/
bbb_ensemble.py
201 lines (149 loc) · 7.29 KB
/
bbb_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
import sys
import numpy as np
import matplotlib.pyplot as plt
import torch
import torch.optim as optim
from torch.optim.lr_scheduler import ReduceLROnPlateau
import utils
import wandb
from pathlib import Path
from models import Classifier_BBB, Classifier_ConvBBB
from datamodules import MiraBestDataModule
#vars = parse_args()
config_dict, config = utils.parse_config('config_augment.txt')
jobid = int(sys.argv[1])
seed = config_dict['training']['seed'] + jobid
data_seed = config_dict['training']['seed_data'] + jobid
torch.manual_seed(seed)
#prior
prior = config_dict['priors']['prior']
prior_var = torch.tensor([float(i) for i in config_dict['priors']['prior_init'].split(',')])[1]
augment = config_dict['data']['augment']
#training
#imsize = config_dict['training']['imsize']
epochs = config_dict['training']['epochs']
nclass = config_dict['training']['num_classes']
learning_rate = torch.tensor(config_dict['training']['lr0'])
momentum = torch.tensor(config_dict['training']['momentum'])
weight_decay = torch.tensor(config_dict['training']['decay'])
reduction = config_dict['training']['reduction']
burnin = config_dict['training']['burnin']
T = config_dict['training']['temp']
kernel_size = config_dict['training']['kernel_size']
pac = config_dict['training']['pac']
base = config_dict['model']['base']
early_stopping = config_dict['model']['early_stopping']
conditioner = config_dict['model']['conditioner']
path_out = config_dict['output']['path_out']
#output
file_path = path_out + str(jobid) + '/'
filename = config_dict['output']['filename_uncert']
wandb_name = 'VI ' + str(jobid)
wandb.init(
project= "Evaluating-VI",
config = {
"seed": seed,
"data_seed": data_seed,
"learning_rate": learning_rate,
# "weight_decay": weight_decay,
# "factor": factor,
# "patience": patience,
"epochs": epochs,
"conditioner": conditioner,
"prior": prior,
"prior_var": prior_var,
"augmentation": augment,
"temp": T,
"optimiser": "Adam",
},
name=wandb_name,
)
#load data
datamodule = MiraBestDataModule(config_dict, config, data_seed)
train_loader, validation_loader, train_sampler, valid_sampler = datamodule.train_val_loader()
test_loader = datamodule.test_loader()
#check if a GPU is available:
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
# print("Device: ",device)
input_ch = 1
out_ch = nclass #y.view(-1)
kernel_size = kernel_size
model = Classifier_ConvBBB(input_ch, out_ch, kernel_size, prior_var, prior).to(device)
for i in range (1):
model = Classifier_ConvBBB(input_ch, out_ch, kernel_size, prior_var, prior).to(device)
optimizer = optim.Adam(model.parameters(), lr = learning_rate)
# scheduler = ReduceLROnPlateau(optimizer=optimizer, mode= 'min', factor=0.95, patience=3, verbose=False)
epoch_trainaccs, epoch_testaccs = [], []
epoch_trainloss, epoch_testloss = [], []
epoch_trainloss_complexity, epoch_testloss_complexity = [], []
epoch_trainloss_loglike, epoch_testloss_loglike = [], []
epoch_trainloss_complexity_conv, epoch_testloss_complexity_conv = [], []
epoch_trainloss_complexity_linear, epoch_testloss_complexity_linear = [], []
epoch_testerr = []
epoch_trainerr = []
_bestacc = 0.
for epoch in range(epochs):
train_loss, train_loss_c, train_loss_l, train_accs, train_complexity_conv,
train_complexity_linear = utils.train(model, train_loader, optimizer, device,
T, burnin, reduction, pac)
test_loss, test_loss_c, test_loss_l, test_accs, test_complexity_conv,
test_complexity_linear = utils.validate(model, validation_loader, device, T, burnin,
reduction, epoch, prior, prior_var, pac, file_path)
epoch_trainaccs.append(np.sum(train_accs)/len(train_sampler))
epoch_testaccs.append(np.sum(test_accs)/len(valid_sampler))
epoch_trainerr.append(100.*(1 - np.sum(train_accs)/len(train_sampler)))
epoch_testerr.append(100.*(1 - np.sum(test_accs)/len(valid_sampler)))
epoch_trainloss.append(np.sum(train_loss)/len(train_sampler))
epoch_testloss.append(np.sum(test_loss)/len(valid_sampler))
epoch_trainloss_complexity.append(np.sum(train_loss_c)/len(train_sampler))
epoch_trainloss_loglike.append(np.sum(train_loss_l)/len(train_sampler))
epoch_testloss_complexity.append(np.sum(test_loss_c)/len(valid_sampler))
epoch_testloss_loglike.append(np.sum(test_loss_l)/len(valid_sampler))
epoch_trainloss_complexity_conv.append(np.sum(train_complexity_conv)/len(train_sampler))
epoch_trainloss_complexity_linear.append(np.sum(train_complexity_linear)/len(train_sampler))
epoch_testloss_complexity_conv.append(np.sum(test_complexity_conv)/len(valid_sampler))
epoch_testloss_complexity_linear.append(np.sum(test_complexity_linear)/len(valid_sampler))
# scheduler.step(epoch_testloss_loglike[-1])
accuracy = epoch_testaccs[-1]
# check early stopping criteria:
if early_stopping and accuracy>_bestacc:
_bestacc = accuracy
torch.save(model.state_dict(), file_path + "model.pt")
torch.save(model.state_dict(), file_path + "model"+str(i)+".pt")
torch.save(optimizer.state_dict(), file_path + "model_optim.pt")
best_acc = accuracy
best_epoch = epoch
wandb.log({"train_loss":epoch_trainloss[epoch],
"train_loglikelihood": epoch_trainloss_loglike[epoch],
"train_complexity": epoch_trainloss_complexity[epoch],
"train_error": epoch_trainerr[epoch],
"val_loss": epoch_testloss[epoch],
"val_loglikelihood": epoch_testloss_loglike[epoch],
"val_complexity": epoch_testloss_complexity[epoch],
"val_error": epoch_testerr[epoch]
})
print('Finished Training')
print("Final validation error: ",100.*(1 - epoch_testaccs[-1]))
if early_stopping:
print("Best validation error: ",100.*(1 - best_acc)," @ epoch: "+str(best_epoch))
if not early_stopping:
torch.save(model.state_dict(), file_path + "model.pt")
# print(100.*(1 - best_acc))
# print(best_epoch)
# best_verr = 100-best_acc
# wandb.log({"best_vloss_epoch": best_epoch, "best_vloss": best_vloss})
wandb.log({"best_err_epoch": best_epoch, "best_err": 100.*(1 - best_acc)})
#calculate test error
model = Classifier_ConvBBB(input_ch, out_ch, kernel_size, prior_var, prior).to(device)
model.load_state_dict(torch.load(file_path+"model.pt"))
test_err= utils.test(model, test_loader, device, T, burnin, reduction, pac)
err_arr = []
for i in range(200):
test_err = utils.test(model, test_loader, device, T, burnin, reduction, pac)
err_arr.append(test_err)
wandb.log({"Mean test error":np.mean(err_arr), "Std test error": np.std(err_arr)})
wandb.finish()
#
#get_samples(model, n_samples = 10000, n_params = 5, log_space = False)