-
Notifications
You must be signed in to change notification settings - Fork 0
/
Ensemble.py
254 lines (221 loc) · 10.2 KB
/
Ensemble.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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
from BaseModels import BaseNeuralNet
from Architectures import PreActResNet18, PreActResNet18_100, WideResNet
from torchvision import datasets, transforms
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
from WongBasedTraining import WongBasedTrainingCIFAR10
from BaseModels import Validator
import torch.cuda as cutorch
import gc
import sys
from datetime import datetime
from AdversarialAttacks import attack_pgd
from utils import (cifar10_mu, cifar10_std, cifar100_mu, cifar100_std)
class Ensemble(Validator):
def __init__(self, weak_learner_type, attack_eps, model_base, weakLearners=[], weakLearnerWeights=[]):
"""
"""
Validator.__init__(self, 'Number of weak learners', attack_eps)
self.weakLearners = weakLearners
self.weakLearnerWeights = weakLearnerWeights
self.weakLearerWeightsTensor = torch.tensor(weakLearnerWeights, requires_grad=False).unsqueeze(1).float().cuda()
self.weak_learner_type = weak_learner_type
self.accuracies['wl_train'] = []
self.accuracies['wl_val'] = []
self.attack_eps = attack_eps
self.model_base = model_base
if model_base == PreActResNet18:
self.num_classes = 10
elif model_base == PreActResNet18_100:
self.num_classes = 100
assert len(self.weakLearners) == 0
def plot_wl_acc(self, path=None):
plt.subplots()
plt.plot(self.train_checkpoints, self.accuracies['wl_train'])
plt.plot(self.train_checkpoints, self.accuracies['wl_val'])
plt.legend(['Train accuracy', 'Val accuracy'])
plt.xlabel(self.xlabel)
plt.ylabel('Weak learner accuracy')
plt.title('Weak learner accuracy')
plt.grid()
if path is not None:
plt.savefig(path, dpi=250)
plt.show()
def addWeakLearner(self, weakLearner, weakLearnerWeight):
self.weakLearners.append(weakLearner)
self.weakLearnerWeights.append(weakLearnerWeight)
self.weakLearnerWeightsTensor = torch.tensor(self.weakLearnerWeights, requires_grad=False).unsqueeze(1).float().cuda()
def getSingleWLPrediction(self, i, wLPredictions, X):
# print("memory allocated:", cutorch.memory_allocated(0), "for WL num ", i)
if not isinstance(self.weakLearners[i], str):
learner = self.weakLearners[i]
else:
learner = self.weak_learner_type(model_base=self.model_base, attack_eps=self.attack_eps)
# print("learner model:", learner.model)
# print("model base:", self.model_base)
learner.model.load_state_dict(torch.load(self.weakLearners[i]))
learner.model = learner.model.to(torch.device('cuda:0'))
learner.model.eval()
prediction = learner.model(X)
wLPredictions[:,:,i] = F.softmax(prediction, dim=1)
# wLPredictions[:,:,i] = prediction
def getWLPredictions(self, X, k):
T = len(self.weakLearners)
# print("total num weak learners:", T)
# print("Start WL Pred memory allocated:", cutorch.memory_allocated(0))
wLPredictions = torch.zeros((X.shape[0], k, T)).cuda()
for i in range(T):
self.getSingleWLPrediction(i, wLPredictions, X)
torch.cuda.empty_cache()
# print("Finishing WL pred memory allocated:", cutorch.memory_allocated(0))
return wLPredictions
def predict(self, X):
# We hope to deprecate schapirePredict since we shouldn't need anymore
# learner = self.weak_learner_type()
# learner.model.load_state_dict(torch.load(self.weakLearners[0]))
# learner.model = learner.model.to(torch.device('cuda:0'))
# learner.model.eval()
# prediction = learner.model(X)
# print("calling modded predict")
# return prediction
return self.schapireContinuousPredict(X, self.num_classes)
def predictUnnormalizedDataCIFAR10(self, X):
X_norm = (X - cifar10_mu) / cifar10_std
return self.schapireContinuousPredict(X_norm, self.num_classes)
def predictUnnormalizedDataCIFAR100(self, X):
X_norm = (X - cifar100_mu) / cifar100_std
return self.schapireContinuousPredict(X_norm, self.num_classes)
def schapirePredict(self, X, k):
print("shouldn't be here")
wLPredictions = None
predictions = np.zeros(len(X))
T = len(self.weakLearners)
wLPredictions = self.getWLPredictions(X, k).argmax(axis=1).transpose(1, 0)
for i in range(len(X)):
F_Tx =[]
for l in range(k):
