def __getitem__(self, index):
LR_path = None
# get LR image
LR_path = self.paths_LR[index]
#img_LR = util.read_img(self.LR_env, LR_path)
img_LR = util.read_img(self.LR_env, LR_path)
H, W, C = img_LR.shape
# change color space if necessary
if self.opt['color']:
img_LR = util.channel_convert(C, self.opt['color'], [img_LR])[0]
# BGR to RGB, HWC to CHW, numpy to tensor
img_LR = util.np2tensor(img_LR, normalize=self.znorm, add_batch=False)
return {'LR': img_LR, 'LR_path': LR_path}
def __getitem__(self, index):
"""Return a data point and its metadata information.
Parameters:
index (int): a random integer for data indexing
Returns a dictionary that contains A and A_paths or
LR and LR_path
A (tensor): an image in the input domain
A_paths (str): the path of the image
"""
# get single image
img_A, A_path = read_single_dataset(self.opt, index, self.A_paths,
self.A_env)
# change color space if necessary
# TODO: move to get_transform()
if self.opt.get('color', None):
img_A = channel_convert(image_channels(img_A), self.opt['color'],
[img_A])[0]
# # apply transforms if any
# default_int_method = get_default_imethod(image_type(img_A))
# transform_params = get_params(self.opt, image_size(img_A))
# A_transform = get_transform(
# self.opt,
# transform_params,
# # grayscale=(input_nc == 1),
# method=default_int_method)
# img_A = A_transform(img_A)
######## Convert images to PyTorch Tensors ########
img_A = self.tensor_transform(img_A)
if self.vars == 'A':
return {'A': img_A, 'A_path': A_path}
else:
return {'LR': img_A, 'LR_path': A_path}
def __getitem__(self, index):
HR_path, LR_path = None, None
scale = self.opt.get('scale', 4)
HR_size = self.opt.get('HR_size', 128)
if HR_size:
LR_size = HR_size // scale
# Default case: tensor will result in the [0,1] range
# Alternative: tensor will be z-normalized to the [-1,1] range
znorm = self.opt.get('znorm', False)
######## Read the images ########
#TODO: check cases where default of 3 channels will be troublesome
image_channels = self.opt.get('image_channels', 3)
# Check if LR Path is provided
if self.paths_LR:
#If LR is provided, check if 'rand_flip_LR_HR' is enabled
if self.opt.get('rand_flip_LR_HR',
None) and self.opt['phase'] == 'train':
LRHRchance = random.uniform(0, 1)
flip_chance = self.opt.get('flip_chance', 0.05)
#print("Random Flip Enabled")
# Normal case, no flipping:
else:
LRHRchance = 0.
flip_chance = 0.
#print("No Random Flip")
# get HR and LR images
# If enabled, random chance that LR and HR images are flipped
# Normal case, no flipping
# If img_LR (LR_path) doesn't exist, use img_HR (HR_path)
if LRHRchance < (1 - flip_chance):
HR_path = self.paths_HR[index]
LR_path = self.paths_LR[index]
if LR_path is None:
LR_path = HR_path
#print("HR kept")
# Flipped case:
# If img_HR (LR_path) doesn't exist, use img_HR (LR_path)
else:
HR_path = self.paths_LR[index]
LR_path = self.paths_HR[index]
if HR_path is None:
HR_path = LR_path
#print("HR flipped")
# Read the LR and HR images from the provided paths
img_LR = util.read_img(self.LR_env, LR_path, out_nc=image_channels)
img_HR = util.read_img(self.HR_env, HR_path, out_nc=image_channels)
# Even if LR dataset is provided, force to generate aug_downscale % of downscales OTF from HR
# The code will later make sure img_LR has the correct size
if self.opt.get('aug_downscale', None):
#aug_downscale = self.opt['aug_downscale']
if np.random.rand() < self.opt['aug_downscale']:
img_LR = img_HR
# If LR is not provided, use HR and modify on the fly
else:
HR_path = self.paths_HR[index]
img_HR = util.read_img(self.HR_env, HR_path, out_nc=image_channels)
img_LR = img_HR
LR_path = HR_path
# tmp_vis(img_HR, False)
# tmp_vis(img_LR, False)
######## Modify the images ########
# HR modcrop in the validation / test phase
if self.opt['phase'] != 'train':
img_HR = util.modcrop(img_HR, scale)
# change color space if necessary
# Note: Changing the LR colorspace here could make it so some colors are introduced when
# doing the augmentations later (ie: with Gaussian or Speckle noise), may be good if the
# model can learn to remove color noise in grayscale images, otherwise move to before
# converting to tensors
# self.opt['color'] For both LR and HR as in the the original code, kept for compatibility
# self.opt['color_HR'] and self.opt['color_LR'] for independent control
if self.opt.get('color', None): # Change both
