-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathrun_utils.py
176 lines (137 loc) · 8.75 KB
/
run_utils.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
import torch
import torch.nn as nn
import numpy as np
import time
import torch.utils.data as data
import torch.optim as optim
from torchvision import models
from train import Trainer
import ood_utils
import layer_utils
import pdb
def weights_init_(m):
if isinstance(m, nn.Linear) or isinstance(m, nn.Conv2d):
nn.init.xavier_uniform_(m.weight, gain=1)
nn.init.constant_(m.bias, 0)
def run(trainer,args, trainloader,valloader, testloader, classes_idx_OOD, classes,classes_idx_ID, idx):
if args.load_checkpoint:
if len(classes_idx_OOD) == 1:
checkpoint = torch.load(args.save_path+'/model_{}.pt'.format(classes[classes_idx_OOD[0]]))
else:
checkpoint = torch.load(args.save_path+'/model_{}.pt'.format(idx))
trainer.model.load_state_dict(checkpoint)
else:
trainer.model.apply(weights_init_)
if args.train:
prev_loss = 1e30
prev_loss_g = 1e30
for z in range(0,1):
print("TRAINING")
if args.dataset1 == 'cifar10_old':
alexnet = models.vgg16(pretrained=True)
output_dim = args.ID_tasks
alexnet.classifier[3] = nn.Linear(4096,1024)
alexnet.classifier[6] = nn.Linear(1024, output_dim)
alexnet_dict = alexnet.state_dict()
model_dict = trainer.model.state_dict()
pretrained_dict = {k : v for k,v in alexnet_dict.items() if k in model_dict}
model_dict.update(pretrained_dict)
trainer.model.load_state_dict(model_dict)
#torch.nn.init.xavier_normal_(trainer.model.classifier[6].weight, gain =1)
#torch.nn.init.xavier_normal_(trainer.model.classifier[3].weight, gain = 1)
scheduler = optim.lr_scheduler.StepLR(trainer.optimizer, step_size = 4, gamma = 0.5)
for epoch in range(args.epochs):
start_time = time.time()
train_loss, train_acc = trainer.optimize(trainloader,classes_idx_OOD, classes_idx_ID)
end_time = time.time()
print(f'\t EPOCH: {epoch+1:.0f} | time elapsed: {end_time - start_time:.3f}')
print(f'\tTrain Loss: {train_loss:.3f} | Train Acc: {train_acc*100:.2f}%')
loss, acc, act_avg_test,_,_ = trainer.evaluate(valloader,classes_idx_OOD,classes_idx_ID, extract_act = False)
print(f'\tTest Loss: {loss:.3f} | Test Acc: {acc*100:.2f}%')
if args.dataset1 =='cifar10':
scheduler.step()
if loss < prev_loss:
prev_loss = loss
train_loss, train_acc = trainer.optimize(valloader,classes_idx_OOD, classes_idx_ID)
if len(classes_idx_OOD) == 1:
torch.save(trainer.model.state_dict(), arg.save_path+'/model_{}.pt'.format(classes[classes_idx_OOD[0]]))
else:
torch.save(trainer.model.state_dict(), args.save_path+'/model_{}.pt'.format(idx))
else:
## do testing
print("TESTING")
def do_ood_eval(trainloader,valloader, testloader,testset, ood_trainset_list,ood_testset_list, trainer, classes, classes_idx_OOD,args,classes_idx_ID, save_k = 0):
## the current implementation only works with a batch size of 1 which is probably not ideal
## 1. use traindata to compute thresholds for each class
_, _, _,activations_list_train, _ = trainer.evaluate(trainloader, classes_idx_OOD, classes_idx_ID, extract_act = True)
if len(classes_idx_OOD) == 1:
np.savez(arg.save_path+'/activations/act_full_train_{}.npz'.format(classes[classes_idx_OOD[0]]),**activations_list_train)
else:
np.savez(args.save_path+'/activations/act_full_train_{}.npz'.format(save_k),**activations_list_train)
if len(classes_idx_OOD) == 1:
activations_list_train = dict(np.load(args.save_path+'/activations/act_full_train_{}.npz'.format(classes[classes_idx_OOD[0]]), allow_pickle = True))
else:
activations_list_train = dict(np.load(args.save_path+'/activations/act_full_train_{}.npz'.format(save_k),allow_pickle = True))
## compute class wise thresholds
thresholds = ood_utils.compute_per_class_thresholds(activations_list_train, trainer, classes,classes_idx_OOD, args.ID_tasks, args.baseline_ood)
