Add dataset loading compatible with finetuning folder structure
This commit is contained in:
parent
07fbb2b72a
commit
5a278993e4
|
@ -0,0 +1,220 @@
|
|||
import os
|
||||
import random
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.utils.data as data
|
||||
import utils.utils_image as util
|
||||
from utils.DTutils import is_natural_patch
|
||||
import itertools
|
||||
|
||||
|
||||
class DatasetFFDNet(data.Dataset):
|
||||
"""
|
||||
# -----------------------------------------
|
||||
# Get L/H/M for denosing on AWGN with a range of sigma.
|
||||
# Only dataroot_H is needed.
|
||||
# -----------------------------------------
|
||||
# e.g., FFDNet, H = f(L, sigma), sigma is noise level
|
||||
# -----------------------------------------
|
||||
"""
|
||||
|
||||
def __init__(self, opt):
|
||||
super(DatasetFFDNet, self).__init__()
|
||||
self.opt = opt
|
||||
self.n_channels = opt['n_channels'] if opt['n_channels'] else 3
|
||||
self.n_channels_datasetload = opt['n_channels_datasetload'] if opt['n_channels_datasetload'] else 3
|
||||
self.patch_size = self.opt['H_size'] if opt['H_size'] else 64
|
||||
self.sigma = opt['sigma'] if opt['sigma'] else [0, 75]
|
||||
self.sigma_min, self.sigma_max = self.sigma[0], self.sigma[1]
|
||||
self.sigma_test = opt['sigma_test'] if opt['sigma_test'] else 0
|
||||
self.use_all_patches = opt['use_all_patches'] if opt['use_all_patches'] else False
|
||||
self.num_patches_per_image = opt['num_patches_per_image'] if opt['num_patches_per_image'] else 100
|
||||
self.skip_natural_patches = opt['skip_natural_patches'] if opt['skip_natural_patches'] else False
|
||||
|
||||
# -------------------------------------
|
||||
# Dataset path contains all H images and subfolders for every single one with one or more
|
||||
# -------------------------------------
|
||||
|
||||
"""
|
||||
Dataset path includes all H images and L-subfolders.
|
||||
Every H image has one L-subfolder assosiated, which
|
||||
contains one or more L representations of the H image.
|
||||
"""
|
||||
|
||||
|
||||
self.paths_H = [f for f in os.listdir(opt['dataroot_H']) if os.path.isfile(f)]
|
||||
#------------------------------------------------------------------------------------------------------
|
||||
# For the above step one can use util.get_image_paths(), but it goes recursevely thought the tree dirs
|
||||
#------------------------------------------------------------------------------------------------------
|
||||
|
||||
paths_H_aux = []
|
||||
self.paths_L = []
|
||||
|
||||
# Iterate over all image paths
|
||||
for H_file in self.paths_H:
|
||||
# filename = os.path.basename(H_file)
|
||||
filename = H_file.split(".png") # TODO: the correct way to do it is with os.path.basename()
|
||||
L_folder = os.path.join(opt['dataroot_H'],filename)
|
||||
# For image at subfolder, append to L paths and repeat current H path
|
||||
for L_file in os.listdir(L_folder):
|
||||
L_filepath = os.path.join(L_folder,L_file)
|
||||
paths_H_aux.append(H_file)
|
||||
self.paths_L.append(L_filepath)
|
||||
|
||||
# Update H paths
|
||||
self.paths_H = paths_H_aux
|
||||
|
||||
# Repeat every image in path list to get more than one patch per image
|
||||
if self.opt['phase'] == 'train':
|
||||
listOfLists = [list(itertools.repeat(path, self.num_patches_per_image)) for path in self.paths_H]
|
||||
self.paths_H = list(itertools.chain.from_iterable(listOfLists))
|
||||
|
||||
listOfLists = [list(itertools.repeat(path, self.num_patches_per_image)) for path in self.paths_L]
|
||||
self.paths_L = list(itertools.chain.from_iterable(listOfLists))
|
||||
|
||||
def __getitem__(self, index):
|
||||
|
||||
# -------------------------------------
|
||||
# get H and L image
|
||||
# -------------------------------------
|
||||
H_path = self.paths_H[index]
|
||||
L_path = self.paths_L[index]
|
||||
|
||||
H_file, L_file = H_path.split('/')[-1], L_path.split('/')[-1]
|
||||
H_name, L_name = H_file.split('.')[0], L_file.split('.')[0]
|
||||
|
||||
assert H_name==L_name, f'Both high and low quality images MUST have same name.\nGot {H_name} and {L_name} respectively.'