img_HR = util.channel_convert(img_HR.shape[2], self.opt['color'],
[img_HR])[0]
img_LR = util.channel_convert(img_LR.shape[2], self.opt['color'],
[img_LR])[0]
if self.opt.get('color_HR', None): # Only change HR
img_HR = util.channel_convert(img_HR.shape[2],
self.opt['color_HR'], [img_HR])[0]
if self.opt.get('color_LR', None): # Only change LR
img_LR = util.channel_convert(img_LR.shape[2],
self.opt['color_LR'], [img_LR])[0]
######## Augmentations ########
#Augmentations during training
if self.opt['phase'] == 'train':
# Note: this is NOT recommended, HR should not be exposed to degradation, as it will
# degrade the model's results, only added because it exist as an option in downstream forks
# HR downscale
if self.opt.get('hr_downscale', None):
ds_algo = self.opt.get('hr_downscale_types', 777)
hr_downscale_amt = self.opt.get('hr_downscale_amt', 2)
if isinstance(hr_downscale_amt, list):
hr_downscale_amt = random.choice(hr_downscale_amt)
# will ignore if 1 or if result is smaller than hr size
if hr_downscale_amt > 1 and img_HR.shape[
0] // hr_downscale_amt >= HR_size and img_HR.shape[
1] // hr_downscale_amt >= HR_size:
#img_HR, hr_scale_interpol_algo = augmentations.scale_img(img_HR, hr_downscale_amt, algo=ds_algo)
img_HR, _ = Scale(img=img_HR,
scale=hr_downscale_amt,
algo=ds_algo)
# Downscales LR to match new size of HR if scale does not match after
if img_LR is not None and (
img_HR.shape[0] // scale != img_LR.shape[0]
or img_HR.shape[1] // scale != img_LR.shape[1]):
#img_LR, lr_scale_interpol_algo = augmentations.scale_img(img_LR, hr_downscale_amt, algo=ds_algo)
img_LR, _ = Scale(img=img_LR,
scale=hr_downscale_amt,
algo=ds_algo)
# Validate there's an img_LR, if not, use img_HR
if img_LR is None:
img_LR = img_HR
print("Image LR: ", LR_path, (
"was not loaded correctly, using HR pair to downscale on the fly."
))
# Check that HR and LR have the same dimensions ratio, else use an option to process
#TODO: add options variable
shape_change = self.opt.get('shape_change',
'reshape_hr') #"reshape_lr"
if img_HR.shape[0] // img_LR.shape[0] != img_HR.shape[
1] // img_LR.shape[1]:
if shape_change == "reshape_lr":
img_LR = transforms.Resize(
(int(img_HR.shape[0] / scale),
int(img_HR.shape[1] / scale)),
interpolation="BILINEAR")(np.copy(img_LR))
elif shape_change == "reshape_hr":
# print("pre: ", img_HR.shape[0], img_HR.shape[1])
nh = img_HR.shape[0] * (2 * img_LR.shape[1]) // (
img_LR.shape[0] + img_LR.shape[1])
nw = img_HR.shape[1] * (2 * img_LR.shape[0]) // (
img_LR.shape[0] + img_LR.shape[1])
#TODO: temporary change to test contextual loss with unaligned LR-HR pairs, forcing them to have the correct scale
#img_HR, _ = augmentations.resize_img(np.copy(img_HR), newdim=(img_LR.shape[1]*scale,img_LR.shape[0]*scale), algo=cv2.INTER_LINEAR)
#img_HR = transforms.Resize((img_LR.shape[0]*scale,img_LR.shape[1]*scale), interpolation="BILINEAR")(np.copy(img_HR))
img_HR = transforms.Resize(
(nh, nw), interpolation="BILINEAR")(np.copy(img_HR))
# print("post: ", img_HR.shape[0], img_HR.shape[1])
img_LR = transforms.Resize(
(int(img_HR.shape[0] / scale),
int(img_HR.shape[1] / scale)),
interpolation="BILINEAR")(np.copy(img_LR))
#TODO: check, this may not make sense
# elif shape_change == "pad_lr":
# LR_pad, _ = get_pad(img_LR, HR_size//scale, fill=0, padding_mode='constant')
# img_LR = LR_pad(np.copy(img_LR))
else: # generate new LR from HR
#TODO: disabled to test cx loss
#print("Warning: img_LR dimensions ratio does not match img_HR dimensions ratio for: ", HR_path)
img_LR = img_HR
# centercrop image
if self.opt['center_crop'] == True:
if img_HR.shape[0] > img_HR.shape[1]:
img_HR = transforms.CenterCrop(
(img_HR.shape[1], img_HR.shape[1]))(np.copy(img_HR))
img_LR = transforms.CenterCrop(
(img_HR.shape[1], img_HR.shape[1]))(np.copy(img_LR))
else:
img_HR = transforms.CenterCrop(
(img_HR.shape[0], img_HR.shape[0]))(np.copy(img_HR))
img_LR = transforms.CenterCrop(
(img_HR.shape[0], img_HR.shape[0]))(np.copy(img_LR))
# resize lr and hr image to given hr dimension
if self.opt['resize_HR_dimension'] == True:
img_HR = transforms.Resize(
(self.opt['HR_size'], self.opt['HR_size']),
interpolation="BILINEAR")(np.copy(img_HR))
img_LR = transforms.Resize(
(self.opt['HR_size'], self.opt['HR_size']),
interpolation="BILINEAR")(np.copy(img_LR))
