-
Notifications
You must be signed in to change notification settings - Fork 1
/
main_seq.py
260 lines (208 loc) · 9.9 KB
/
main_seq.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
"""ensLoss with sequential training"""
# Authors: Ben Dai <bendai@cuhk.edu.hk>
# License: MIT License
## basics
import numpy as np
import pandas as pd
import random
import torch
from itertools import combinations
## dataloader
from loader import openml_data, img_data
from torch.utils.data import DataLoader
## models
import img_models
## Train
from train import Trainer, Trainer_seq
## args; print config, figure, out
import argparse
import pprint
import sys
from base import pairwise_ttest, line
## log to wandb
import wandb
def main(config, filename='PCam', n_trials=5, wandb_log=False):
## wandb log
if wandb_log:
wandb.init(project="ensLoss-seq", name=filename+'-'+config['model']['net'])
## Reproducibility
torch.manual_seed(0)
random.seed(0)
np.random.seed(0)
Acc = {'trial': [], 'loss': [], 'test_acc': [], 'test_auc': []}
path_={'method': [], 'loss': [], 'epoch': [], 'train_acc': [], 'test_acc': []}
for h in range(n_trials):
train_data, test_data = img_data(name=filename)
train_loader = DataLoader(dataset=train_data, batch_size=config['batch_size'], shuffle=True)
test_loader = DataLoader(dataset=test_data, batch_size=32)
## get some random training data
# dataiter = iter(train_loader)
# X_batch, y_batch = next(dataiter)
## ensLoss ##
# if 'ensLoss' in config['loss_list']:
# model = getattr(img_models, config['model']['net'])(num_classes=1)
# model.to(config['device'])
# if h==0:
# ## print the model in the first trial
# print(model)
# print('\n-- TRAIN ensLoss --\n')
# trainer_ = Trainer(model=model, loss='ensLoss', period=config['ensLoss_per_epochs'],
# config=config, device=config['device'],
# train_loader=train_loader, val_loader=test_loader)
# path_, acc_test, auc_test = trainer_.train(path_)
# Acc['trial'].append(h)
# Acc['loss'].append('ensLoss')
# Acc['test_acc'].append(acc_test)
# Acc['test_auc'].append(auc_test)
# BCE loss ##
if 'BCE' in config['loss_list']:
print('\n-- TRAIN BCE --\n')
model = getattr(img_models, config['model']['net'])(num_classes=1)
model.to(config['device'])
## Stage 1 train: BCE loss
trainer_ = Trainer(model=model, loss='BCELoss',
config=config, device=config['device'],
train_loader=train_loader, val_loader=test_loader)
path_, acc_test, auc_test = trainer_.train(path_)
torch.save(model.state_dict(), './save/seq/seq_model.pt')
model.load_state_dict(torch.load('./save/seq/seq_model.pt', weights_only=True))
## Stage 2 train: ensloss
trainer_seq = Trainer_seq(model=model, loss='ensLoss',
config=config, device=config['device'],
train_loader=train_loader, val_loader=test_loader)
path_, acc_test, auc_test = trainer_seq.train(path_)
path_['method'].extend(['BCE+ensLoss']*int((config['trainer']['epochs'] + config['trainer']['seq_epochs']) / config['trainer']['val_per_epochs']))
Acc['trial'].append(h)
Acc['loss'].append('BCE+ensLoss')
Acc['test_acc'].append(acc_test)
Acc['test_auc'].append(auc_test)
# # Hinge loss ##
# if 'Hinge' in config['loss_list']:
# print('\n-- TRAIN Hinge --\n')
# model = getattr(img_models, config['model']['net'])(num_classes=1)
# model.to(config['device'])
# trainer_ = Trainer_seq(model=model, loss='Hinge',
# config=config, device=config['device'],
# train_loader=train_loader, val_loader=test_loader)
# path_, acc_test, auc_test = trainer_.train(path_)
# Acc['trial'].append(h)
# Acc['loss'].append('Hinge+ensLoss')
# Acc['test_acc'].append(acc_test)
# Acc['test_auc'].append(auc_test)
# path_ = pd.DataFrame(path_)
# Acc = pd.DataFrame(Acc)
# # Plot
# mean_pd = path_.groupby(['epoch', 'loss'], as_index=False).mean()
# mean_pd = mean_pd.melt(id_vars=['epoch', 'loss'], var_name='type', value_name='mean')
# std_pd = path_.groupby(['epoch', 'loss'], as_index=False).std()
# std_pd = std_pd.melt(id_vars=['epoch', 'loss'], var_name='type', value_name='std')
# std_pd['std'] = std_pd['std'] / np.sqrt(n_trials)
# path_stat = pd.merge(mean_pd, std_pd, on=['epoch', 'loss', 'type'], suffixes=("", ""))
# fig = line(
# data_frame = path_stat,
# x = 'epoch',
# y = 'mean',
# error_y = 'std',
# error_y_mode = 'band',
# color = 'loss',
