-
Notifications
You must be signed in to change notification settings - Fork 13
/
data.py
executable file
·212 lines (156 loc) · 6.38 KB
/
data.py
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
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
from typing import Callable, Dict
import torch
from torch.utils.data import Dataset
import torchvision.transforms.functional as F
from torchvision import transforms
import pytorch_lightning as pl
from collections.abc import Iterable
# image reader writer
from pathlib import Path
from PIL import Image
from typing import Tuple
def read_image(filepath: Path, mode: str = None) -> Image:
with open(filepath, 'rb') as file:
image = Image.open(file)
return image.convert(mode)
image2tensor = transforms.ToTensor()
tensor2image = transforms.ToPILImage()
def write_image(image: Image, filepath: Path):
filepath.parent.mkdir(parents=True, exist_ok=True)
image.save(str(filepath))
def read_image_tensor(filepath: Path, mode: str = 'RGB') -> torch.Tensor:
return image2tensor(read_image(filepath, mode))
def write_image_tensor(input: torch.Tensor, filepath: Path):
write_image(tensor2image(input), filepath)
def get_valid_indices(H: int, W: int, patch_size: int, random_overlap: int = 0):
vih = torch.arange(random_overlap, H-patch_size -
random_overlap+1, patch_size)
viw = torch.arange(random_overlap, W-patch_size -
random_overlap+1, patch_size)
if random_overlap > 0:
rih = torch.randint_like(vih, -random_overlap, random_overlap)
riw = torch.randint_like(viw, -random_overlap, random_overlap)
vih += rih
viw += riw
vi = torch.stack(torch.meshgrid(vih, viw)).view(2, -1).t()
return vi
def cut_patches(input: torch.Tensor, indices: Tuple[Tuple[int, int]], patch_size: int, padding: int = 0):
# TODO use slices to get all patches at the same time ?
patches_l = []
for n in range(len(indices)):
patch = F.crop(input, *(indices[n]-padding),
*(patch_size+padding*2,)*2)
patches_l.append(patch)
patches = torch.cat(patches_l, dim=0)
return patches
def prepare_data(data_path: Path, read_func: Callable = read_image_tensor) -> Dict:
"""
Takes a data_path of a folder which contains subfolders with input, target, etc.
lablelled by the same names.
:param data_path: Path of the folder containing data
:param read_func: function that reads data and returns a tensor
"""
data_dict = {}
subdir_names = ["target", "input", "mask"] # ,"helper"
# checks only files for which there is an target
# TODO check for images
name_ls = [file.name for file in (
data_path / "target").iterdir() if file.is_file()]
subdirs = [data_path / sdn for sdn in subdir_names]
for sd in subdirs:
if sd.is_dir():
data_ls = []
files = [sd / name for name in name_ls]
for file in files:
tensor = read_func(file)
H, W = tensor.shape[-2:]
data_ls.append(tensor)
# TODO check that all sizes match
data_dict[sd.name] = torch.stack(data_ls, dim=0)
data_dict['name'] = name_ls
data_dict['len'] = len(data_dict['name'])
data_dict['H'] = H
data_dict['W'] = W
return data_dict
# TODO an image is loaded whenever a patch is needed, this may be a bottleneck
class DataDictLoader():
def __init__(self, data_dict: Dict,
batch_size: int = 16,
max_length: int = 128,
shuffle: bool = False):
"""
"""
self.batch_size = batch_size
self.shuffle = shuffle
self.batch_size = batch_size
self.data_dict = data_dict
self.dataset_len = data_dict['len']
self.len = self.dataset_len if max_length is None else min(
self.dataset_len, max_length)
# Calculate # batches
num_batches, remainder = divmod(self.len, self.batch_size)
if remainder > 0:
num_batches += 1
self.num_batches = num_batches
def __iter__(self):
if self.shuffle:
r = torch.randperm(self.dataset_len)
self.data_dict = {k: v[r] if isinstance(
v, Iterable) else v for k, v in self.data_dict.items()}
self.i = 0
return self
def __next__(self):
if self.i >= self.len:
raise StopIteration
batch = {k: v[self.i:self.i+self.batch_size]
if isinstance(v, Iterable) else v for k, v in self.data_dict.items()}
self.i += self.batch_size
return batch
def __len__(self):
return self.num_batches
class PatchDataModule(pl.LightningDataModule):
def __init__(self, data_dict,
patch_size: int = 2**5,
batch_size: int = 2**4,
patch_num: int = 2**6):
super().__init__()
self.data_dict = data_dict
self.H, self.W = data_dict['H'], data_dict['W']
self.len = data_dict['len']
self.batch_size = batch_size
self.patch_size = patch_size
self.patch_num = patch_num
def dataloader(self, data_dict, **kwargs):
return DataDictLoader(data_dict, **kwargs)
def train_dataloader(self):
patches = self.cut_patches()
return self.dataloader(patches, batch_size=self.batch_size, shuffle=True,
max_length=self.patch_num)
def val_dataloader(self):
return self.dataloader(self.data_dict, batch_size=1)
def test_dataloader(self):
return self.dataloader(self.data_dict) # TODO batch size
def cut_patches(self):
# TODO cycle once
patch_indices = get_valid_indices(
self.H, self.W, self.patch_size, self.patch_size//4)
dd = {k: cut_patches(
v, patch_indices, self.patch_size) for k, v in self.data_dict.items()
if isinstance(v, torch.Tensor)
}
threshold = 0.1
mask_p = torch.mean(
dd.get('mask', torch.ones_like(dd['input'])), dim=(-1, -2, -3))
masked_idx = (mask_p > threshold).nonzero(as_tuple=True)[0]
dd = {k: v[masked_idx] for k, v in dd.items()}
dd['len'] = len(masked_idx)
dd['H'], dd['W'] = (self.patch_size,)*2
return dd
class ImageDataset(Dataset):
def __init__(self, file_paths: Iterable, read_func: Callable = read_image_tensor):
self.file_paths = file_paths
def __getitem__(self, idx: int) -> dict:
file = self.file_paths[idx]
return read_image_tensor(file), file.name
def __len__(self) -> int:
return len(self.file_paths)