-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset.py
217 lines (169 loc) · 8.52 KB
/
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
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
213
214
215
216
217
import torch
import numpy as np
from tqdm import tqdm
from collections import defaultdict
class SignalGameDataset(torch.utils.data.Dataset):
def __init__(self, dataset, num_imgs, vision, task, classes=None, same_class_prob=0.0, seed=1):
self.dataset = dataset
self.num_imgs = num_imgs
self.vision = vision
self.task = task
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.same_class_prob = same_class_prob
np.random.seed(seed=seed)
self.img_features, self.img_class_labels = self._extract_img_features() # (dataset_size, embed_size)
self._create_class_index()
def __len__(self):
return len(self.dataset)
def randint_exclude(self, target_idx):
"""
Get distractor indices which are different than target index.
"""
random_idxs = []
# sample distractors until you get enough
while len(random_idxs) != self.num_imgs-1:
random_float = np.random.random_sample()
if random_float < self.same_class_prob: # sample distractor with the same class as target
target_class_label = self.img_class_labels[target_idx]
# sample an index from all indices from a given class
random_idx = np.random.choice(self.class_idxs[target_class_label.item()], size=None)
else: # sample distractor with any class
random_idx = np.random.randint(low=0, high=self.__len__())
# if the distractor index is different than target, add it to the list of indices
if random_idx != target_idx:
random_idxs.append(random_idx)
return random_idxs
def __getitem__(self, item):
# get random images
random_idxs = [item] + self.randint_exclude(item)
random_imgs = self.img_features[random_idxs]
# get a random permutation of integers from 0 to num_imgs-1
permutation = torch.randperm(self.num_imgs)
# set the target image as the first random image
sender_imgs = random_imgs[0].unsqueeze(dim=0)
# permute random images for the receiver
receiver_imgs = random_imgs[permutation]
# set the label
sg_target = permutation.argmin()
if self.task == "standard":
target = sg_target
elif self.task=="img_clas":
random_img_labels = self.img_class_labels[random_idxs]
# image class label of the target image
target_class = random_img_labels[0].squeeze()
# array of booleans that are true if the class of a given image is the same as the class of the target image
is_target_class = (random_img_labels[permutation] == target_class).float().squeeze()
# concatenate the signaling game label with the multilabel binary image classification label
target = torch.cat((sg_target.unsqueeze(dim=0), is_target_class))
elif self.task=="target_clas":
random_img_labels = self.img_class_labels[random_idxs]
# image class label of the target image
target_class = random_img_labels[0]
# concatenate the signaling game label with the multiclass target image classification label
target = torch.cat((sg_target.unsqueeze(dim=0), target_class))
else:
assert False, "Wrong task name"
return sender_imgs, target, receiver_imgs
def _extract_img_features(self):
"""
We have to have to extract image features by making a forward pass through the pretrained vision model.
We can't do it for all images at once as there's too many of them and we would run out of memory,
so we do it in batches.
"""
# read images and labels from dataset into separate arrays
imgs, img_class_labels = zip(*self.dataset)
imgs = torch.stack(list(imgs)) # (dataset_size, 3, 32, 32)
img_class_labels = torch.Tensor(img_class_labels).reshape(-1, 1) # (dataset_size, 1)
# if you run out of memory or your laptop freezes, decrease this number and try again
VISION_BATCH_SIZE = 1000
vision_loader = torch.utils.data.DataLoader(imgs, shuffle=False, batch_size=VISION_BATCH_SIZE, num_workers=0)
# extract features from images with a vision model
img_features = []
for img in tqdm(vision_loader, desc="Extracting features in batches"):
img = img.to(self.device)
with torch.no_grad():
img_features.append(self.vision(img))
img_features = torch.cat(img_features) # (dataset_size, 64)
return img_features, img_class_labels
def _create_class_index(self):
self.class_idxs = defaultdict(list)
for i, label in enumerate(self.img_class_labels):
self.class_idxs[label.item()].append(i)
class GaussianNoiseDataset(torch.utils.data.Dataset):
def __init__(self, num_imgs, vision, task, dataset_size, classes=None, seed=1, im_size=(3, 32, 32), mean=0.0, std_dev=1.0):
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.mean = mean
self.std_dev = std_dev
self.num_imgs = num_imgs
self.im_size = im_size
self.dataset_size = dataset_size
self.task = task
self.dataset = self._init_dataset()
self.vision = vision
np.random.seed(seed=seed)
self.img_features = self._extract_img_features() # (dataset_size, embed_size)
def __len__(self):
return len(self.dataset)
def randint_exclude(self, target_idx):
"""
Get distractor indices which are different than target index.
"""
random_idxs = []
# sample distractors until you get enough
while len(random_idxs) != self.num_imgs-1:
random_float = np.random.random_sample()
random_idx = np.random.randint(low=0, high=self.__len__())
# if the distractor index is different than target, add it to the list of indices
if random_idx != target_idx:
random_idxs.append(random_idx)
return random_idxs
def __getitem__(self, item):
# get random images
random_idxs = [item] + self.randint_exclude(item)
random_imgs = self.img_features[random_idxs]
# get a random permutation of integers from 0 to num_imgs-1
permutation = torch.randperm(self.num_imgs)
# set the target image as the first random image
sender_imgs = random_imgs[0].unsqueeze(dim=0)
# permute random images for the receiver
receiver_imgs = random_imgs[permutation]
# set the label
sg_target = permutation.argmin()
if self.task == "standard":
target = sg_target
elif self.task=="img_clas":
is_target_class = torch.zeros(2)
target = torch.cat((sg_target.unsqueeze(dim=0), is_target_class))
elif self.task=="target_clas":
target_class = torch.zeros(1).float()
target = torch.cat((sg_target.unsqueeze(dim=0), target_class))
else:
assert False, "Wrong task name"
return sender_imgs, target, receiver_imgs
def _init_dataset(self):
samples = []
for i in range(self.dataset_size):
noise_vec = np.random.normal(loc=self.mean, scale=self.std_dev, size=self.im_size)
# What label to give to these noise vectors?
samples.append((torch.from_numpy(noise_vec).float(), 0))
return samples
def _extract_img_features(self):
"""
We have to extract image features by making a forward pass through the pretrained vision model.
We can't do it for all images at once as there's too many of them and we would run out of memory,
so we do it in batches.
"""
# read images from dataset into a single array
imgs = [img for img, label in self.dataset]
imgs = torch.stack(imgs)
# if you run out of memory or your laptop freezes, decrease this number and try again
VISION_BATCH_SIZE = 1000
vision_loader = torch.utils.data.DataLoader(imgs, shuffle=False, batch_size=VISION_BATCH_SIZE, num_workers=0)
# extract features from images with a vision model
img_features = []
for img in tqdm(vision_loader, desc="Extracting features in batches"):
img = img.to(self.device)
with torch.no_grad():
img_features.append(self.vision(img))
img_features = torch.cat(img_features)
return img_features