# Random Crop (reduce computing cost and adjust images to correct size first)
if img_HR.shape[0] > HR_size or img_HR.shape[1] > HR_size:
#Here the scale should be in respect to the images, not to the training scale (in case they are being scaled on the fly)
scaleor = img_HR.shape[0] // img_LR.shape[0]
img_HR, img_LR = augmentations.random_crop_pairs(
img_HR, img_LR, HR_size, scaleor)
#TODO: test to be removed
# img_HR = transforms.Resize((HR_size-100,img_HR.shape[1]), interpolation="BILINEAR")(np.copy(img_HR))
# img_LR = transforms.Resize((LR_size-(100//scale),img_LR.shape[1]), interpolation="BILINEAR")(np.copy(img_LR))
# Or if the HR images are too small, Resize to the HR_size size and fit LR pair to LR_size too
#TODO: add options variable
dim_change = self.opt.get('dim_change', 'pad')
if img_HR.shape[0] < HR_size or img_HR.shape[1] < HR_size:
if dim_change == "resize":
#TODO: temp disabled to test
#print("Warning: Image: ", HR_path, " size does not match HR size: (", HR_size,"). The image size is: ", img_HR.shape)
# rescale HR image to the HR_size
# img_HR, _ = augmentations.resize_img(np.copy(img_HR), newdim=(HR_size,HR_size), algo=cv2.INTER_LINEAR)
img_HR = transforms.Resize(
(HR_size, HR_size),
interpolation="BILINEAR")(np.copy(img_HR))
# rescale LR image to the LR_size (The original code discarded the img_LR and generated a new one on the fly from img_HR)
#img_LR, _ = augmentations.resize_img(np.copy(img_LR), newdim=(LR_size,LR_size), algo=cv2.INTER_LINEAR)
img_LR = transforms.Resize(
(LR_size, LR_size),
interpolation="BILINEAR")(np.copy(img_LR))
elif dim_change == "pad":
# if img_LR is img_HR, padding will be wrong, downscaling LR before padding
if img_LR.shape[
0] != img_HR.shape[0] // scale or img_LR.shape[
1] != img_HR.shape[1] // scale:
ds_algo = 777 # default to matlab-like bicubic downscale
if self.opt.get(
'lr_downscale', None
): # if manually set and scale algorithms are provided, then:
ds_algo = self.opt.get('lr_downscale_types', 777)
if self.opt.get('lr_downscale', None) and self.opt.get(
'dataroot_kernels', None
) and 999 in self.opt["lr_downscale_types"]:
ds_kernel = self.ds_kernels
else:
ds_kernel = None
img_LR, _ = Scale(img=img_LR,
scale=scale,
algo=ds_algo,
ds_kernel=ds_kernel)
HR_pad, fill = get_pad(img_HR,
HR_size,
fill='random',
padding_mode=self.opt.get(
'pad_mode', 'constant'))
img_HR = HR_pad(np.copy(img_HR))
LR_pad, _ = get_pad(img_LR,
HR_size // scale,
fill=fill,
padding_mode=self.opt.get(
'pad_mode', 'constant'))
img_LR = LR_pad(np.copy(img_LR))
# (Randomly) scale LR (from HR) during training if :
# - LR dataset is not provided
# - LR dataset is not in the correct scale
# - Also to check if LR is not at the correct scale already (if img_LR was changed to img_HR)
if img_LR.shape[0] != LR_size or img_LR.shape[1] != LR_size:
ds_algo = 777 # default to matlab-like bicubic downscale
if self.opt.get(
'lr_downscale', None
): # if manually set and scale algorithms are provided, then:
ds_algo = self.opt.get('lr_downscale_types', 777)
else: # else, if for some reason img_LR is too large, default to matlab-like bicubic downscale
#if not self.opt['aug_downscale']: #only print the warning if not being forced to use HR images instead of LR dataset (which is a known case)
print(
"LR image is too large, auto generating new LR for: ",
LR_path)
if self.opt.get('lr_downscale', None) and self.opt.get(
'dataroot_kernels',
None) and 999 in self.opt["lr_downscale_types"]:
ds_kernel = self.ds_kernels #KernelDownscale(scale, self.kernel_paths, self.num_kernel)
else:
ds_kernel = None
#img_LR, scale_interpol_algo = augmentations.scale_img(img_LR, scale, algo=ds_algo)
#print("algo")
#print(ds_algo)
img_LR, _ = Scale(img=img_LR,
scale=scale,
algo=ds_algo,
ds_kernel=ds_kernel)
# resize = get_resize(size=LR_size, scale=scale, ds_algo=ds_algo)
# img_LR = resize(np.copy(img_LR))
#"""
# Rotations. 'use_flip' = 180 or 270 degrees (mirror), 'use_rot' = 90 degrees, 'HR_rrot' = random rotations +-45 degrees
if (self.opt['use_flip'] or self.opt['use_rot']) and self.opt.get(
'hr_rrot', None):
if np.random.rand() > 0.5:
img_LR, img_HR = util.augment([img_LR, img_HR], self.opt['use_flip'], \
self.opt['use_rot'])
else:
if np.random.rand(
) > 0.5: # randomize the random rotations, so half the images are the original
