def _evaluate_denoise(sr_model : BaseSuperResolutionModel, validation_dir, scale_pred=False):
print("Validating %s model" % sr_model.model_name)
predict_path = "val_predict/"
if not os.path.exists(predict_path):
os.makedirs(predict_path)
validation_path_set5 = validation_dir + "set5/"
validation_path_set14 = validation_dir + "set14/"
validation_dirs = [validation_path_set5, validation_path_set14]
for val_dir in validation_dirs:
image_fns = [name for name in os.listdir(val_dir)]
nb_images = len(image_fns)
print("Validating %d images from path %s" % (nb_images, val_dir))
total_psnr = 0.0
for impath in os.listdir(val_dir):
t1 = time.time()
# Input image
y = img_utils.imread(val_dir + impath, mode='RGB')
width, height, _ = y.shape
if ((width // sr_model.scale_factor) % 4 != 0) or ((height // sr_model.scale_factor) % 4 != 0) \
or (width % 2 != 0) or (height % 2 != 0):
width = ((width // sr_model.scale_factor) // 4) * 4 * sr_model.scale_factor
height = ((height // sr_model.scale_factor) // 4) * 4 * sr_model.scale_factor
print("Model %s require the image size to be divisible by 4. New image size = (%d, %d)" % \
(sr_model.model_name, width, height))
y = img_utils.imresize(y, (width, height), interp='bicubic')
y = y.astype('float32')
y = np.expand_dims(y, axis=0)
x_temp = y.copy()
if sr_model.type_scale_type == "tanh":
x_temp = (x_temp - 127.5) / 127.5
y = (y - 127.5) / 127.5
else:
x_temp /= 255.
y /= 255.
img = img_utils.imresize(x_temp[0], (width // sr_model.scale_factor, height // sr_model.scale_factor),
interp='bicubic', mode='RGB')
if not sr_model.type_true_upscaling:
img = img_utils.imresize(img, (width, height), interp='bicubic')
x = np.expand_dims(img, axis=0)
if K.image_dim_ordering() == "th":
x = x.transpose((0, 3, 1, 2))
y = y.transpose((0, 3, 1, 2))
sr_model.model = sr_model.create_model(height, width, load_weights=True)
if sr_model.evaluation_func is None:
if sr_model.uses_learning_phase:
sr_model.evaluation_func = K.function([sr_model.model.layers[0].input, K.learning_phase()],
[sr_model.model.layers[-1].output])
else:
sr_model.evaluation_func = K.function([sr_model.model.layers[0].input],
[sr_model.model.layers[-1].output])
if sr_model.uses_learning_phase:
y_pred = sr_model.evaluation_func([x, 0])[0][0]
else:
y_pred = sr_model.evaluation_func([x])[0][0]
if scale_pred:
if sr_model.type_scale_type == "tanh":
y_pred = (y_pred + 1) * 127.5
else:
y_pred *= 255.
if sr_model.type_scale_type == 'tanh':
y = (y + 1) / 2
psnr_val = psnr(y[0], np.clip(y_pred, 0, 255) / 255)
total_psnr += psnr_val
t2 = time.time()
print("Validated image : %s, Time required : %0.2f, PSNR value : %0.4f" % (impath, t2 - t1, psnr_val))
generated_path = predict_path + "%s_%s_generated.png" % (sr_model.model_name, os.path.splitext(impath)[0])
if K.image_dim_ordering() == "th":
y_pred = y_pred.transpose((1, 2, 0))
y_pred = np.clip(y_pred, 0, 255).astype('uint8')
img_utils.imsave(generated_path, y_pred)
print("Average PRNS value of validation images = %00.4f \n" % (total_psnr / nb_images))
def evaluate(self, validation_dir, scale_factor, target_size=256, small_train_images=False):
"""
Evaluates the model on the Set5 Validation images
"""
if self.model == None: self.create_model(load_weights=True, small_train_images=small_train_images)
if self.evaluation_func is None:
self.evaluation_func = K.function([self.model.layers[0].input],
[self.model.layers[-1].output])
predict_path = "val_predict/"
if not os.path.exists(predict_path):
os.makedirs(predict_path)
image_fns = [name for name in os.listdir(validation_dir)]
nb_images = len(image_fns)
print("Validating %d images" % (nb_images))
total_psnr = 0.0
for impath in os.listdir(validation_dir):
t1 = time.time()
# Input image
y = img_utils.imread(validation_dir + impath, mode='RGB')
width, height, _ = y.shape
if self.model_name in self.denoise_models:
# Denoise models require precise width and height, divisible by 4
if ((width // scale_factor) % 4 != 0) or ((height // scale_factor) % 4 != 0):
width = ((width // scale_factor) // 4) * 4 * scale_factor
height = ((height // scale_factor) // 4) * 4 * scale_factor
print("Model %s require the image size to be divisible by 4. New image size = (%d, %d)" % \
(self.model_name, width, height))
y = img_utils.imresize(y, (width, height), interp='bicubic')
y = y.astype('float32') / 255.
