-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
154 lines (114 loc) · 5.72 KB
/
eval.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
import os
import time
import copy
import argparse
import numpy as np
import torch
import torch.nn as nn
from utils import get_dataset, get_network, get_eval_pool, ParamDiffAug, epoch
from eval_utils import get_eval_lrs
import random
from absl import logging
from ema_pytorch import EMA
def set_logger(log_level='info', fname=None):
import logging as _logging
handler = logging.get_absl_handler()
formatter = _logging.Formatter('%(asctime)s - %(filename)s - %(message)s')
handler.setFormatter(formatter)
logging.set_verbosity(log_level)
if fname is not None:
handler = _logging.FileHandler(fname)
handler.setFormatter(formatter)
logging.get_absl_logger().addHandler(handler)
def eval_synset(it_run, net, testloader, args):
net = net.to(args.device)
ema = EMA(net, beta=0.995, power=1, update_after_step=0, update_every=1)
criterion = nn.CrossEntropyLoss().to(args.device)
start = time.time()
with torch.no_grad():
loss_test, acc_test, auc_test, fscore_test = epoch('test', testloader, ema, None, criterion, args, aug=False)
time_test = time.time() - start
logging.info(f'Evaluate_{it_run}: TestAcc:\t{acc_test}, AUC:\t{auc_test}, Fscore:\t{fscore_test}, eval time = {int(time_test)} s ')
return net, acc_test, auc_test, fscore_test
def eval_loop(testloader=None, model_eval_pool=[], channel=3, num_classes=9, im_size=(32, 32), args=None):
curr_acc_dict = {}
curr_std_dict = {}
eval_pool_dict = get_eval_lrs(args)
max_epoch = {'ConvNet': 2700, 'ResNet18': 2900}
for model_eval in model_eval_pool:
start_model = time.time()
logging.info(f'-------------------------\nTrain and Evaluation: model = {model_eval}')
accs_test = []
aucs_test = []
fscores_test = []
for it_run in range(args.num_run):
net_model = get_network(model_eval, channel, num_classes, im_size, width=args.width,
depth=args.depth, args=args).to(args.device) # get a random model
model_path = os.path.join(args.save_path, f'{model_eval}{max_epoch[model_eval]}_{it_run}.pth')
logging.info(f'Load model from {model_path}')
net_model.load_state_dict(torch.load(model_path))
args.lr_net = eval_pool_dict[model_eval]
trained_net, acc_score, auc_score, fscore = eval_synset(it_run, net_model, testloader, args=args)
accs_test.append(acc_score)
aucs_test.append(auc_score)
fscores_test.append(fscore)
logging.info(f'accs: {accs_test}')
logging.info(f'aucs: {aucs_test}')
logging.info(f'fscores: {fscores_test}')
accs_test = np.array(accs_test)
acc_test_mean = np.mean(accs_test)
acc_test_std = np.std(accs_test)
curr_acc_dict[model_eval] = acc_test_mean
curr_std_dict[model_eval] = acc_test_std
aucs_test = np.array(aucs_test)
auc_test_mean = np.mean(aucs_test)
fscores_test = np.array(fscores_test)
fscores_test_mean = np.mean(fscores_test)
logging.info(f'Evaluate Acc {len(accs_test)} random {model_eval}, '
f'mean = {acc_test_mean} std = {acc_test_std}\n-------------------------')
logging.info(f'Evaluate AUC {len(aucs_test)} random {model_eval}, '
f'mean = {auc_test_mean} std = {np.std(aucs_test)}\n-------------------------')
logging.info(f'Evaluate FScore {len(fscores_test)} random {model_eval}, '
f'mean = {fscores_test_mean} std = {np.std(fscores_test)}\n-------------------------')
end_model = time.time()
used_model = time.strftime("%H:%M:%S", time.gmtime(end_model - start_model))
logging.info(f'The passed time of training model {model_eval} is {used_model}')
logging.info({
'Accuracy/Avg_All'.format(model_eval): np.mean(np.array(list(curr_acc_dict.values()))),
'Std/Avg_All'.format(model_eval): np.mean(np.array(list(curr_std_dict.values())))})
curr_acc_dict.pop("{}".format(args.model))
curr_std_dict.pop("{}".format(args.model))
logging.info({
'Accuracy/Avg_Cross'.format(model_eval): np.mean(np.array(list(curr_acc_dict.values()))),
'Std/Avg_Cross'.format(model_eval): np.mean(np.array(list(curr_std_dict.values())))})
def main(args):
torch.random.manual_seed(0)
np.random.seed(0)
random.seed(0)
args.batch_train = 64
args.batch_test = 64
args.res = 256
args.distributed = True
args.dsa_param = ParamDiffAug()
args.dsa = False if args.dsa_strategy in ['none', 'None'] else True
run_dir = '20240225-163550-chocolate-universe-381'
args.save_path = os.path.join(args.save_path, run_dir)
start_time = time.time()
set_logger(log_level='info', fname=os.path.join(args.save_path, 'output_eval.log'))
channel, im_size, num_classes, class_names, mean, std, dst_train, dst_test, testloader, loader_train_dict, class_map, class_map_inv = get_dataset(
args.dataset, args.data_path, args.batch_train, args.res, args=args)
model_eval_pool = get_eval_pool(args.eval_mode, args.model, args.model)
logging.info('training begins')
logging.info(f'Hyper-parameters: \n {args.__dict__}')
logging.info(f'Evaluation model pool: {model_eval_pool}')
eval_loop(testloader=testloader, model_eval_pool=model_eval_pool,
channel=channel, num_classes=num_classes, im_size=im_size, args=args)
end_time = time.time()
used_time = time.strftime("%H:%M:%S", time.gmtime(end_time - start_time))
logging.info(f'The passed time is {used_time}')
print('Training finished')
if __name__ == '__main__':
import shared_args
parser = shared_args.add_shared_args()
args = parser.parse_args()
main(args)