-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain_cifar.py
289 lines (228 loc) · 10.2 KB
/
train_cifar.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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
# -*- coding: utf-8 -*-
"""
Created on Tue Aug 10 07:52:53 2021
@author: JeanMichelAmath
"""
from architectures.resnet import CifarResNet, BasicBlock
import augmix.augmentations
import torch
from torch.utils.data import DataLoader
import torch.nn.functional as F
from torchvision import datasets
from torchvision import transforms
import numpy as np
import torch.nn as nn
from isda.isda import ISDALoss, ISDALossFull, ISDALossPosNeg
from isda.isda_utils import train_isda
import matplotlib.pyplot as plt
import os
from pathlib import Path
import torchvision
from isda.isda_utils import Full_layer
from architectures.lenet import EnsembleNeuralNet
from training_loop_utils import train_model
def get_device():
if torch.cuda.is_available():
device = 'cuda:0'
else:
device = 'cpu'
return device
device = get_device()
CORRUPTIONS = [
'gaussian_noise', 'shot_noise', 'impulse_noise', 'defocus_blur',
'glass_blur', 'motion_blur', 'zoom_blur', 'snow', 'frost', 'fog',
'brightness', 'contrast', 'elastic_transform', 'pixelate',
'jpeg_compression'
]
from augmix.augmix_utils import test_c, AugMixDataset, train_augmix, test
train_transform = transforms.Compose(
[transforms.RandomHorizontalFlip(),
transforms.RandomCrop(32, padding=4)])
preprocess = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize([0.5] * 3, [0.5] * 3)])
test_transform = preprocess
n_epochs = 15
######################## TRAIN AUGMIX ########################
# Load datasets
train_data = datasets.CIFAR10('../data/cifar', train=True, transform=train_transform, download=False)
test_data = datasets.CIFAR10('../data/cifar', train=False, transform=test_transform, download=False)
# the specifics of AugMix happens here, in the custom dataset but also later in the train_augmix function
# where the mixture is computed
train_data = AugMixDataset(train_data, preprocess, no_jsd=False)
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=128,
shuffle=True)
test_loader = torch.utils.data.DataLoader(
test_data,
batch_size=128,
shuffle=True)
base_c_path = '../data/CIFAR-10-C/'
print('TRAIN AUGMIX')
model = CifarResNet(BasicBlock, [1,1,1]).to(device)
# model.fc = nn.Sequential()
lr_steps = n_epochs * len(train_loader)
optimizer = torch.optim.SGD(model.parameters(), lr=0.2, momentum=0.9, weight_decay=5e-4)
scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=0, max_lr=0.2,
step_size_up=lr_steps / 2, step_size_down=lr_steps / 2)
augmix_loss = []
for epoch in range(n_epochs):
# begin_time = time.time()
train_loss_ema = train_augmix(model, train_loader, optimizer, scheduler, no_jsd=False)
test_loss, test_acc = test(model, test_loader)
augmix_loss.append(train_loss_ema)
print('Epoch {} - AugMix loss: {} - Test loss: {} - Test Acc: {}'.format(epoch, train_loss_ema, test_loss, test_acc))
test_acc_augmix = test(model, test_loader)
test_c_acc_augmix = test_c(model, test_data, base_c_path)
######################## TRAIN ISDA, ISDA FULL, ISDA POSNEG, CROSS-ENTROPY ########################
train_data = datasets.CIFAR10('../data/cifar', train=True, transform=preprocess, download=False)
test_data = datasets.CIFAR10('../data/cifar', train=False, transform=test_transform, download=False)
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=128,
shuffle=True)
test_loader = torch.utils.data.DataLoader(
test_data,
batch_size=128,
shuffle=True)
def evaluate_loss_function(criterion, device, epochs, train_loader, test_loader, test_data, CE=False):
model = CifarResNet(BasicBlock, [1,1,1]).to(device)
feature_nb = model.fc.in_features
fc = Full_layer(feature_nb, class_num=10).to(device)
# we empty the last layer with a sequential layer doing nothing (pass-through)
model.fc = nn.Sequential()
lr_steps = epochs * len(train_loader)
optimizer = torch.optim.SGD([{'params': model.parameters()},
{'params': fc.parameters()}],
lr=0.2,
momentum=0.9,
weight_decay=5e-4)
scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=0, max_lr=0.2,
step_size_up=lr_steps / 2, step_size_down=lr_steps / 2)
loss_isda = []
acc_isda = []
for epoch in range(0, epochs):
# train for one epoch
if CE:
loss, acc = train_isda(train_loader, model, fc, criterion, optimizer, epoch, device, scheduler, loss_isda=False)
else:
loss, acc = train_isda(train_loader, model, fc, criterion, optimizer, epoch, device, scheduler)
loss_isda.append(loss.ave)
acc_isda.append(acc.ave)
model.fc = fc
iid_acc = test(model, test_loader)
ood_acc = test_c(model, test_data, base_c_path)
return iid_acc, ood_acc
ce_criterion = nn.CrossEntropyLoss()
print('CE Loss')
ce_acc, ce_ood_acc = evaluate_loss_function(ce_criterion, device, n_epochs, train_loader, test_loader, test_data, CE=True)
isda_loss = ISDALoss(64, class_num=10)
print('ISDA Loss')
isda_acc, isda_ood_acc = evaluate_loss_function(isda_loss, device, n_epochs, train_loader, test_loader, test_data)
