conditional_score_based_diffusion/create_dataset.py at true_master · Mehdi-Letafati/conditional_score_based_diffusion · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
#testing code

# load and show an image with Pillow
from PIL import Image
import torch
import numpy as np
from iunets.layers import InvertibleDownsampling2D
from torchvision.utils import make_grid, save_image
from pathlib import Path
import torch
from torch import nn
import os
from tqdm import tqdm
import matplotlib.pyplot as plt
from argparse import ArgumentParser

def permute_channels(haar_image, forward=True):
        permuted_image = torch.zeros_like(haar_image)
        if forward:
            for i in range(4):
                if i == 0:
                    k = 1
                elif i == 1:
                    k = 0
                else:
                    k = i
                for j in range(3):
                    permuted_image[:, 3*k+j, :, :] = haar_image[:, 4*j+i, :, :]
        else:
            for i in range(4):
                if i == 0:
                    k = 1
                elif i == 1:
                    k = 0
                else:
                    k = i

                for j in range(3):
                    permuted_image[:,4*j+k,:,:] = haar_image[:, 3*i+j, :, :]

        return permuted_image

def normalise(x, value_range=None):
    if value_range is None:
        x -= x.min()
        x /= x.max()
    else:
        x -= value_range[0]
        x /= value_range[1]
    return x

def create_train_val_test_index_dict(total_num_images, split):
        #return a dictionary that maps each index to the corresponding phase dataset (train, val, test)
        indices = np.arange(total_num_images)
        np.random.shuffle(indices) #in-place operation
        phase_dataset = {}
        for counter, index in enumerate(indices):
            if counter < split[0]*total_num_images:
                folder = 'train'
            elif counter < (split[0]+split[1])*total_num_images:
                folder = 'val'
            else:
                folder = 'test'
            phase_dataset[index] = folder
        return phase_dataset

def create_level_folders(base_image_dir, dataset, target_resolution, levels):
    for i in range(0, levels+1):
        intermediate_resolution = target_resolution // 2**i
        Path(os.path.join(base_image_dir, dataset+'_'+str(intermediate_resolution))).mkdir(parents=True, exist_ok=True)
        Path(os.path.join(base_image_dir, dataset+'_'+str(intermediate_resolution), 'train')).mkdir(parents=True, exist_ok=True)
        Path(os.path.join(base_image_dir, dataset+'_'+str(intermediate_resolution), 'val')).mkdir(parents=True, exist_ok=True)
        Path(os.path.join(base_image_dir, dataset+'_'+str(intermediate_resolution), 'test')).mkdir(parents=True, exist_ok=True)

def center_crop(img, crop_left, crop_right, crop_top, crop_bottom):
    width, height = img.size
    left = crop_left
    right = width - crop_right
    top = crop_top
    bottom = height - crop_bottom
    return img.crop((left, top, right, bottom))


def create_haar_dataset(base_image_dir, dataset, target_resolution, levels, split):
    create_level_folders(base_image_dir, dataset, target_resolution, levels)

    haar_transform = InvertibleDownsampling2D(3, stride=2, method='cayley', init='haar', learnable=False)

    haar_level_ranges={}
    approx_level_ranges={}

    total_num_images = len([path for path in Path(os.path.join(base_image_dir, dataset)).iterdir()])
    phase_dataset = create_train_val_test_index_dict(total_num_images, split)

    for counter, img_file in tqdm(enumerate(sorted(os.listdir(os.path.join(base_image_dir, dataset))))):
        image = Image.open(os.path.join(base_image_dir, dataset, img_file))

        print('dimensions: (%d,%d)'%(image.size[0], image.size[1]))
        if image.size[0]!=178 and image.size[1]!=218:
            os.remove(os.path.join(base_image_dir, dataset, img_file))
            continue

        if dataset in ['celeba', 'celebA']:
            image = center_crop(image, 9, 9, 39, 19)

        assert image.size[0]==image.size[1], 'Image size is not square, revisit the data generation code. Image dimension: (%d, %d)' % (image.size[0], image.size[1])

        try:
            assert image.size[0] == target_resolution
        except AssertionError:
            print('image.size[0] is not equal to target resolution: %d - %d' % (image.size[0], target_resolution))

        if image.size[0] > target_resolution:
            image = image.resize((target_resolution, target_resolution))

        image = torch.from_numpy(np.array(image)).float().unsqueeze(0)
        image = normalise(image, value_range=[0, 255])
        image = image.permute(0, 3, 1, 2)

        save_file = os.path.join(base_image_dir, dataset+'_'+str(target_resolution), phase_dataset[counter], img_file.split('.')[0]+'.png')
        image_grid = make_grid(image, nrow=1, normalize=False)
        save_image(tensor=image_grid, fp=save_file)

        if 0 in approx_level_ranges.keys():
            approx_level_ranges[0].append([image.min(), image.max()])
        else:
            approx_level_ranges[0] = [[image.min(), image.max()]]

        for i in range(1, levels+1):
            intermediate_resolution = target_resolution // 2**i #intermediate resolution
            haar_image = haar_transform(image)
            if i in haar_level_ranges.keys():
                haar_level_ranges[i].append([haar_image.min(), haar_image.max()])
            else:
                haar_level_ranges[i] = [[haar_image.min(), haar_image.max()]]

            permuted_haar_image = permute_channels(haar_image)
            image = permuted_haar_image[:, :3, :, :]

            if i in approx_level_ranges.keys():
                approx_level_ranges[i].append([image.min(), image.max()])
            else:
                approx_level_ranges[i] = [[image.min(), image.max()]]

            save_file = os.path.join(base_image_dir, dataset+'_'+str(intermediate_resolution), phase_dataset[counter], img_file.split('.')[0]+'.npy')
            np.save(file=save_file, arr=np.squeeze(image, axis=0))

        counter+=1


    print('----------- Haar Transform ranges ---------')
    for level in haar_level_ranges.keys():
        min_maxs = np.array(haar_level_ranges[level])
        minimum, maximum= np.mean(min_maxs[:, 0]), np.mean(min_maxs[:, 1])
        print('level: %d - min: %.3f - max: %.3f' % (level, minimum, maximum))

    print('------- Approximation coefficient ranges --------')
    for level in approx_level_ranges.keys():
        min_maxs = np.array(approx_level_ranges[level])
        minimum, maximum= np.mean(min_maxs[:, 0]), np.mean(min_maxs[:, 1])
        print('level: %d - min: %.3f - max: %.3f' % (level, minimum, maximum))

def create_dataset(config):
    base_image_dir = config.data.base_dir
    dataset = config.data.dataset
    target_resolution = config.data.target_resolution
    levels = config.data.max_haar_depth
    split = config.data.split
    create_haar_dataset(base_image_dir, dataset, target_resolution, levels, split)