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
|
import os
import pandas as pd
from pathlib import Path
import PIL
import pytorch_lightning as pl
from PIL import Image
from torch.utils.data import Dataset, DataLoader, random_split
from torchvision import transforms
class CSVDataModule(pl.LightningDataModule):
def __init__(self,
batch_size,
data_root,
tokenizer,
instance_prompt,
class_data_root=None,
class_prompt=None,
size=512,
repeats=100,
interpolation="bicubic",
identifier="*",
center_crop=False,
collate_fn=None):
super().__init__()
self.data_root = data_root
self.tokenizer = tokenizer
self.instance_prompt = instance_prompt
self.class_data_root = class_data_root
self.class_prompt = class_prompt
self.size = size
self.repeats = repeats
self.identifier = identifier
self.center_crop = center_crop
self.interpolation = interpolation
self.collate_fn = collate_fn
self.batch_size = batch_size
def prepare_data(self):
metadata = pd.read_csv(f'{self.data_root}/list.csv')
image_paths = [os.path.join(self.data_root, f_path) for f_path in metadata['image'].values]
captions = [caption for caption in metadata['caption'].values]
skips = [skip for skip in metadata['skip'].values]
self.data_full = [(img, cap) for img, cap, skip in zip(image_paths, captions, skips) if skip != "x"]
def setup(self, stage=None):
train_set_size = int(len(self.data_full) * 0.8)
valid_set_size = len(self.data_full) - train_set_size
self.data_train, self.data_val = random_split(self.data_full, [train_set_size, valid_set_size])
train_dataset = CSVDataset(self.data_train, self.tokenizer, instance_prompt=self.instance_prompt,
class_data_root=self.class_data_root,
class_prompt=self.class_prompt, size=self.size, repeats=self.repeats,
interpolation=self.interpolation, identifier=self.identifier,
center_crop=self.center_crop)
val_dataset = CSVDataset(self.data_val, self.tokenizer, instance_prompt=self.instance_prompt,
class_data_root=self.class_data_root,
class_prompt=self.class_prompt, size=self.size, interpolation=self.interpolation,
identifier=self.identifier, center_crop=self.center_crop)
self.train_dataloader_ = DataLoader(train_dataset, batch_size=self.batch_size,
shuffle=True, collate_fn=self.collate_fn)
self.val_dataloader_ = DataLoader(val_dataset, batch_size=self.batch_size, collate_fn=self.collate_fn)
def train_dataloader(self):
return self.train_dataloader_
def val_dataloader(self):
return self.val_dataloader_
class CSVDataset(Dataset):
def __init__(self,
data,
tokenizer,
instance_prompt,
class_data_root=None,
class_prompt=None,
size=512,
repeats=1,
interpolation="bicubic",
identifier="*",
center_crop=False,
):
self.data = data
self.tokenizer = tokenizer
self.instance_prompt = instance_prompt
self.num_instance_images = len(self.data)
self._length = self.num_instance_images * repeats
self.identifier = identifier
if class_data_root is not None:
self.class_data_root = Path(class_data_root)
self.class_data_root.mkdir(parents=True, exist_ok=True)
self.class_images = list(Path(class_data_root).iterdir())
self.num_class_images = len(self.class_images)
self._length = max(self.num_class_images, self.num_instance_images)
self.class_prompt = class_prompt
else:
self.class_data_root = None
self.interpolation = {"linear": PIL.Image.LINEAR,
"bilinear": PIL.Image.BILINEAR,
"bicubic": PIL.Image.BICUBIC,
"lanczos": PIL.Image.LANCZOS,
}[interpolation]
self.image_transforms = transforms.Compose(
[
transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5]),
]
)
self.cache = {}
def __len__(self):
return self._length
def get_example(self, i):
image_path, text = self.data[i % self.num_instance_images]
if image_path in self.cache:
return self.cache[image_path]
example = {}
instance_image = Image.open(image_path)
if not instance_image.mode == "RGB":
instance_image = instance_image.convert("RGB")
text = text.format(self.identifier)
example["prompts"] = text
example["instance_images"] = instance_image
example["instance_prompt_ids"] = self.tokenizer(
self.instance_prompt,
padding="do_not_pad",
truncation=True,
max_length=self.tokenizer.model_max_length,
).input_ids
if self.class_data_root:
class_image = Image.open(self.class_images[i % self.num_class_images])
if not class_image.mode == "RGB":
class_image = class_image.convert("RGB")
example["class_images"] = class_image
example["class_prompt_ids"] = self.tokenizer(
self.class_prompt,
padding="do_not_pad",
truncation=True,
max_length=self.tokenizer.model_max_length,
).input_ids
self.cache[image_path] = example
return example
def __getitem__(self, i):
example = {}
unprocessed_example = self.get_example(i)
example["prompts"] = unprocessed_example["prompts"]
example["instance_images"] = self.image_transforms(unprocessed_example["instance_images"])
example["instance_prompt_ids"] = unprocessed_example["instance_prompt_ids"]
if self.class_data_root:
example["class_images"] = self.image_transforms(unprocessed_example["class_images"])
example["class_prompt_ids"] = unprocessed_example["class_prompt_ids"]
return example
|