isda_loss_full = ISDALossFull(64, class_num=10, rank=[0,1,2,3], use_mu=True)
print('ISDA Full Rank 1 - 3')
isda_full_acc, isda_full_ood_acc = evaluate_loss_function(isda_loss_full, device, n_epochs, train_loader, test_loader, test_data)
isda_loss_pn = ISDALossPosNeg(64, class_num=10, rank=[0,1,2,3])
print('ISDA Pos Neg Rank 1-3')
isda_loss_pn_acc, isda_loss_pn_ood_acc = evaluate_loss_function(isda_loss_pn, device, n_epochs, train_loader, test_loader, test_data)
######################## TRAIN ########################
train_data = datasets.CIFAR10('../data/cifar', train=True, transform=preprocess, download=False)
test_data = datasets.CIFAR10('../data/cifar', train=False, transform=test_transform, download=False)
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=128,
shuffle=True)
test_loader = torch.utils.data.DataLoader(
test_data,
batch_size=128,
shuffle=True)
print('TRAIN DEEP ENSEMBLES')
## Train an ensemble of NN
def train_ensemble(N, n_epochs, trainLoader, no_classes):
# Here our goal is to train a neural net N times from scratch, and return the list of trained neural net
ensembles = []
for i in range(N):
model = CifarResNet(BasicBlock, [1,1,1]).to(device)
lr_steps = n_epochs * len(train_loader)
optimizer = torch.optim.SGD(model.parameters(), lr=0.2, momentum=0.9, weight_decay=5e-4)
scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=0, max_lr=0.2,
step_size_up=lr_steps / 2, step_size_down=lr_steps / 2)
crit = nn.CrossEntropyLoss()
_ = train_model(model, trainLoader, n_epochs, crit, optimizer, no_classes, device, scheduler)
ensembles.append(model)
return ensembles
ensembles = train_ensemble(5, n_epochs, train_loader, no_classes=10)
ensemble_nn = EnsembleNeuralNet(ensembles)
ens_acc = test(ensemble_nn, test_loader)
ens_ood_acc = test_c(ensemble_nn, test_data, base_c_path)
######################## TRAIN FAST GRADIENT SIGN METHOD ########################
print('TRAIN FAST GRADIENT SIGN METHOD')
model = CifarResNet(BasicBlock, [1,1,1]).to(device)
lr_steps = n_epochs * len(train_loader)
optimizer = torch.optim.SGD(model.parameters(), lr=0.2, momentum=0.9, weight_decay=5e-4)
scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=0, max_lr=0.2,
step_size_up=lr_steps / 2, step_size_down=lr_steps / 2)
crit = nn.CrossEntropyLoss()
_ = train_model(model, train_loader, n_epochs, crit, optimizer, 10, device, scheduler, fgsm=True)
fgsm_acc = test(model, test_loader)
fgsm_ood_acc = test_c(model, test_data, base_c_path)
######################## TRAIN MIXUP ########################
print('TRAIN MIXUP')
model = CifarResNet(BasicBlock, [1,1,1]).to(device)
lr_steps = n_epochs * len(train_loader)
optimizer = torch.optim.SGD(model.parameters(), lr=0.2, momentum=0.9, weight_decay=5e-4)
scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=0, max_lr=0.2,
step_size_up=lr_steps / 2, step_size_down=lr_steps / 2)
def mixup_log_loss(prediction, label):
loss = nn.LogSoftmax()
log_loss = loss(prediction) * label
return -log_loss.mean()
crit = mixup_log_loss
_ = train_model(model, train_loader, n_epochs, crit, optimizer, 10, device, scheduler, mixup=True)
mixup_acc = test(model, test_loader)
mixup_ood_acc = test_c(model, test_data, base_c_path)
######################## TRAIN SYNTHETIC AUGMENTATION ########################
print('TRAIN SYNTHETIC AUGMENTATION')
transf = transforms.ToTensor()# Turn PIL Image to torch.Tensor
current_dir = os.getcwd()
style_set_path = Path(os.path.join(current_dir,'adain/style/style_set/'))
style_dataset = torchvision.datasets.ImageFolder(style_set_path, transform=transf)
style_loader = DataLoader(style_dataset, batch_size=128, shuffle=True,)
model = CifarResNet(BasicBlock, [1,1,1]).to(device)
lr_steps = n_epochs * len(train_loader)
optimizer = torch.optim.SGD(model.parameters(), lr=0.2, momentum=0.9, weight_decay=5e-4)
scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, base_lr=0, max_lr=0.2,
step_size_up=lr_steps / 2, step_size_down=lr_steps / 2)
crit = nn.CrossEntropyLoss()
_ = train_model(model, train_loader, n_epochs, crit, optimizer, 10, device, scheduler, style_loader=style_loader)
adain_acc = test(model, test_loader)
adain_ood_acc = test_c(model, test_data, base_c_path)
######################## EVALUATE RESULTS ########################
print('-------------------')
print('Cifar-10-Acc')
print('\n')
print('CE :', ce_acc[1])
print('Deep Ensembles :', ens_acc[1])
print('FGSM :', fgsm_acc[1])
print('AugMix: ', test_acc_augmix[1])
print('Mixup: ', mixup_acc[1])
print('Style Transfer: ', adain_acc[1])
print('ISDA : ', isda_acc[1])
print('ISDA_Full : ', isda_full_acc[1])
print('ISDA_PosNeg : ', isda_loss_pn_acc[1])
print('\n')
print('-------------------')
print('Cifar-10-C Acc')
print('\n')
print('CE :', ce_ood_acc)
print('Deep Ensembles :', ens_ood_acc)
print('FGSM :', fgsm_ood_acc)
print('AugMix: ', test_c_acc_augmix)
print('Mixup: ', mixup_ood_acc)
print('Style Transfer: ', adain_ood_acc)
print('ISDA : ', isda_ood_acc)
print('ISDA_Full : ', isda_full_ood_acc)
print('ISDA_PosNeg : ', isda_loss_pn_ood_acc)
print('\n')
print('-------------------')