if not args.cont_learner:
## 2. with these thresholds evuate the test set accuracy
test_acc = ood_utils.compute_test_Acc(testloader, thresholds, trainer, classes, classes_idx_OOD, args.save_path, args.ID_tasks,classes_idx_ID, args.baseline_ood, save_k)
# test_loss, test_accuracy = evaluate_with_thresh(testloader, thresholds, trainer.model)
# ## 3. evaluate ood detection data ## with varying data amounts
ood_acc = ood_utils.compute_ood_Acc(ood_trainset_list, thresholds, trainer, classes, classes_idx_OOD, args.save_path, args.ID_tasks, args.multiple_dataset,classes_idx_ID,args.baseline_ood,save_k)
return test_acc, ood_acc
else:
avg_act_all_layers, layer_indices = layer_utils.return_all_layer_activations(trainer, testloader)
test_acc_full = []
ood_acc_full = []
lr_mult = 2 # set to 2 for fmnist
for k in range(0, len(ood_trainset_list),1):
curr_ood_data = ood_trainset_list[k]
# loss, acc, activations,activations_list_test, labels_list = trainer.evaluate(testloader, [classes_idx_OOD[k],0],classes_idx_ID, extract_act = False)
#
# print(f'\tTest Loss: {loss:.3f} | Test Acc: {acc*100:.2f}%')
## incremental test data ood detetion
if args.full_pipeline:
testloader = data.DataLoader(testset, batch_size = 1, shuffle = False, num_workers = 2)
for p in range(0,3):
percent = 0.1*p
test_ood_acc = ood_utils.compute_incremental_test_acc(testloader, thresholds, trainer, classes, classes_idx_OOD, args.save_path, len(classes_idx_ID),classes_idx_ID,percent,save_k)
print("total in OOD accuracy for sample test data", test_ood_acc)
sample_data_len = int(0.1*len(curr_ood_data))
rem_data_len = len(curr_ood_data) - sample_data_len
sample_ood_data, remaining_ood_data = data.random_split(curr_ood_data,[sample_data_len, rem_data_len])
ood_acc = ood_utils.compute_ood_Acc([sample_ood_data], thresholds, trainer, classes,[classes_idx_OOD[k]], args.save_path, len(classes_idx_ID), args.multiple_dataset, classes_idx_ID, 0)
print("ood_acc :", ood_acc)
else:
ood_acc = 100
if ood_acc > 30:
# pdb.set_trace()
batch_size = 32
classes_idx_ID = np.array(np.insert(classes_idx_ID,len(classes_idx_ID),classes_idx_OOD[k]))
#
testset = [testset, ood_testset_list[k]]
testset = data.ConcatDataset(testset)
testloader = data.DataLoader(testset, batch_size = batch_size, shuffle = False, num_workers = 2)
current_act_avg_layers, test_acc_list,ood_acc_list,trainer = ood_utils.continual_learner(trainer, curr_ood_data, ood_testset_list[k], testloader, avg_act_all_layers, layer_indices, batch_size,classes_idx_OOD[k],classes_idx_ID,lr_mult)
test_acc_full.append(test_acc_list)
ood_acc_full.append(ood_acc_list)
ood_trainloader = data.DataLoader(curr_ood_data, batch_size = 1, shuffle = False, num_workers = 2)
_, _, _, activations_list_new_class = trainer.ood_evaluate(ood_trainloader)
thresholds = ood_utils.update_thresholds(thresholds, activations_list_new_class,trainer,classes_idx_ID)
new_avg = current_act_avg_layers
#
for i in range(len(new_avg)-1):
# new_avg[i] = (current_act_avg_layers[i] + avg_act_all_layers[i]*(1/(lr_mult*10)))
new_avg[i] = (current_act_avg_layers[i] + avg_act_all_layers[i])
avg_act_all_layers = new_avg
lr_mult = 2 # 2 for fmnist
# np.savez('cont_learner_with_another_nosp.npz', test_acc=test_acc_full, ood_acc= ood_acc_full)
if args.full_pipeline:
testloader = data.DataLoader(testset, batch_size = 1, shuffle = False, num_workers = 2)
for p in range(0,3):
percent = 0.1*p
test_ood_acc = ood_utils.compute_incremental_test_acc(testloader, thresholds, trainer, classes,classes_idx_OOD, args.save_path, len(classes_idx_ID),classes_idx_ID, percent,save_k)
print("total in OOD accuracy for sample test data", test_ood_acc)
print("Exiting Continual Learner experiment")
exit() ## the continula learner exits after running 1 experiment