-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdata_loader.py
174 lines (154 loc) · 5.66 KB
/
data_loader.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
''' Digit
'''
import torch
import torch.nn.functional as F
from torch.utils.data import Dataset, TensorDataset
from torchvision import transforms
from torchvision.datasets import MNIST, USPS, SVHN, CIFAR10
import os
import pickle
import numpy as np
from scipy.io import loadmat
from PIL import Image
from tools.autoaugment import SVHNPolicy, CIFAR10Policy
from tools.randaugment import RandAugment
class myTensorDataset(Dataset):
def __init__(self, x, y, transform=None, twox=False):
self.x = x
self.y = y
self.transform = transform
self.twox = twox
def __len__(self):
return len(self.x)
def __getitem__(self, index):
x = self.x[index]
y = self.y[index]
if self.transform is not None:
x = self.transform(x)
if self.twox:
x2 = self.transform(x)
return (x, x2), y
return x, y
HOME = os.environ['HOME']
def resize_imgs(x, size):
'''
x [n, 28, 28]
size int
'''
resize_x = np.zeros([x.shape[0], size, size])
for i, im in enumerate(x):
im = Image.fromarray(im)
im = im.resize([size, size], Image.ANTIALIAS)
resize_x[i] = np.asarray(im)
return resize_x
def load_mnist(split='train', translate=None, twox=False, ntr=None, autoaug=None, channels=3):
path = f'data/digits/mnist-{split}.pkl'
if not os.path.exists(path):
dataset = MNIST(f'{HOME}/.pytorch/MNIST', train=(split=='train'), download=True)
x, y = dataset.data, dataset.targets
if split=='train':
x, y = x[0:10000], y[0:10000]
x = torch.tensor(resize_imgs(x.numpy(), 32))
x = (x.float()/255.).unsqueeze(1).repeat(1,3,1,1)
with open(path, 'wb') as f:
pickle.dump([x, y], f)
with open(path, 'rb') as f:
x, y = pickle.load(f)
if channels == 1:
x = x[:,0:1,:,:]
if ntr is not None:
x, y = x[0:ntr], y[0:ntr]
if (translate is None) and (autoaug is None):
dataset = TensorDataset(x, y)
return dataset
if translate is not None:
dataset = myTensorDataset(x, y, transform=translate, twox=twox)
return dataset
transform = [transforms.ToPILImage()]
if translate is not None:
transform.append(transforms.RandomAffine(0, [translate, translate]))
if autoaug is not None:
if autoaug == 'AA':
transform.append(SVHNPolicy())
elif autoaug == 'RA':
transform.append(RandAugment(3,4))
transform.append(transforms.ToTensor())
transform = transforms.Compose(transform)
dataset = myTensorDataset(x, y, transform=transform, twox=twox)
return dataset
def load_usps(split='train', channels=3):
path = f'data/digits/usps-{split}.pkl'
if not os.path.exists(path):
dataset = USPS(f'{HOME}/.pytorch/USPS', train=(split=='train'), download=True)
x, y = dataset.data, dataset.targets
x = torch.tensor(resize_imgs(x, 32))
x = (x.float()/255.).unsqueeze(1).repeat(1,3,1,1)
y = torch.tensor(y)
with open(path, 'wb') as f:
pickle.dump([x, y], f)
with open(path, 'rb') as f:
x, y = pickle.load(f)
if channels == 1:
x = x[:,0:1,:,:]
dataset = TensorDataset(x, y)
return dataset
def load_svhn(split='train', channels=3):
dataset = SVHN(f'{HOME}/.pytorch/SVHN', split=split, download=True)
x, y = dataset.data, dataset.labels
x = x.astype('float32')/255.
x, y = torch.tensor(x), torch.tensor(y)
if channels == 1:
x = x.mean(1, keepdim=True)
dataset = TensorDataset(x, y)
return dataset
def load_svhn_tf(split='train', channels=3, translate =None):
dataset = SVHN(f'{HOME}/.pytorch/SVHN', split=split, transform= translate, download=True)
x, y = dataset.data, dataset.labels
x = x.astype('float32')/255.
x, y = torch.tensor(x), torch.tensor(y)
if channels == 1:
x = x.mean(1, keepdim=True)
dataset = TensorDataset(x, y)
return dataset
def load_syndigit(split='train', channels=3):
path = f'data/digits/synth_{split}_32x32.mat'
data = loadmat(path)
x, y = data['X'], data['y']
x = np.transpose(x, [3, 2, 0, 1]).astype('float32')/255.
y = y.squeeze()
x, y = torch.tensor(x), torch.tensor(y)
if channels == 1:
x = x.mean(1, keepdim=True)
dataset = TensorDataset(x, y)
return dataset
def load_mnist_m(split='train', channels=3):
path = f'data/digits/mnist_m-{split}.pkl'
with open(path, 'rb') as f:
x, y = pickle.load(f)
x, y = torch.tensor(x.astype('float32')/255.), torch.tensor(y)
if channels==1:
x = x.mean(1, keepdim=True)
dataset = TensorDataset(x, y)
return dataset
def load_mnist_m_tf(split='train', channels=3):
path = f'data/digit/mnist_m-{split}.pkl'
with open(path, 'rb') as f:
x, y = pickle.load(f)
x, y = torch.tensor(x.astype('float32')/255.), torch.tensor(y)
if channels==1:
x = x.mean(1, keepdim=True)
dataset = TensorDataset(x, y)
return dataset
if __name__=='__main__':
dataset = load_mnist(split='train')
print('mnist train', len(dataset))
dataset = load_mnist('test')
print('mnist test', len(dataset))
dataset = load_mnist_m('test')
print('mnsit_m test', len(dataset))
dataset = load_svhn(split='test')
print('svhn', len(dataset))
dataset = load_usps(split='test')
print('usps', len(dataset))
dataset = load_syndigit(split='test')
print('syndigit', len(dataset))