F_Tx.append(sum([self.weakLearnerWeights[t] * (1 if wLPredictions[t][i] == l else 0) for t in range(T)]))
predictions[i] = np.argmax(np.array(F_Tx))
return predictions
def schapireContinuousPredict(self, X, k):
wLPredictions = None
T = len(self.weakLearners)
wLPredictions = self.getWLPredictions(X, k)
# print(wlPre)
weights = self.weakLearnerWeightsTensor
assert(wLPredictions.shape == (len(X), k, T))
# print("weights shape:", weights.shape)
# print("T:", T)
assert(weights.shape == (T, 1))
output = torch.matmul(wLPredictions, weights).squeeze(2)
del wLPredictions
assert(output.shape == (len(X), k))
output = F.normalize(output, p=1, dim=1)
output = torch.log(output)
return output
def gradOptWeights(self, train_loader, num_samples=1000, train_eps=0.127):
# gradient optimize weights
# get ~100 adv examples
# calculate cross-entropy loss
# calc gradient with respect to weights
print("weights before opt:", self.weakLearnerWeights)
print("tensor before opt:", self.weakLearnerWeightsTensor)
self.toggleWeightGrad(True)
optim = torch.optim.Adam([self.weakLearnerWeightsTensor], lr=0.01)
total_samples = 0
for _, data in enumerate(train_loader):
X, y = data[0].cuda(), data[1].cuda()
X_adv = attack_pgd(X, y, train_eps, self.predict)
if total_samples > num_samples:
break
total_samples += len(X)
output = self.predict(X_adv)
optim.zero_grad()
loss = F.cross_entropy(output, y)
loss.backward()
optim.step()
self.toggleWeightGrad(False)
self.weakLearnerWeights = self.weakLearnerWeightsTensor.squeeze(1).tolist()
print("weights after opt:", self.weakLearnerWeights)
print("tensor after opt:", self.weakLearnerWeightsTensor)
def toggleWeightGrad(self, option=True):
self.weakLearnerWeightsTensor.requires_grad = option
def get_sum(self, dicts):
ans = {}
for d in dicts:
for k in d:
if isinstance(k, str) and k == 'train' or k == 'val':
if k not in ans:
ans[k] = d[k]
else:
ans[k] += d[k]
else:
if k not in ans:
ans[k] = np.array(d[k])
else:
ans[k] += np.array(d[k])
return ans
def get_mean(self, dicts):
ans = {}
for d in dicts:
for k in d:
if isinstance(k, str) and k == 'train' or k == 'val':
if k not in ans:
ans[k] = d[k]
else:
ans[k] += d[k]
else:
if k not in ans:
ans[k] = np.array(d[k])
else:
ans[k] += np.array(d[k])
for k in ans:
ans[k] = ans[k] / len(dicts)
return ans
def record_accuracies(self, progress, train_loader, test_loader, numsamples_train, numsamples_val, val_attacks, attack_iters, restarts, dataset_name):
# record train
if numsamples_train >0:
train_batch_size = train_loader.batch_size
self.train_checkpoints.append(progress)
# sum losses
# average accuracies
curSample = 0
train_loss_dicts = []
train_acc_dicts = []
for i, data in enumerate(train_loader):
curSample += train_batch_size
if curSample >= numsamples_train: break
losses, accuracies = self.calc_accuracies(data[0].cuda(), data[1].cuda(), data_type='train', dataset_name=dataset_name)
train_loss_dicts.append(losses)
train_acc_dicts.append(accuracies)
self.losses['train'].append(self.get_sum(train_loss_dicts)['train'])
self.accuracies['train'].append(self.get_mean(train_acc_dicts)['train'])
# record val / adversarial
val_batch_size = test_loader.batch_size
self.val_checkpoints.append(progress)
curSample = 0
val_loss_dicts = []
val_acc_dicts = []
for i, data in enumerate(test_loader):
curSample += val_batch_size
if curSample >= numsamples_val: break
losses, accuracies = self.calc_accuracies(data[0].cuda(), data[1].cuda(), data_type='val', val_attacks=val_attacks, attack_iters=attack_iters, restarts=restarts, dataset_name=dataset_name)
val_loss_dicts.append(losses)
val_acc_dicts.append(accuracies)
print(accuracies)
val_loss_dict = self.get_sum(val_loss_dicts)
val_acc_dict = self.get_mean(val_acc_dicts)
self.losses['val'].append(val_loss_dict['val'])
self.accuracies['val'].append(val_acc_dict['val'])
for k in val_loss_dict:
if isinstance(k, str): continue
attack = k
if len(self.losses[attack.__name__]) == 0:
self.losses[attack.__name__] = [[] for i in range(len(self.attack_eps))]
self.accuracies[attack.__name__] = [[] for i in range(len(self.attack_eps))]
for i in range(len(self.attack_eps)):
self.losses[attack.__name__][i].append(val_loss_dict[attack][i])
self.accuracies[attack.__name__][i].append(val_acc_dict[attack][i])