-
Notifications
You must be signed in to change notification settings - Fork 2
/
main.py
163 lines (138 loc) · 6.81 KB
/
main.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
import os
import argparse
import ssl
ssl._create_default_https_context = ssl._create_unverified_context
from tqdm import tqdm
import pandas as pd
import torch
from torchvision import transforms
from torch.utils.data import DataLoader
from data.stl10 import STL10PAIR
from model.resnet import ResNet18
from tools.normalize import Normalize
from model.npc_model import NonParametricClassifier
from tools.averagetracker import AverageTracker
from tools.tools import check_clustering_metrics
from losses.crossview_contrastive_Loss import crossview_contrastive_Loss
from losses.Loss_ID import Loss_ID
from losses.Loss_FMI import Loss_FMI
def parse():
parser = argparse.ArgumentParser()
parser.add_argument("-g", "--gpus", type=str, default="4")
parser.add_argument("-n", "--num_workers", type=int, default=8)
parser.add_argument('--batch_size', default=256, type=int, help='Number of images in each mini-batch')
parser.add_argument('--epochs', default=2000, type=int, help='Number of sweeps over the dataset to train')
parser.add_argument("--pretrained", default="", type=str, help="path to pretrained models")
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
return args
def main():
args = parse()
batch_size, epochs = args.batch_size, args.epochs
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_transform = transforms.Compose([
transforms.RandomResizedCrop(size=96),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomApply([transforms.ColorJitter(0.4, 0.4, 0.4, 0.1)], p=0.8),
transforms.RandomGrayscale(p=0.2),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
train_data= STL10PAIR(root='data', split='labeled', transform=train_transform, download=True)
train_loader = DataLoader(train_data, batch_size=batch_size, shuffle=True, num_workers=16, pin_memory=True,
drop_last=True)
low_dim = 128
net = ResNet18(low_dim=low_dim)
norm = Normalize(2)
npc = NonParametricClassifier(input_dim=low_dim,
output_dim=len(train_data),
tau=1.0,
momentum=0.5)
loss_id = Loss_ID(tau2=2.0)
loss_fmi = Loss_FMI()
net, norm = net.to(device), norm.to(device)
npc, loss_id, loss_fmi = npc.to(device), loss_id.to(device), loss_fmi.to(device)
# lr = 0.03
optimizer = torch.optim.SGD(net.parameters(),
lr=0.05,
momentum=0.9,
weight_decay=5e-4,
nesterov=False,
dampening=0)
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer,
[500, 950, 1350, 2050, 2350, 2750, 3350, 3750, 4250, 4550],
gamma=0.5)
if torch.cuda.is_available():
net = torch.nn.DataParallel(net,
device_ids=range(len(
args.gpus.split(","))))
torch.backends.cudnn.benchmark = True
trackers = {n: AverageTracker() for n in ["loss", "loss_id", "loss_fmi", "loss_imi"]}
if os.path.exists(args.pretrained):
print('Restart from checkpoint {}'.format(args.pretrained))
checkpoint = torch.load(args.pretrained, map_location='cpu')
optimizer.load_state_dict(checkpoint['optimizer'])
net.load_state_dict(checkpoint['model'])
net.cuda()
start_epoch = checkpoint['epoch']
else:
print('No checkpoint file at {}'.format(args.pretrained))
start_epoch = 0
net = net.cuda()
file_name = "demo.py"
results = {'Epochs': [], 'loss_id': [],'loss_fmi':[],'loss_imi':[], 'k_means_acc': [], 'k_means_nmi': [],
'k_means_ari': []}
save_name_pre = '{}_{}_{}'.format(file_name, batch_size, epochs)
if not os.path.exists('results'):
os.mkdir('results')
tmp = 1
for epoch in range(start_epoch+1, epochs + 1):
net.train()
total_loss, total_num, train_bar = 0.0, 0, tqdm(train_loader)
for pos_1, pos_2, target, index in train_bar:
optimizer.zero_grad()
inputs_1 = pos_1.to(device, dtype=torch.float32, non_blocking=True)
inputs_2 = pos_2.to(device, dtype=torch.float32, non_blocking=True)
indexes = index.to(device, non_blocking=True)
features_1 = norm(net(inputs_1))
features_2 = norm(net(inputs_2))
outputs = npc(features_1, indexes)
loss_imi = crossview_contrastive_Loss(features_1, features_2)
loss_id = loss_id(outputs, indexes)
loss_fmi = loss_fmi(features_1)
#loss_fmi = loss_fmi(features_2)
tot_loss = loss_id + 0.00001 * loss_fmi + 0.000001 * loss_imi
tot_loss.backward()
optimizer.step()
# track loss
trackers["loss"].add(tot_loss.item())
trackers["loss_id"].add(loss_id.item())
trackers["loss_imi"].add(loss_imi.item())
trackers["loss_fmi"].add(loss_fmi.item())
lr_scheduler.step()
# logging
# postfix = {name: t.avg() for name, t in trackers.items()}
# epoch_bar.set_postfix(**postfix)
# for t in trackers.values():
# t.reset()
# check clustering acc
# torch.cuda.empty_cache()
if (epoch == 0) or (((epoch + 1) % 10) == 0):
acc, nmi, ari = check_clustering_metrics(npc, train_loader)
print("Epoch:{} Loss_id:{} Loss_mr:{} Loss_cl:{} Kmeans ACC, NMI, ARI = {}, {}, {}".format(epoch+1, loss_id, loss_fmi, loss_imi, acc, nmi, ari))
results['Epochs'].append(tmp * 10)
results['loss_id'].append(loss_id.item())
results['loss_fmi'].append(loss_fmi.item())
results['loss_imi'].append(loss_imi.item())
results['k_means_acc'].append(acc)
results['k_means_nmi'].append(nmi)
results['k_means_ari'].append(ari)
# Checkpoint
print('Checkpoint ...')
torch.save({'optimizer': optimizer.state_dict(), 'model': net.state_dict(),
'epoch': epoch + 1}, args.pretrained)
tmp = tmp + 1
# save statistics df = pd.DataFrame.from_dict(d, orient='index')
data_frame = pd.DataFrame.from_dict(data=results, orient='index')
data_frame.to_csv('results/{}_statistics.csv'.format(save_name_pre))
if __name__ == "__main__":
main()