from sys import argv, stderr
from LRNN import gen_model
def die(msg):
print(msg, file=stderr)
exit(1)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
np.random.seed(0x5EED)
if len(argv) < 4:
die('Usage: {} [model path] [output path] [Gaussian noise variance]'.format(argv[0]))
model = gen_model((None, None, 3), 'parallel')
model.load_weights(argv[1])
datadir = './data/test-denoise'
filenames = os.listdir(datadir)
outputdir = argv[2]
if not os.path.exists(outputdir):
os.makedirs(outputdir)
gauss_var = float(argv[3])
for filename in filenames:
pathname = os.path.join(datadir, filename)
img = skimage.io.imread(pathname)
r, c = img.shape[:2]
if patch.ndim == 2:
patch = skimage.color.gray2rgb(patch)
images[n_sample] = patch
filter_images[n_sample] = denoise_bilateral(
patch,
sigma_color=0.1,
sigma_spatial=2,
multichannel=True)
n_sample = n_sample + 1
yield (images[:n_sample], filter_images[:n_sample])
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
np.random.seed(0x5EED)
model = gen_model((96, 96, 3), 'cascade')
model.compile(optimizer='Adamax', loss=Keras.losses.mse)
# Keras.utils.plot_model(model, to_file='model.png')
batch_size = 32
patchsize = 96
n_epoch = 30
datadir = './data/train2014'
pathnames = [os.path.join(datadir, f) for f in os.listdir(datadir)]
n_sample = len(pathnames)
steps_per_epoch = int((n_sample + batch_size - 1) / batch_size)
steps_per_epoch = 500
n_epoch = 300000
valdir = './data/test2014'
from LRNN import gen_model
def die(msg):
print(msg, file=stderr)
exit(1)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
np.random.seed(0x5EED)
if len(argv) < 3:
die('Usage: {} [model path] [output path]'.format(argv[0]))
model = gen_model((None, None, 3), 'cascade')
model.load_weights(argv[1])
datadir = './data/test-denoise'
filenames = os.listdir(datadir)
outputdir = argv[2]
if not os.path.exists(outputdir):
os.makedirs(outputdir)
for filename in filenames:
pathname = os.path.join(datadir, filename)
img = skimage.io.imread(pathname)
r, c = img.shape[:2]
if (r % 16 != 0) or (c % 16 != 0):
print('Image dimension is not 16-divisible: ' + filename)
for _ in range(4):
patch = gen_patch(img, patchsize)
patch = skimage.util.img_as_float(patch)
if patch.ndim == 2:
patch = skimage.color.gray2rgb(patch)
images[n_sample] = patch
gauss_var = (0.01 + 0.09 * np.random.random())**2
noisy_images[n_sample] = random_noise(patch, var=gauss_var)
n_sample = n_sample + 1
yield (noisy_images[:n_sample], images[:n_sample])
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
np.random.seed(0x5EED)
model = gen_model((96, 96, 3), 'parallel')
model.compile(optimizer='Adamax', loss=Keras.losses.mse)
# Keras.utils.plot_model(model, to_file='model.png')
batch_size = 32
patchsize = 96
n_epoch = 30
datadir = './data/train2014'
pathnames = [os.path.join(datadir, f) for f in os.listdir(datadir)]
n_sample = len(pathnames)
steps_per_epoch = int((n_sample + batch_size - 1) / batch_size)
steps_per_epoch = 100
n_epoch = 300000
valdir = './data/test2014'
if len(argv) < 2:
raise ValueError("Usage: {} [cascade | parallel] [start iter]".format(argv[0]))
if len(argv) > 2:
start_from_iter = int(argv[2])
if start_from_iter <= 0:
raise ValueError("start_from_iter should be larger than 0")
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
np.random.seed(0x5EED)
batch_size = 32
patchsize = 96
n_epoch = 30
model = gen_model((patchsize, patchsize, 4), argv[1])
model.compile(optimizer='Adamax', loss=Keras.losses.mse)
# Keras.utils.plot_model(model, to_file='model.png')
datadir = './data/train2014'
pathnames = [os.path.join(datadir, f) for f in os.listdir(datadir)]
n_sample = len(pathnames)
steps_per_epoch=200
n_epoch=300000
valdir = './data/test2014'
valnames = [os.path.join(valdir, f) for f in os.listdir(valdir)]
outputdir = './model-color'
if not os.path.exists(outputdir):
os.makedirs(outputdir)
def degrade(img, p):
r, c = img.shape[:2]
mask = np.random.choice([True, False], size=(r, c), p=[1-p, p])
out = img.copy()
out[mask] = 0
return out
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
np.random.seed(0x5EED)
if len(argv) < 4:
die('Usage: {} [model path] [output path] [cascade | parallel]'.format(argv[0]))
model = gen_model((None, None, 4), argv[3])
model.load_weights(argv[1])
datadir = './data/test-denoise'
filenames = os.listdir(datadir)
outputdir = argv[2]
if not os.path.exists(outputdir):
os.makedirs(outputdir)
for filename in filenames:
pathname = os.path.join(datadir, filename)
img = skimage.io.imread(pathname)
r, c = img.shape[:2]
if (r % 16 != 0) or (c % 16 != 0):
print('Image dimension is not 16-divisible: ' + filename)