img_HR, img_LR = augmentations.random_rotate_pairs(
img_HR, img_LR, HR_size, scale)
elif (self.opt['use_flip'] or
self.opt['use_rot']) and not self.opt.get('hr_rrot', None):
# augmentation - flip, rotate
img_LR, img_HR = util.augment([img_LR, img_HR], self.opt['use_flip'], \
self.opt['use_rot'])
elif self.opt.get('hr_rrot', None):
if np.random.rand(
) > 0.5: # randomize the random rotations, so half the images are the original
img_HR, img_LR = augmentations.random_rotate_pairs(
img_HR, img_LR, HR_size, scale)
# Final sizes checks
# if the resulting HR image size so far is too large or too small, resize HR to the correct size and downscale to generate a new LR on the fly
# if the resulting LR so far does not have the correct dimensions, also generate a new HR-LR image pair on the fly
if img_HR.shape[0] != HR_size or img_HR.shape[
1] != HR_size or img_LR.shape[
0] != LR_size or img_LR.shape[1] != LR_size:
#if img_HR.shape[0] != HR_size or img_HR.shape[1] != HR_size:
#TODO: temp disabled to test
#print("Image: ", HR_path, " size does not match HR size: (", HR_size,"). The image size is: ", img_HR.shape)
#if img_LR.shape[0] != LR_size or img_LR.shape[0] != LR_size:
#TODO: temp disabled to test
#print("Image: ", LR_path, " size does not match LR size: (", HR_size//scale,"). The image size is: ", img_LR.shape)
# rescale HR image to the HR_size (should not be needed in LR case, but something went wrong before, just for sanity)
#img_HR, _ = augmentations.resize_img(np.copy(img_HR), newdim=(HR_size,HR_size), algo=cv2.INTER_LINEAR)
img_HR = transforms.Resize(
(HR_size, HR_size),
interpolation="BILINEAR")(np.copy(img_HR))
# if manually provided and scale algorithms are provided, then use it, else use matlab imresize to generate LR pair
ds_algo = self.opt.get('lr_downscale_types', 777)
#img_LR, _ = augmentations.scale_img(img_HR, scale, algo=ds_algo)
img_LR, _ = Scale(img_HR, scale, algo=ds_algo)
# Below are the On The Fly augmentations
# Apply "auto levels" to images
#rand_levels = (1 - self.opt.get('rand_auto_levels', 0)) # Randomize for augmentation
rand_levels = self.opt.get('rand_auto_levels', 0)
#rand_percent = self.opt.get('rand_auto_per', 10)
#if self.opt.get('auto_levels', None) and np.random.rand() > rand_levels:
if self.opt.get('auto_levels', None) == 'HR' or self.opt.get(
'auto_levels', None) == 'Both':
#img_HR = augmentations.simplest_cb(img_HR, znorm=znorm) #TODO: now images are processed in the [0,255] range
#img_HR = augmentations.simplest_cb(img_HR)
img_HR = transforms.FilterColorBalance(
p=rand_levels, percent=10, random_params=True)(img_HR)
if self.opt.get('auto_levels', None) == 'LR' or self.opt.get(
'auto_levels', None) == 'Both':
#img_LR = augmentations.simplest_cb(img_LR, znorm=znorm) #TODO: now images are processed in the [0,255] range
#img_LR = augmentations.simplest_cb(img_LR)
img_LR = transforms.FilterColorBalance(
p=rand_levels, percent=10, random_params=True)(img_LR)
# Apply unsharpening mask to LR images
'''
rand_unsharp = (1 - self.opt.get('lr_rand_unsharp', 0)) # Randomize for augmentation
if self.opt.get('lr_unsharp_mask', None) and np.random.rand() > rand_unsharp:
#img_LR = augmentations.unsharp_mask(img_LR, znorm=znorm) #TODO: now images are processed in the [0,255] range
img_LR = augmentations.unsharp_mask(img_LR)
'''
if self.opt.get('lr_unsharp_mask', None):
lr_rand_unsharp = self.opt.get('lr_rand_unsharp', 0)
img_LR = transforms.FilterUnsharp(p=lr_rand_unsharp)(img_LR)
# Apply unsharpening mask to HR images
'''
rand_unsharp = (1 - self.opt.get('hr_rand_unsharp', 0)) # Randomize for augmentation
if self.opt.get('hr_unsharp_mask', None) and np.random.rand() > rand_unsharp:
#img_HR = augmentations.unsharp_mask(img_HR, znorm=znorm) #TODO: now images are processed in the [0,255] range
img_HR = augmentations.unsharp_mask(img_HR)
'''
if self.opt.get('hr_unsharp_mask', None):
hr_rand_unsharp = self.opt.get('hr_rand_unsharp', 0)
img_HR = transforms.FilterUnsharp(p=hr_rand_unsharp)(img_HR)
# Add noise to HR if enabled AND noise types are provided (for noise2noise and similar)
if self.opt.get('hr_noise', None):
noise_option = get_noise(self.opt.get('hr_noise_types', None))
if noise_option:
#img_HR, hr_noise_algo = augmentations.noise_img(img_HR, noise_types=self.opt['hr_noise_types'])
img_HR = noise_option(img_HR)