y = np.expand_dims(y, axis=0)
x_width = width if not small_train_images else width // scale_factor
x_height = height if not small_train_images else height // scale_factor
x_temp = y.copy()
img = img_utils.gaussian_filter(x_temp[0], sigma=0.01)
img = img_utils.imresize(img, (x_width // scale_factor, x_height // scale_factor), interp='bicubic')
if not small_train_images:
img = img_utils.imresize(img, (x_width, x_height), interp='bicubic')
x = np.expand_dims(img, axis=0)
if K.image_dim_ordering() == "th":
x = x.transpose((0, 3, 1, 2))
y = y.transpose((0, 3, 1, 2))
y_pred = self.evaluation_func([x])[0][0]
psnr_val = psnr(y[0], np.clip(y_pred, 0, 255) / 255)
total_psnr += psnr_val
t2 = time.time()
print("Validated image : %s, Time required : %0.2f, PSNR value : %0.4f" % (impath, t2 - t1, psnr_val))
generated_path = predict_path + "%s_generated.png" % (os.path.splitext(impath)[0])
if K.image_dim_ordering() == "th":
y_pred = y_pred.transpose((1, 2, 0))
y_pred = np.clip(y_pred, 0, 255).astype('uint8')
img_utils.imsave(generated_path, y_pred)
print("Average PRNS value of validation images = %00.4f" % (total_psnr / nb_images))
def evaluate(self,
validation_dir,
scale_factor,
target_size=256,
small_train_images=False):
"""
Evaluates the model on the Set5 Validation images
"""
if self.model == None:
self.create_model(load_weights=True,
small_train_images=small_train_images)
if self.evaluation_func is None:
self.evaluation_func = K.function([self.model.layers[0].input],
[self.model.layers[-1].output])
predict_path = "val_predict/"
if not os.path.exists(predict_path):
os.makedirs(predict_path)
image_fns = [name for name in os.listdir(validation_dir)]
nb_images = len(image_fns)
print("Validating %d images" % (nb_images))
total_psnr = 0.0
for impath in os.listdir(validation_dir):
t1 = time.time()
# Input image
y = img_utils.imread(validation_dir + impath, mode='RGB')
width, height, _ = y.shape
if self.model_name in self.denoise_models:
# Denoise models require precise width and height, divisible by 4
if ((width // scale_factor) % 4 != 0) or (
(height // scale_factor) % 4 != 0):
width = ((width // scale_factor) // 4) * 4 * scale_factor
height = ((height // scale_factor) // 4) * 4 * scale_factor
print("Model %s require the image size to be divisible by 4. New image size = (%d, %d)" % \
(self.model_name, width, height))
y = img_utils.imresize(y, (width, height),
interp='bicubic')
y = y.astype('float32') / 255.
y = np.expand_dims(y, axis=0)
x_width = width if not small_train_images else width // scale_factor
x_height = height if not small_train_images else height // scale_factor
x_temp = y.copy()
img = img_utils.gaussian_filter(x_temp[0], sigma=0.01)
img = img_utils.imresize(
img, (x_width // scale_factor, x_height // scale_factor),
interp='bicubic')
if not small_train_images:
img = img_utils.imresize(img, (x_width, x_height),
interp='bicubic')
x = np.expand_dims(img, axis=0)
if K.image_dim_ordering() == "th":
x = x.transpose((0, 3, 1, 2))
y = y.transpose((0, 3, 1, 2))
y_pred = self.evaluation_func([x])[0][0]
psnr_val = psnr(y[0], np.clip(y_pred, 0, 255) / 255)
total_psnr += psnr_val
t2 = time.time()
print(
"Validated image : %s, Time required : %0.2f, PSNR value : %0.4f"
% (impath, t2 - t1, psnr_val))
generated_path = predict_path + "%s_generated.png" % (
os.path.splitext(impath)[0])
if K.image_dim_ordering() == "th":
y_pred = y_pred.transpose((1, 2, 0))
y_pred = np.clip(y_pred, 0, 255).astype('uint8')
img_utils.imsave(generated_path, y_pred)
print("Average PRNS value of validation images = %00.4f" %
(total_psnr / nb_images))
def evaluate(self, validation_dir, small_train_images=False):
print("Validating %s model" % self.model_name)
predict_path = "val_predict/"
if not os.path.exists(predict_path):
os.makedirs(predict_path)
validation_path_set5 = validation_dir + "set5/"
validation_path_set14 = validation_dir + "set14/"
validation_dirs = [validation_path_set5, validation_path_set14]
for val_dir in validation_dirs:
image_fns = [name for name in os.listdir(val_dir)]
nb_images = len(image_fns)
print("Validating %d images from path %s" % (nb_images, val_dir))
total_psnr = 0.0
for impath in os.listdir(val_dir):
t1 = time.time()
# Input image
y = img_utils.imread(val_dir + impath, mode='RGB')
width, height, _ = y.shape
y = y.astype('float32') / 255.
y = np.expand_dims(y, axis=0)
x_temp = y.copy()
img = img_utils.imresize(
x_temp[0],
(width // self.scale_factor, height // self.scale_factor),
interp='bicubic')
if not small_train_images:
img = img_utils.imresize(img, (width, height),
interp='bicubic')
x = np.expand_dims(img, axis=0)
if K.image_dim_ordering() == "th":
x = x.transpose((0, 3, 1, 2))
y = y.transpose((0, 3, 1, 2))
self.model = self.create_model(height,
width,
load_weights=True)
self.evaluation_func = K.function(
[self.model.layers[0].input],
[self.model.layers[-1].output])
y_pred = self.evaluation_func([x])[0][0]
psnr_val = psnr(y[0], np.clip(y_pred, 0, 255) / 255)
total_psnr += psnr_val
t2 = time.time()
print(
"Validated image : %s, Time required : %0.2f, PSNR value : %0.4f"
% (impath, t2 - t1, psnr_val))
generated_path = predict_path + "%s_%s_generated.png" % (
self.model_name, os.path.splitext(impath)[0])
if K.image_dim_ordering() == "th":
y_pred = y_pred.transpose((1, 2, 0))
y_pred = np.clip(y_pred, 0, 255).astype('uint8')
img_utils.imsave(generated_path, y_pred)
print("Average PRNS value of validation images = %00.4f \n" %
(total_psnr / nb_images))
