.png')[0],
glob.glob(dir_path + '/LDR/007.png')[0]
] # just only CanonCamera dataset
filename_root = os.path.basename(dir_path)
print('filename', filename_root)
save_path = dir_outpath[0] + '/' + filename_root
if not os.path.exists(save_path):
os.makedirs(save_path)
try:
x_input_1 = img_io.readLDR(frams[0], (sy, sx), True, FLAGS.scaling)
x_input_4 = img_io.readLDR(frams[1], (sy, sx), True, FLAGS.scaling)
x_input_7 = img_io.readLDR(frams[2], (sy, sx), True, FLAGS.scaling)
y_predict_dm_1 = sess_dm.run(pred_placehoder_dm,
feed_dict={
x_dm: x_input_1 * 1,
x_local_dm: x_input_1 * 1
})
y_predict_dm_4 = sess_dm.run(pred_placehoder_dm,
feed_dict={
x_dm: x_input_4 * 1,
x_local_dm: x_input_4 * 1
})
y_predict_dm_7 = sess_dm.run(pred_placehoder_dm,
feed_dict={
frames = [FLAGS.im_dir]
if os.path.isdir(FLAGS.im_dir):
frames = [
os.path.join(FLAGS.im_dir, name)
for name in sorted(os.listdir(FLAGS.im_dir))
if os.path.isfile(os.path.join(FLAGS.im_dir, name))
]
x = tf.placeholder(tf.float32, shape=[1, sy, sx, 3])
net = ldr2hdr.model(x)
y = ldr2hdr.get_final(net, x)
sess = tf.InteractiveSession()
load_params = tl.files.load_npz(name=FLAGS.params)
tl.files.assign_params(sess, load_params, net)
if not os.path.exists(FLAGS.out_dir):
os.makedirs(FLAGS.out_dir)
k = 0
for i in range(len(frames)):
x_buffer = img_io.readLDR(frames[i], (sy, sx), True, 1.0)
y_predict = sess.run([y], feed_dict={x: x_buffer})
y_gamma = np.power(np.maximum(y_predict, 0.0), 0.5)
k += 1
outname = './output' + frames[i][7:-4] + '.exr'
img_io.writeEXR(y_predict, outname)
sess.close()