|
||||
|
||||
img_H = util.imread_uint(H_path, self.n_channels_datasetload)
|
||||
|
||||
|
||||
img_L = util.imread_uint(L_path, self.n_channels_datasetload)[:,:,:2]
|
||||
|
||||
# Temp solution for blanking images
|
||||
L_v, L_h = img_L.shape[:2]
|
||||
|
||||
if L_v==1000 and L_h==1800:
|
||||
img_L = img_L[(1000-900)//2:-(1000-900)//2,(1800-1600)//2:-(1800-1600)//2,:]
|
||||
|
||||
# Get module of complex image, stretch and to uint8
|
||||
# img_L = img_L.astype('float')
|
||||
# img_L = np.abs(img_L[:,:,0]+1j*img_L[:,:,1])
|
||||
# img_L = 255*(img_L - img_L.min())/(img_L.max() - img_L.min())
|
||||
# img_L = img_L.astype('uint8')
|
||||
|
||||
if self.opt['phase'] == 'train':
|
||||
"""
|
||||
# --------------------------------
|
||||
# get L/H/M patch pairs
|
||||
# --------------------------------
|
||||
"""
|
||||
H, W = img_H.shape[:2]
|
||||
|
||||
if self.use_all_patches or (img_H.shape[0] <= self.patch_size) or (img_H.shape[1] <= self.patch_size):
|
||||
|
||||
# ---------------------------------
|
||||
# Start or continue image patching
|
||||
# ---------------------------------
|
||||
img_patch_index = index % self.num_patches_per_image # Resets to 0 every time index overflows num_patches
|
||||
|
||||
# Upper-left corner of patch
|
||||
h_index = self.patch_size * ( (img_patch_index * self.patch_size) // W)
|
||||
w_index = self.patch_size * ( ( (img_patch_index * self.patch_size) % W ) // self.patch_size)
|
||||
|
||||
# Dont exceed the image limit
|
||||
h_index = min(h_index, H - self.patch_size)
|
||||
w_index = min(w_index, W - self.patch_size)
|
||||
|
||||
### Keep text patches only (non-natural images)
|
||||
if self.skip_natural_patches:
|
||||
|
||||
# Check if selected patch is natural, based on RGB entropy
|
||||
is_natural = is_natural_patch(img_H[h_index:h_index + self.patch_size, w_index:w_index + self.patch_size, :])
|
||||
|
||||
# If natural, select random patch and keep trying until non-natural or reaching max attempts
|
||||
attempt = 0
|
||||
max_attempts = 10
|
||||
while is_natural and (attempt < max_attempts):
|
||||
h_index = random.randint(0, max(0, H - self.patch_size))
|
||||
w_index = random.randint(0, max(0, W - self.patch_size))
|
||||
is_natural = is_natural_patch(img_H[h_index:h_index + self.patch_size, w_index:w_index + self.patch_size, :])
|
||||
attempt += 1
|
||||
|
||||
|
||||
else:
|
||||
# ---------------------------------
|
||||
# randomly crop the patch
|
||||
# ---------------------------------
|
||||
h_index = random.randint(0, max(0, H - self.patch_size))
|
||||
w_index = random.randint(0, max(0, W - self.patch_size))
|
||||
|
||||
# Ground-truth as channels mean
|
||||
patch_H = np.mean(img_H[h_index:h_index + self.patch_size, w_index:w_index + self.patch_size, :],axis=2)
|
||||
|
||||
# Get the patch from the simulation
|
||||
patch_L = img_L[h_index:h_index + self.patch_size, w_index:w_index + self.patch_size,:]
|
||||
|
||||
# ---------------------------------
|
||||
# HWC to CHW, numpy(uint) to tensor
|
||||
# ---------------------------------
|
||||
img_H = util.uint2tensor3(patch_H)
|
||||
img_L = util.uint2tensor3(patch_L)
|
||||
|
||||
# ---------------------------------
|
||||
# get noise level
|
||||
# ---------------------------------
|
||||
noise_level = torch.FloatTensor([int(np.random.uniform(self.sigma_min, self.sigma_max))])/255.0
|
||||
# noise_level = torch.FloatTensor([np.random.randint(self.sigma_min, self.sigma_max)])/255.0
|
||||
if (self.sigma_max != 0):
|
||||
# ---------------------------------
|
||||
# add noise
|
||||
# ---------------------------------
|
||||
noise = torch.randn(img_L.size()).mul_(noise_level).float()
|
||||
img_L.add_(noise)
|
||||
|
||||
else:
|
||||
"""
|
||||
# --------------------------------
|
||||
# get L/H/sigma image pairs
|
||||
# --------------------------------
|
||||
"""
|
||||
|
||||
# Ground-truth as mean value of RGB channels
|
||||
img_H = np.mean(img_H,axis=2)
|
||||
img_H = img_H[:,:,np.newaxis]
|
||||
|
||||
# ---------------------------------
|
||||
# HWC to CHW, numpy(uint) to tensor
|
||||
# ---------------------------------
|
||||
img_H = util.uint2tensor3(img_H)
|
||||
img_L = util.uint2tensor3(img_L)
|
||||
|
||||
# img_H = util.uint2single(img_H)
|
||||
|
||||
# img_L = util.uint2single(img_L)
|
||||
|
||||
|
||||
# ---------------------------------
|
||||
# get noise level
|
||||
# ---------------------------------
|
||||
|
||||
noise_level = torch.FloatTensor([int(self.sigma_test)])/255.0
|
||||
if self.sigma_test != 0:
|
||||
|
||||
# noise_level = torch.FloatTensor([np.random.randint(self.sigma_min, self.sigma_max)])/255.0
|
||||
|
||||
# ---------------------------------
|
||||
# add noise
|
||||
# ---------------------------------
|
||||
noise = torch.randn(img_L.size()).mul_(noise_level).float()
|
||||
img_L.add_(noise)
|
||||
|
||||
|
||||
|
||||
noise_level = noise_level.unsqueeze(1).unsqueeze(1)
|
||||
|
||||
|
||||
return {'L': img_L, 'H': img_H, 'C': noise_level, 'L_path': L_path, 'H_path': H_path}
|
||||
|
||||
def __len__(self):
|
||||
return len(self.paths_H)
|
Loading…
Reference in New Issue