#TODO: check if necessary
# else:
# print("Noise types 'hr_noise_types' not defined. Skipping OTF noise for HR.")
# Create color fringes
# Caution: Can easily destabilize a model
# Only applied to a small % of the images. Around 20% and 50% appears to be stable.
if self.opt.get('lr_fringes', None):
lr_fringes_chance = self.opt['lr_fringes_chance'] if self.opt[
'lr_fringes_chance'] else 0.4
if np.random.rand() > (1. - lr_fringes_chance):
img_LR = augmentations.translate_chan(img_LR)
#"""
#v LR blur AND blur types are provided, else will skip
if self.opt.get('lr_blur', None):
blur_option = get_blur(self.opt.get('lr_blur_types', None))
if blur_option:
#img_LR, blur_algo, blur_kernel_size = augmentations.blur_img(img_LR, blur_algos=blur_algos)
img_LR = blur_option(img_LR)
#TODO: check if necessary
# else:
# print("Blur types 'lr_blur_types' not defined. Skipping OTF blur.")
#"""
#"""
#v LR primary noise: Add noise to LR if enabled AND noise types are provided, else will skip
if self.opt.get('lr_noise', None):
noise_option = get_noise(self.opt.get('lr_noise_types', None),
self.noise_patches)
if noise_option:
#img_LR, noise_algo = augmentations.noise_img(img_LR, noise_types=self.opt['lr_noise_types'])
img_LR = noise_option(img_LR)
#TODO: check if necessary
# else:
# print("Noise types 'lr_noise_types' not defined. Skipping OTF noise.")
#v LR secondary noise: Add additional noise to LR if enabled AND noise types are provided, else will skip
if self.opt.get('lr_noise2', None):
noise_option = get_noise(self.opt.get('lr_noise_types2', None),
self.noise_patches)
if noise_option:
#img_LR, noise_algo2 = augmentations.noise_img(img_LR, noise_types=self.opt['lr_noise_types2'])
img_LR = noise_option(img_LR)
#TODO: check if necessary
# else:
# print("Noise types 'lr_noise_types2' not defined. Skipping OTF secondary noise.")
#"""
#"""
#v LR cutout / LR random erasing (for inpainting/classification tests)
if self.opt.get(
'lr_cutout',
None) and (self.opt.get('lr_erasing', None) != True):
# img_LR = augmentations.cutout(img_LR, img_LR.shape[0] // 2)
img_LR = transforms.Cutout(p=1, mask_size=img_LR.shape[0] //
2)(img_LR)
elif self.opt.get(
'lr_erasing',
None) and (self.opt.get('lr_cutout', None) != True):
# img_LR = augmentations.random_erasing(img_LR)
img_LR = transforms.RandomErasing(p=1)(img_LR, mode=[3])
elif self.opt.get('lr_cutout', None) and self.opt.get(
'lr_erasing', None):
if np.random.rand(
) > 0.5: #only do cutout or erasing, not both at the same time
#img_LR = augmentations.cutout(img_LR, img_LR.shape[0] // 2, p=0.5)
img_LR = transforms.Cutout(p=1,
mask_size=img_LR.shape[0] //
2)(img_LR)
else:
#img_LR = augmentations.random_erasing(img_LR, p=0.5, modes=[3])
img_LR = transforms.RandomErasing(p=1)(img_LR, mode=[3])
#"""
# For testing and validation
if self.opt['phase'] != 'train':
# Randomly downscale LR if enabled
if self.opt['lr_downscale']:
# if self.opt['lr_downscale_types']:
# img_LR, scale_interpol_algo = augmentations.scale_img(img_LR, scale, algo=self.opt['lr_downscale_types'])
# else: # Default to matlab-like bicubic downscale
# img_LR, scale_interpol_algo = augmentations.scale_img(img_LR, scale, algo=777)
img_LR, _ = Scale(img_LR,
scale,
algo=self.opt.get('lr_downscale_types', 777))
