def main(_):
if not os.path.exists(FLAGS.outdir):
os.mkdir(FLAGS.outdir)
if (os.path.exists(FLAGS.prunedlist_path)):
prunedlist = np.loadtxt(FLAGS.prunedlist_path, dtype=np.int64)
else:
prunedlist = [0] * 16
#network = EDSR(FLAGS.layers, FLAGS.featuresize, FLAGS.scale, FLAGS.channels)
network = EDSR(FLAGS.layers, FLAGS.featuresize, FLAGS.scale,
FLAGS.channels, FLAGS.channels, prunedlist)
network.buildModel()
network.resume(FLAGS.reusedir, 1999)
hr_list, lr_imgs, groundtruth_imgs = data_for_predict(
FLAGS.datadir, FLAGS.groundtruth, FLAGS.postfixlen)
if groundtruth_imgs:
psnr_list = []
time_list = []
fo = open(FLAGS.outdir + '/psnr.csv', 'w')
fo.writelines("file, PSNR\n")
for lr_img, groundtruth_img, hr_name in zip(lr_imgs, groundtruth_imgs,
hr_list):
start = time.time()
out = network.predict([lr_img])
# out = enhance_predict([lr_img],network)
use_time = time.time() - start
time_list.append(use_time)
tl.vis.save_image(out[0], FLAGS.outdir + '/' + hr_name)
psnr = utils.psnr_np(groundtruth_img, out[0], scale=8)
print('%s : %.6f' % (hr_name, psnr))
psnr_list.append(psnr)
fo.writelines("%s, %.6f\n" % (hr_name, psnr))
print(np.mean(psnr_list))
print(np.mean(time_list))
fo.writelines("%d, Average,0, %.6f" % (-1, np.mean(psnr_list)))
fo.close()
else:
for i in tqdm(range((len(hr_list)))):
out = network.predict([lr_imgs[i]])
tl.vis.save_image(out[0], FLAGS.outdir + '/' + hr_list[i])
def main(_):
if not os.path.exists(FLAGS.outdir):
os.mkdir(FLAGS.outdir)
network = EDSR(FLAGS.n_channels)
network.build_model(FLAGS.n_res_blocks, FLAGS.n_features, FLAGS.scale)
network.resume(FLAGS.reusedir, None)
hr_list, lr_imgs, groundtruth_imgs = data_for_predict(
FLAGS.datadir, FLAGS.grouthtruth, FLAGS.postfixlen)
# Valid
if groundtruth_imgs:
psnr_list = []
time_list = []
fo = open(FLAGS.outdir + '/psnr.csv', 'w')
fo.writelines("file, PSNR\n")
for lr_img, groundtruth_img, hr_name in zip(lr_imgs, groundtruth_imgs,
hr_list):
start = time.time()
out = network.predict([lr_img])
# out = enhance_predict([lr_img],network)
use_time = time.time() - start
time_list.append(use_time)
tl.vis.save_image(out[0], FLAGS.outdir + '/' + hr_name)
psnr = utils.psnr_np(groundtruth_img, out[0], scale=4)
print('%s : %.6f' % (hr_name, psnr))
psnr_list.append(psnr)
fo.writelines("%s, %.6f\n" % (hr_name, psnr))
print(np.mean(psnr_list))
print(np.mean(time_list))
fo.writelines("Average psnr,0, %.6f" % (np.mean(psnr_list)))
fo.writelines("Average runtime,0, %.6f" % (np.mean(time_list)))
fo.close()
# Test
else:
for i in tqdm(range((len(hr_list)))):
# out = network.predict([lr_imgs[i]])
out = enhance_predict([lr_imgs[i]], network)
tl.vis.save_image(out[0], FLAGS.outdir + '/' + hr_list[i])
def main(_):
img_files = sorted(os.listdir(FLAGS.datadir))
lr_imgs = tl.visualize.read_images(img_files, FLAGS.datadir)
mean_list = []
dict = {}
for i in range(len(img_files)):
name_lr, postfix = os.path.splitext(img_files[i])
name_hr = name_lr
hr_img = tl.visualize.read_image(name_hr + postfix, FLAGS.grouthtruth)
mean = utils.psnr_np(hr_img, lr_imgs[i])
print('%d -> %s : %.6f' % (i, name_lr + postfix, mean))
mean_list.append(mean)
dict[name_hr] = mean
dict_sort_by_value = OrderedDict(sorted(dict.items(), key=lambda x: x[1]))
mean = np.mean(mean_list)
with open(FLAGS.datadir + '/' + FLAGS.record, 'w') as file:
for k, v in dict_sort_by_value.items():
file.write('%s : %.6f\n' % (k, v))
file.write('Average: %.6f\n' % (mean))
print('Average: %.6f' % (mean))
def main(_):
if os.path.exists(FLAGS.outdir):
shutil.rmtree(FLAGS.outdir)
os.mkdir(FLAGS.outdir)
img_files = sorted(os.listdir(FLAGS.datadir))
lr_imgs, hr_imgs, lr_pos, hr_pos = utils.get_image_set(
img_files,
input_dir=FLAGS.datadir,
ground_truth_dir=FLAGS.groundtruth,
hr_image_size=0,
scale=FLAGS.scale,
postfix_len=FLAGS.postfixlen)
hr_norm_imgs = utils.normalize_color(hr_imgs)
network = WaveletSR(FLAGS.layers,
FLAGS.featuresize,
FLAGS.scale,
FLAGS.waveletimgsize,
FLAGS.hrimgsize,
channels=3)
network.buildModel()
network.resume(FLAGS.reusedir, global_step=FLAGS.step)
level = FLAGS.hrimgsize // (FLAGS.scale * FLAGS.waveletimgsize)
fo = open(FLAGS.outdir + '/psnr.csv', 'w')
fo.writelines("seq, file, L1, PSNR\n")
mean_list = []
for i in range(len(img_files)):
size, _, _ = np.shape(lr_imgs[i])
size_hr, _, _ = np.shape(hr_imgs[i])
target_imgs = utils.get_dwt_images(
[hr_norm_imgs[i]],
img_size=1 + (size_hr // (FLAGS.scale * math.pow(2, level - 1))),
wavelet=FLAGS.wavelet)
input_imgs = utils.get_dwt_images(
[lr_imgs[i]], img_size=1 + (size // level),
wavelet=FLAGS.wavelet) if level > 1 else [lr_imgs[i]]
#output = predict(i, input_imgs, network, target_imgs, fo)
output = ensem_predict(input_imgs, network)
output_img = make_same_shape(hr_imgs[i], output[0])
print(
'%dth composed image, loss = %.6f, min = %.6f, max = %.6f, mean = %.6f, var = %.6f\n'
%
(i,
np.mean(np.abs(hr_imgs[i] / 255.0, output_img.astype(
np.float32))), np.min(output_img), np.max(output_img),
np.mean(output_img), math.sqrt(np.var(output_img))))
output_img = np.clip(output_img, 0, 1)
output_img = output_img * 255 + 0.5
mean = utils.psnr_np(hr_imgs[i], output_img, scale=FLAGS.scale)
#fo.writelines("%s, %.6f\n"%(img_files[i], mean))
fo.writelines(
"%d, %s, %.6f, %.6f\n" %
(i, img_files[i], np.mean(np.abs(hr_imgs[i] - output_img)), mean))
mean_list.append(mean)
tl.vis.save_image(output_img, FLAGS.outdir + '/' + img_files[i])
fo.writelines("%d, Average,0, %.6f\n" % (-1, np.mean(mean_list)))
fo.close()
return