-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathsis_training_util.py
220 lines (181 loc) · 8.06 KB
/
sis_training_util.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
218
219
220
"""Util for training on SIS/backselect subsets."""
import glob
import numpy as np
import os
import pandas as pd
import torch
from tqdm import tqdm
import sis_analysis_util
import sis_util
from sufficient_input_subsets import sis
def load_preds_file(preds_filepath):
preds_df = pd.read_csv(
preds_filepath,
header=None,
names=['dataset_idx', 'pred_class_idx', 'confidence'],
)
return preds_df
def create_random_mask(shape, pixels_to_keep):
"""Create random mask of pixels.
If masking all channels per pixel, use `shape` (1, 32, 32).
"""
mask = np.zeros(shape, dtype=bool)
# Set pixels_to_keep pixels of mask to True (uniformly at random).
mask_flat = np.ravel(mask, order='C')
true_idxs = np.random.choice(
mask_flat.shape[0], replace=False, size=pixels_to_keep)
mask_flat[true_idxs] = True
mask = mask_flat.reshape(shape, order='C')
assert mask.sum() == pixels_to_keep
return mask
class BackselectSubsetsCIFARDataset(torch.utils.data.Dataset):
"""Dataset of backselect pixel subsets on CIFAR.
Args:
sis_dir (str): Path to directory containing SIS.
original_dataset: Dataset object containing original images.
frac_to_keep (float): Fraction of each images to retain (rest masked).
fully_masked_image (array): Array containing fully masked images.
Shape should be either (3, 32, 32) or (1, 32, 32) (or broadcastable
to images).
true_labels (bool): If True, use true labels as targets for each image
(rather than predicted labels). If False, must specify
`preds_filepath`. Default is True.
preds_filepath (str): Path to file containing predictions on original
images. See `inference.py` on format details and to generate preds
file for saved models. Default is None.
transform (function): Transform to apply to final images.
target_transform (function): Transform to apply to targets.
"""
def __init__(self, sis_dir, original_dataset, frac_to_keep,
fully_masked_image, true_labels=True, preds_filepath=None,
transform=None, target_transform=None):
self.sis_dir = sis_dir
self.original_dataset = original_dataset
self.frac_to_keep = frac_to_keep
self.fully_masked_image = fully_masked_image
self.pixels_to_keep = int(np.ceil(1024 * self.frac_to_keep))
self.true_labels = true_labels
self.preds_filepath = preds_filepath
self.transform = transform
self.target_transform = target_transform
self.masked_images = None
self.targets = None
if not (preds_filepath or true_labels):
raise ValueError(
'If not specifying `preds_filepath`, must use `true_labels`.')
if preds_filepath and true_labels:
raise ValueError(
'Must specify only one of `preds_filepath` and `true_labels`.')
# Load labels from predictions or true labels.
self._load_labels()
# Load masks from SIS data.
self._initialize_masked_images()
assert self.masked_images.shape[0] == len(self.original_dataset)
assert self.masked_images.shape[0] == self.targets.shape[0]
def _load_labels(self):
if self.true_labels:
labels = np.array(self.original_dataset.targets)
else:
preds_df = load_preds_file(self.preds_filepath)
labels = preds_df['pred_class_idx'].values
self.targets = labels
def _initialize_masked_images(self):
masked_images = []
for i in tqdm(range(len(self.original_dataset))):
image, _ = self.original_dataset[i]
image = image.cpu().numpy()
matched_files = glob.glob(
os.path.join(self.sis_dir, '*_%d.npz' % i))
assert len(matched_files) == 1
sr = sis_util.load_sis_result(matched_files[0])
bs_mask = sis_analysis_util.backselect_mask_from_sis_result(
sr, self.pixels_to_keep)
img_masked = sis.produce_masked_inputs(
image, self.fully_masked_image, [bs_mask])[0]
masked_images.append(img_masked)
masked_images = np.array(masked_images)
self.masked_images = masked_images
def __len__(self):
return len(self.original_dataset)
def __getitem__(self, index):
"""
Args:
index (int): Index
Returns:
tuple: (image, target) where target is index of the target class.
"""
img = self.masked_images[index]
img = torch.from_numpy(img)
target = self.targets[index]
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target
class RandomSubsetsCIFARDataset(torch.utils.data.Dataset):
"""Dataset of random pixel subsets on CIFAR.
Random subsets are retained in the dataset so that they remain the same
each time an image is called. Can be seeded so that subsets are the same
across initializations of RandomSubsetsCIFARDataset.
Args:
original_dataset: Dataset object containing original images.
frac_to_keep (float): Fraction of each images to retain (rest masked).
fully_masked_image (array): Array containing fully masked images.
Shape should be either (3, 32, 32) or (1, 32, 32) (or broadcastable
to images).
random_seed (int): Random seed. Default is None (no seed).
transform (function): Transform to apply to final images.
target_transform (function): Transform to apply to targets.
"""
# Random seed constants for train/test splits. Different seeds so the
# random subsets are different for the splits.
# Using the seeding in training scripts enables different models to
# train on the same random subsets.
RANDOM_SEED_TRAIN = 1234
RANDOM_SEED_TEST = 5678
def __init__(self, original_dataset, frac_to_keep, fully_masked_image,
random_seed=None, transform=None, target_transform=None):
self.original_dataset = original_dataset
self.frac_to_keep = frac_to_keep
self.fully_masked_image = fully_masked_image
self.random_seed = random_seed
self.pixels_to_keep = int(np.ceil(1024 * self.frac_to_keep))
self.transform = transform
self.target_transform = target_transform
self.masked_images = None
if random_seed is not None:
np.random.seed(random_seed)
# Use same labels as original dataset.
self.targets = np.array(self.original_dataset.targets)
# Create random masked images. Store the masked images so that random
# subsets remain the same each time an image is called.
self._initialize_masked_images()
assert self.masked_images.shape[0] == len(self.original_dataset)
assert self.masked_images.shape[0] == self.targets.shape[0]
def _initialize_masked_images(self):
masked_images = []
for i in tqdm(range(len(self.original_dataset))):
image, _ = self.original_dataset[i]
image = image.cpu().numpy()
random_mask = create_random_mask((1, 32, 32), self.pixels_to_keep)
img_masked = np.where(random_mask, image, self.fully_masked_image)
masked_images.append(img_masked)
masked_images = np.array(masked_images)
self.masked_images = masked_images
def __len__(self):
return len(self.original_dataset)
def __getitem__(self, index):
"""
Args:
index (int): Index
Returns:
tuple: (image, target) where target is index of the target class.
"""
img = self.masked_images[index]
img = torch.from_numpy(img)
target = self.targets[index]
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target