# Alternative position for changing the colorspace of LR.
# if self.opt['color_LR']: # Only change LR
# img_LR = util.channel_convert(img_LR.shape[2], self.opt['color'], [img_LR])[0]
# Debug #TODO: use the debugging functions to visualize or save images instead
# Save img_LR and img_HR images to a directory to visualize what is the result of the on the fly augmentations
# DO NOT LEAVE ON DURING REAL TRAINING
# self.output_sample_imgs = True
if self.opt['phase'] == 'train':
if self.output_sample_imgs:
import os
# LR_dir, im_name = os.path.split(LR_path)
HR_dir, im_name = os.path.split(HR_path)
#baseHRdir, _ = os.path.split(HR_dir)
#debugpath = os.path.join(baseHRdir, os.sep, 'sampleOTFimgs')
# debugpath = os.path.join(os.path.split(LR_dir)[0], 'sampleOTFimgs')
debugpath = os.path.join('D:/tmp_test', 'sampleOTFimgs')
#print(debugpath)
if not os.path.exists(debugpath):
os.makedirs(debugpath)
import uuid
hex = uuid.uuid4().hex
cv2.imwrite(debugpath + "\\" + im_name + hex + '_LR.png',
img_LR) #random name to save
cv2.imwrite(debugpath + "\\" + im_name + hex + '_HR.png',
img_HR) #random name to save
# cv2.imwrite(debugpath+"\\"+im_name+hex+'_HR1.png',img_HRn1) #random name to save
######## Convert images to PyTorch Tensors ########
""" # for debugging
if (img_HR.min() < -1):
describe_numpy(img_HR, all=True)
print(HR_path)
if (img_HR.max() > 1):
describe_numpy(img_HR, all=True)
print(HR_path)
if (img_LR.min() < -1):
describe_numpy(img_LR, all=True)
print(LR_path)
if (img_LR.max() > 1):
describe_numpy(img_LR, all=True)
print(LR_path)
#"""
# check for grayscale images #TODO: should not be needed anymore
if len(img_HR.shape) < 3:
img_HR = img_HR[..., np.newaxis]
if len(img_LR.shape) < 3:
img_LR = img_LR[..., np.newaxis]
# tmp_vis(img_HR, False)
# tmp_vis(img_LR, False)
# edge-connect
if self.opt['training_with_canny'] == True:
img_HR_gray = cv2.cvtColor(img_HR, cv2.COLOR_BGR2GRAY)
img_HR_canny = cv2.Canny(img_HR_gray, 100, 150)
img_HR_gray = torch.from_numpy(img_HR_gray).unsqueeze(0)
img_HR_canny = torch.from_numpy(img_HR_canny).unsqueeze(0)
if self.opt['training_with_canny_SR'] == True:
img_LR_gray = cv2.cvtColor(img_LR, cv2.COLOR_BGR2GRAY)
img_LR_canny = cv2.Canny(img_LR_gray, 100, 150)
img_LR_canny = torch.from_numpy(img_LR_canny).unsqueeze(0)
if self.opt['noise_estimation'] == True:
ds_kernel = torch.from_numpy(getattr(
ds_kernel, 'used_kernel')).unsqueeze(0).float()
noise_est = noise_estimate(img_LR, 4)
sigma = torch.tensor(noise_est).float().view([1, 1, 1])
img_HR = util.np2tensor(img_HR, normalize=znorm,
add_batch=False) #.astype('uint8').clip(0,255)
img_LR = util.np2tensor(img_LR, normalize=znorm, add_batch=False)
if LR_path is None:
LR_path = HR_path
if self.opt['training_with_canny'] == True:
return {
'LR': img_LR,
'HR': img_HR,
'LR_path': LR_path,
'HR_path': HR_path,
'img_HR_gray': img_HR_gray,
'img_HR_canny': img_HR_canny
}
elif self.opt['training_with_canny_SR'] == True:
return {
'LR': img_LR,
'HR': img_HR,
'LR_path': LR_path,
'HR_path': HR_path,
'img_LR_canny': img_LR_canny
}
if self.opt['noise_estimation'] == True:
return {
'LR': img_LR,
'HR': img_HR,
'LR_path': LR_path,
'HR_path': HR_path,
'ds_kernel': ds_kernel,
'sigma': sigma
}
else:
return {
'LR': img_LR,
'HR': img_HR,
'LR_path': LR_path,
'HR_path': HR_path
}
def __getitem__(self, index):
"""Return a data point and its metadata information.
