-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathscenarios.py
179 lines (133 loc) · 8.02 KB
/
scenarios.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
"""
A script that partitions the dataset for transferability scenarios
"""
# basics
import numpy as np
from PIL import Image
# torch...
import torch
# custom libs
import utils
# ------------------------------------------------------------------------------
# Misc. functions
# ------------------------------------------------------------------------------
def update_numpy(acc, term, func):
if acc is None:
acc = term
else:
acc = func((acc, term))
return acc
def get_class_wise_lists(n_classes_cifar10, return_test=False):
if not return_test:
class_wise_dataset = []
for n_class in range(n_classes_cifar10):
train_data, train_labels, _, _ = af.get_cifar10_class_data(n_class) # don't use
class_wise_dataset.append((train_data, train_labels))
return class_wise_dataset
else:
class_wise_dataset = []
test_class_wise_dataset = []
for n_class in range(n_classes_cifar10):
train_data, train_labels, test_data, test_labels = af.get_cifar10_class_data(n_class) # don't use
class_wise_dataset.append((train_data, train_labels))
test_class_wise_dataset.append((test_data, test_labels))
return class_wise_dataset, test_class_wise_dataset
# ------------------------------------------------------------------------------
# Scenario related...
# ------------------------------------------------------------------------------
def scenario_1_split(int_percentages=None):
np.random.seed(0)
"""
Scenario 1) Train CIFAR10 models that use 10%, 25%, 50% of the full training set.
Chooses p% of data in each class (and corresponding labels)
Parameter int_percentages contains percentages as integers, NOT FLOATS!
Returns:
- percent_loaders (dict): each key p% contains an af.ManualData object containing p% of dataset (p% from each label)
* Loader data contains p% of images (p% of class 0, ..., p% of class 9) - consecutive
* Loader labels (np.ndarray): contains p% of labels (p% 0s, ..., p% 9s) - consecutive
"""
if int_percentages is None:
int_percentages = [10, 25, 50, 100]
print('Running scenario_1_split\n')
n_classes_cifar10 = 10
# get a list containing CIFAR10 data class by class (class k at index k)
class_wise_dataset = get_class_wise_lists(n_classes_cifar10)
percent_loaders = {}
for p in int_percentages:
subset_data = None
subset_labels = None
for n_class in range(n_classes_cifar10):
crt_train_data, crt_train_labels = class_wise_dataset[n_class]
count = crt_train_data.shape[0]
how_many_2_choose = int(count * p / 100.0)
indexes = np.random.choice(np.arange(count), how_many_2_choose, replace=False)
subset_data = update_numpy(acc=subset_data, term=np.copy(crt_train_data[indexes]), func=np.vstack)
subset_labels = update_numpy(acc=subset_labels, term=np.copy(crt_train_labels[indexes]), func=np.hstack)
# end for n_class
print(f'p={p}, data: {subset_data.shape}, labels: {subset_labels.shape}\n')
percent_loaders[p] = af.ManualData(data=subset_data, labels=subset_labels)
# end for p
np.random.seed(af.get_random_seed())
return percent_loaders
def scenario_2_split(int_classes=None):
np.random.seed(0)
"""
Scenario 2) Split CIFAR10 training set into non-overlapping 5 classes - 5 classes, 6 - 6 and 7 - 7.
Parameter int_classes_left:
- each value c is used to generate the two datasets that contain c classes
Returns:
- percent_loaders (dict): each key c contains a pair of af.ManualData meaning ( Dataset w c classes, another dataset c classes)
* Loader data contains p% of images (p% of class 0, ..., p% of class 9) - consecutive
* Loader labels (np.ndarray): contains p% of labels (p% 0s, ..., p% 9s) - consecutive
"""
if int_classes is None:
int_classes = [5, 6, 7]
print('Running scenario_2_split\n')
n_classes_cifar10 = 10
# get a list containing CIFAR10 data class by class (class k at index k)
class_wise_dataset, test_class_wise_dataset = get_class_wise_lists(n_classes_cifar10, return_test=True)
all_classes = np.arange(n_classes_cifar10)
class_loaders = {}
for classes in int_classes:
num_class_overlap = 2*(classes - 5)
class_indexes_overlap = np.random.choice(all_classes, num_class_overlap, replace=False)
left_unique_classes = np.random.choice([x for x in all_classes if x not in class_indexes_overlap], classes-num_class_overlap, replace=False)
right_unique_classes = [x for x in all_classes if (x not in class_indexes_overlap) and (x not in left_unique_classes)]
class_indexes_left = np.array(list(left_unique_classes) + list(class_indexes_overlap))
class_indexes_right = np.array(list(right_unique_classes) + list(class_indexes_overlap))
print(class_indexes_left)
print(class_indexes_right)
subset_data_left, subset_labels_left = None, None
subset_data_right, subset_labels_right = None, None
subset_test_data_left, subset_test_labels_left = None, None
subset_test_data_right, subset_test_labels_right = None, None
label_left = 0
label_right = 0
for n_class in all_classes:
crt_train_data, crt_train_labels = class_wise_dataset[n_class]
crt_test_data, crt_test_labels = test_class_wise_dataset[n_class]
if n_class in class_indexes_left:
new_train_labels = np.ones(crt_train_labels.shape) * label_left # we have to relabel the dataset because pytorch expects labels as 0,1,2,3,...
subset_data_left = update_numpy(acc=subset_data_left, term=np.copy(crt_train_data), func=np.vstack)
subset_labels_left = update_numpy(acc=subset_labels_left, term=np.copy(new_train_labels), func=np.hstack)
new_test_labels = np.ones(crt_test_labels.shape) * label_left
subset_test_data_left = update_numpy(acc=subset_test_data_left, term=np.copy(crt_test_data), func=np.vstack)
subset_test_labels_left = update_numpy(acc=subset_test_labels_left, term=np.copy(new_test_labels), func=np.hstack)
label_left += 1
if n_class in class_indexes_right:
new_train_labels = np.ones(crt_train_labels.shape) * label_right # we have to relabel the dataset because pytorch expects labels as 0,1,2,3,...
subset_data_right = update_numpy(acc=subset_data_right, term=np.copy(crt_train_data), func=np.vstack)
subset_labels_right = update_numpy(acc=subset_labels_right, term=np.copy(new_train_labels), func=np.hstack)
new_test_labels = np.ones(crt_test_labels.shape) * label_right
subset_test_data_right = update_numpy(acc=subset_test_data_right, term=np.copy(crt_test_data), func=np.vstack)
subset_test_labels_right = update_numpy(acc=subset_test_labels_right, term=np.copy(new_test_labels), func=np.hstack)
label_right += 1
# end for n_class
print(f'{classes}: train - data-left: {subset_data_left.shape}, labels-left: {subset_labels_left.shape}, data-right: {subset_data_right.shape}, labels-right: {subset_labels_right.shape}\n')
print(f'{classes}: test - data-left: {subset_test_data_left.shape}, labels-left: {subset_test_labels_left.shape}, data-right: {subset_test_data_right.shape}, labels-right: {subset_test_labels_right.shape}\n')
loaders_left = (af.ManualData(data=subset_data_left, labels=subset_labels_left), af.ManualData(data=subset_test_data_left, labels=subset_test_labels_left))
loaders_right = (af.ManualData(data=subset_data_right, labels=subset_labels_right), af.ManualData(data=subset_test_data_right, labels=subset_test_labels_right))
class_loaders[classes] = (loaders_left, loaders_right)
np.random.seed(af.get_random_seed())
# end for class_left, class_right
return class_loaders