# line_dash='type',
# line_dash_map={'test_acc': 'solid', 'train_acc': 'dot'},
# title = f'Ave Test Acc in Epochs',
# )
# # fig.show()
# # Hypothesis Testing
# try:
# p_less = pairwise_ttest(df=Acc, val_col='test_acc', group_col='loss', alternative='less').round(5)
# p_less = p_less[p_less['B'] == 'ensLoss']
# p_greater = pairwise_ttest(df=Acc, val_col='test_acc', group_col='loss', alternative='greater').round(5)
# p_greater = p_greater[p_greater['B'] == 'ensLoss']
# p_less_auc = pairwise_ttest(df=Acc, val_col='test_auc', group_col='loss', alternative='less').round(5)
# p_less_auc = p_less_auc[p_less_auc['B'] == 'ensLoss']
# p_greater_auc = pairwise_ttest(df=Acc, val_col='test_auc', group_col='loss', alternative='greater').round(5)
# p_greater_auc = p_greater_auc[p_greater_auc['B'] == 'ensLoss']
# except:
# pass
# res_acc = Acc.groupby('loss').agg({'test_acc': ['mean', 'std']})
# res_acc[('test_acc', 'std')] /= np.sqrt(n_trials)
# res_acc = res_acc.T.round(4)
# res_auc = Acc.groupby('loss').agg({'test_auc': ['mean', 'std']})
# res_auc[('test_auc', 'std')] /= np.sqrt(n_trials)
# res_auc = res_auc.T.round(4)
# ## Save outcome
# orig_stdout = sys.stdout
# out_file = open('out_seq.txt', 'a+')
# sys.stdout = out_file
# print('\n#### %s - model: %s ####\n' %(filename, config['model']['net']))
# # print('\n Step Size: %s \n' %config['optimizer'])
# print('\n-- CONFIG --\n')
# pprint.pprint(config, width=1)
# print('\n-- Performance --\n')
# print((res_acc.round(4)).to_markdown())
# print('\n')
# print((res_auc.round(4)).to_markdown())
# print('\n-- Testing --\n')
# print(p_less.round(4).to_markdown())
# print('\n')
# print(p_greater.round(4).to_markdown())
# print(p_less_auc.round(4).to_markdown())
# print('\n')
# print(p_greater_auc.round(4).to_markdown())
# if wandb_log:
# wandb.log({"test_acc_curve": fig,
# "perf": Acc.groupby('loss')['test_acc'].agg(['mean', 'std']),
# "path": path_,
# "perf_table": Acc,
# "p_less": p_less,
# "p_greater": p_greater,
# "p_less_auc": p_less_auc,
# "p_greater_auc": p_greater_auc,
# })
# wandb.finish()
# out_file.close()
# sys.stdout = orig_stdout
if __name__=='__main__':
# PARSE THE ARGS
parser = argparse.ArgumentParser(description='ensLoss sequential Training')
parser.add_argument('-B', '--batch', default=128, type=int,
help='batch size of the training set')
parser.add_argument('-e', '--epoch', default=50, type=int,
help='number of epochs to train')
parser.add_argument('-F', '--filename', default="CIFAR", type=str,
help='filename of the dataset')
parser.add_argument('-N', '--net', default="ResNet50", type=str,
help='the neural net of the classification')
parser.add_argument('-SE', '--seq_epoch', default=100, type=int,
help='number of ens epochs to train')
parser.add_argument('-R', '--n_trials', default=5, type=int,
help='number of trials for the experiments')
parser.add_argument('--log', default=True, action=argparse.BooleanOptionalAction,
help='if save the training process in wandb')
args = parser.parse_args()
config = {
'loss_list': ['BCE', 'Hinge'],
'dataset' : args.filename,
'model': {'net': args.net},
'save_model': False,
'batch_size': args.batch,
'ensLoss_per_epochs': -1,
'trainer': {'epochs': args.epoch, 'seq_epochs': args.seq_epoch, 'val_per_epochs': 5},
'optimizer': {'lr': 1e-3, 'type': 'SGD', 'weight_decay': 5e-4,
'lr_scheduler': 'CosineAnnealingLR', 'args': {'T_max': args.epoch+args.seq_epoch}},
'device': torch.device("cuda:0" if torch.cuda.is_available() else "cpu")}
filename = args.filename
n_trials = args.n_trials
wandb_log = args.log
if filename == 'CIFAR':
## for a multi-class classification dataset
## Currently, we only support experiments that are based on the CIFAR10 dataset.
for (u,v) in combinations(range(10), 2):
filename_tmp = filename+str(u)+str(v)
main(config=config, filename=filename_tmp, n_trials=n_trials, wandb_log=wandb_log)
else:
## for a binary classification dataset
main(config=config, filename=filename, n_trials=n_trials, wandb_log=wandb_log)
## Image dataset
# CIFAR10: https://www.cs.toronto.edu/~kriz/cifar.html
# PCam: https://github.com/basveeling/pcam
# python main_seq.py -B=128 -e=100 -F="PCam" -R=5 --log
# python main_seq.py -B=128 -e=200 -F="CIFAR35" -N="VGG16" -R=1 --no-log