Parameters:
index (int): a random integer for data indexing
Returns a dictionary that contains A, B, A_paths and B_paths
(or LR, HR, LR_paths and HR_paths)
A (tensor): an image in the input domain
B (tensor): its corresponding image in the target domain
A_paths (str): paths A images
B_paths (str): paths B images (can be same as A_paths if
using single images)
"""
scale = self.opt.get('scale')
######## Read the images ########
if self.AB_paths:
img_A, img_B, A_path, B_path = read_split_single_dataset(
self.opt, index, self.AB_paths, self.AB_env)
else:
img_A, img_B, A_path, B_path = read_imgs_from_path(
self.opt, index, self.A_paths, self.B_paths, self.A_env,
self.B_env)
######## Modify the images ########
# for the validation / test phases
if self.opt['phase'] != 'train':
img_type = image_type(img_B)
# B/HR modcrop
img_B = modcrop(img_B, scale=scale, img_type=img_type)
# modcrop and downscale A/LR if enabled
if self.opt['lr_downscale']:
img_A = modcrop(img_A, scale=scale, img_type=img_type)
# TODO: 'pil' images will use default method for scaling
img_A, _ = Scale(img_A,
scale,
algo=self.opt.get('lr_downscale_types', 777),
img_type=img_type)
# change color space if necessary
# TODO: move to get_transform()
color_B = self.opt.get('color', None) or self.opt.get('color_HR', None)
if color_B:
img_B = channel_convert(image_channels(img_B), color_B, [img_B])[0]
color_A = self.opt.get('color', None) or self.opt.get('color_LR', None)
if color_A:
img_A = channel_convert(image_channels(img_A), color_A, [img_A])[0]
######## Augmentations ########
#Augmentations during training
if self.opt['phase'] == 'train':
default_int_method = get_default_imethod(image_type(img_A))
# random HR downscale
img_A, img_B = random_downscale_B(img_A=img_A,
img_B=img_B,
opt=self.opt)
# validate there's an img_A, if not, use img_B
if img_A is None:
img_A = img_B
print(
f"Image A: {A_path} was not loaded correctly, using B pair to downscale on the fly."
)
# validate proper dimensions between paired images, generate A if needed
img_A, img_B = paired_imgs_check(img_A,
img_B,
opt=self.opt,
ds_kernels=self.ds_kernels)
# get and apply the paired transformations below
transform_params = get_params(scale_opt(self.opt, scale),
image_size(img_A))
A_transform = get_transform(
scale_opt(self.opt, scale),
transform_params,
# grayscale=(input_nc == 1),
method=default_int_method)
B_transform = get_transform(
self.opt,
scale_params(transform_params, scale),
# grayscale=(output_nc == 1),
method=default_int_method)
img_A = A_transform(img_A)
img_B = B_transform(img_B)
# Below are the On The Fly augmentations
# get and apply the unpaired transformations below
a_aug_params, b_aug_params = get_unpaired_params(self.opt)
a_augmentations = get_augmentations(
self.opt,
params=a_aug_params,
noise_patches=self.noise_patches,
)
b_augmentations = get_augmentations(
self.opt,
params=b_aug_params,
noise_patches=self.noise_patches,
)
img_A = a_augmentations(img_A)
img_B = b_augmentations(img_B)
# Alternative position for changing the colorspace of A/LR.
# color_A = self.opt.get('color', None) or self.opt.get('color_A', None)
# if color_A:
# img_A = channel_convert(image_channels(img_A), color_A, [img_A])[0]
######## Convert images to PyTorch Tensors ########
tensor_transform = get_totensor(self.opt,
params=self.totensor_params,
toTensor=True,
grayscale=False)
img_A = tensor_transform(img_A)
img_B = tensor_transform(img_B)
if A_path is None:
A_path = B_path
if self.vars == 'AB':
return {'A': img_A, 'B': img_B, 'A_path': A_path, 'B_path': B_path}
else:
return {
'LR': img_A,
'HR': img_B,
'LR_path': A_path,
'HR_path': B_path
}
def __getitem__(self, index):
"""Return a data point and its metadata information.
Parameters:
index (int): a random integer for data indexing
Returns a dictionary that contains A, B, A_paths and B_paths
A (tensor): an image in the input domain
B (tensor): its corresponding image in the target domain
A_paths (str): paths A images
B_paths (str): paths B images
"""
scale = self.opt.get('scale')
######## Read the images ########
img_A, A_path = read_single_dataset(opt=self.opt,
index=index,
paths=self.A_paths,
env=self.A_env,
idx_case='inrange',
d_size=self.A_size)
img_B, B_path = read_single_dataset(opt=self.opt,
index=index,
paths=self.B_paths,
env=self.B_env,
idx_case=self.idx_case,
d_size=self.B_size)
######## Modify the images ########
# change color space if necessary
# TODO: move to get_transform()
color_B = self.opt.get('color', None) or self.opt.get('color_B', None)
