forked from lllyasviel/ControlNet
/
tutorial_dataset.py
38 lines (29 loc) · 1.14 KB
/
tutorial_dataset.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
import json
import cv2
import numpy as np
from torch.utils.data import Dataset
class MyDataset(Dataset):
def __init__(self):
self.data = []
with open('./training/record.json', 'rt') as f:
for line in f:
self.data.append(json.loads(line))
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
item = self.data[idx]
filenames = [item['source'], item['source1'], item['target']]
prompt = item['prompt']
sources = []
for i in range(2):
source = cv2.imread('./training/' + filenames[i])
# OpenCV reads images in BGR order.
source = cv2.cvtColor(source, cv2.COLOR_BGR2RGB)
# Normalize source images to [0, 1].
source = source.astype(np.float32) / 255.0
sources.append(source)
target = cv2.imread('./training/' + filenames[2])
target = cv2.cvtColor(target, cv2.COLOR_BGR2RGB)
# Normalize target images to [-1, 1].
target = (target.astype(np.float32) / 127.5) - 1.0
return dict(jpg=target, txt=prompt, hint1=sources[0], hint2=sources[1])