if color_B:
img_B = channel_convert(image_channels(img_B), color_B, [img_B])[0]
color_A = self.opt.get('color', None) or self.opt.get('color_A', None)
if color_A:
img_A = channel_convert(image_channels(img_A), color_A, [img_A])[0]
# apply image transformation
default_int_method = get_default_imethod(image_type(img_A))
# get first set of random params
transform_params_A = get_params(scale_opt(self.opt, scale),
image_size(img_A))
# get second set of random params
transform_params_B = get_params(self.opt, image_size(img_B))
A_transform = get_transform(
scale_opt(self.opt, scale),
transform_params_A,
# grayscale=(input_nc == 1),
method=default_int_method)
B_transform = get_transform(
self.opt,
transform_params_B,
# grayscale=(output_nc == 1),
method=default_int_method)
img_A = A_transform(img_A)
img_B = B_transform(img_B)
#TODO: not needed initially, but available
# get and apply the unpaired transformations below
# a_aug_params, b_aug_params = get_unpaired_params(self.opt)
# a_augmentations = get_augmentations(
# self.opt,
# params=a_aug_params,
# noise_patches=self.noise_patches,
# )
# b_augmentations = get_augmentations(
# self.opt,
# params=b_aug_params,
# noise_patches=self.noise_patches,
# )
# img_A = a_augmentations(img_A)
# img_B = b_augmentations(img_B)
######## Convert images to PyTorch Tensors ########
tensor_transform = get_totensor(self.opt,
params=self.totensor_params,
toTensor=True,
grayscale=False)
img_A = tensor_transform(img_A)
img_B = tensor_transform(img_B)
if self.vars == 'AB':
return {'A': img_A, 'B': img_B, 'A_path': A_path, 'B_path': B_path}
else:
return {
'LR': img_A,
'HR': img_B,
'LR_path': A_path,
'HR_path': B_path
}
def __getitem__(self, index):
HR_path, LR_path = None, None
scale = self.opt['scale']
HR_size = self.opt['HR_size']
# get HR image
HR_path = self.paths_HR[index]
img_HR = util.read_img(self.HR_env, HR_path)
# modcrop in the validation / test phase
if self.opt['phase'] != 'train':
img_HR = util.modcrop(img_HR, scale)
# change color space if necessary
if self.opt['color']:
img_HR = util.channel_convert(img_HR.shape[2], self.opt['color'],
[img_HR])[0]
# get LR image
if self.paths_LR:
LR_path = self.paths_LR[index]
img_LR = util.read_img(self.LR_env, LR_path)
else: # down-sampling on-the-fly
# randomly scale during training
if self.opt['phase'] == 'train':
random_scale = random.choice(self.random_scale_list)
H_s, W_s, _ = img_HR.shape
def _mod(n, random_scale, scale, thres):
rlt = int(n * random_scale)
rlt = (rlt // scale) * scale
return thres if rlt < thres else rlt
H_s = _mod(H_s, random_scale, scale, HR_size)
W_s = _mod(W_s, random_scale, scale, HR_size)
img_HR = cv2.resize(np.copy(img_HR), (W_s, H_s),
interpolation=cv2.INTER_LINEAR)
# force to 3 channels
if img_HR.ndim == 2:
img_HR = cv2.cvtColor(img_HR, cv2.COLOR_GRAY2BGR)
H, W, _ = img_HR.shape
# using matlab imresize
img_LR = util.imresize_np(img_HR, 1 / scale, True)
if img_LR.ndim == 2:
img_LR = np.expand_dims(img_LR, axis=2)
if self.opt['phase'] == 'train':
# if the image size is too small
H, W, _ = img_HR.shape
if H < HR_size or W < HR_size:
img_HR = cv2.resize(np.copy(img_HR), (HR_size, HR_size),
interpolation=cv2.INTER_LINEAR)
# using matlab imresize
img_LR = util.imresize_np(img_HR, 1 / scale, True)
if img_LR.ndim == 2:
img_LR = np.expand_dims(img_LR, axis=2)
H, W, C = img_LR.shape
LR_size = HR_size // scale
# randomly crop
rnd_h = random.randint(0, max(0, H - LR_size))
rnd_w = random.randint(0, max(0, W - LR_size))
img_LR = img_LR[rnd_h:rnd_h + LR_size, rnd_w:rnd_w + LR_size, :]
rnd_h_HR, rnd_w_HR = int(rnd_h * scale), int(rnd_w * scale)
img_HR = img_HR[rnd_h_HR:rnd_h_HR + HR_size,
rnd_w_HR:rnd_w_HR + HR_size, :]
# augmentation - flip, rotate
img_LR, img_HR = util.augment([img_LR, img_HR], self.opt['use_flip'], \
self.opt['use_rot'])
# change color space if necessary
if self.opt['color']:
img_LR = util.channel_convert(
C, self.opt['color'],
[img_LR])[0] # TODO during val no definetion
# BGR to RGB, HWC to CHW, numpy to tensor
if img_HR.shape[2] == 3:
img_HR = img_HR[:, :, [2, 1, 0]]
img_LR = img_LR[:, :, [2, 1, 0]]
elif img_LR.shape[2] == 4:
img_HR = img_HR[:, :, [2, 1, 0, 3]]
img_LR = img_LR[:, :, [2, 1, 0, 3]]
img_HR = torch.from_numpy(
np.ascontiguousarray(np.transpose(img_HR, (2, 0, 1)))).float()
img_LR = torch.from_numpy(
np.ascontiguousarray(np.transpose(img_LR, (2, 0, 1)))).float()
if LR_path is None:
LR_path = HR_path
return {
'LR': img_LR,
'HR': img_HR,
'LR_path': LR_path,
'HR_path': HR_path
}
