-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
180 lines (157 loc) · 7.31 KB
/
train.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
import os
os.environ['PYTHONHASHSEED']=str(0)
import random
random.seed(0)
import numpy as np
np.random.seed(0)
import torch
import torch.nn.functional as F
torch.manual_seed(0)
import time
from dataset import Dataset
from model import *
from utils import *
from resnet import *
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print(device)
def train_step(args, epoch, dataset, pred_model, optimizer, scheduler, metric_history):
st = time.time()
train_acc = 0.0
train_loss = 0.0
sm_metric = 0
pred_model.train()
for images, labels in dataset.train_dl:
labels = labels.to(device)
images = images.to(device)
optimizer.zero_grad()
scores, _, r, labels = pred_model(images, y=labels, dataset=dataset)
ce_loss = F.cross_entropy(scores, labels.long())
loss = ce_loss
loss.backward()
optimizer.step()
train_loss += loss.item()
train_acc += calc_accuracy(scores, labels)
sm_metric += len(torch.where(torch.amax(torch.softmax(scores, 1), 1) == 1)[0])
if args.CLAMP_MAX is not None and args.CLAMP_MIN is not None:
with torch.no_grad():
pred_model.distnet.fc1.weight.clamp_(args.CLAMP_MIN, args.CLAMP_MAX)
train_loss = train_loss/(len(dataset.train_dl))
train_acc = train_acc/(len(dataset.train_dl))
metric_history['train_acc'].append(train_acc)
test_acc = 0.0
pred_model.eval()
wrong_idx = []
with torch.no_grad():
if args.REGULAR == False:
# calculate latent space centroids
centroids = torch.zeros(args.CLASSES, args.LATENT_DIM)
class_counts = torch.zeros(args.CLASSES)
for images, labels in dataset.centroid_dl:
images = images.to(device)
ls = pred_model.calc_latent(images)
class_idxs = [np.where(labels == i)[0] for i in range(0, args.CLASSES)]
for c, idxs in enumerate(class_idxs):
centroids[c, :] += torch.sum(ls[idxs], 0).cpu()
class_counts[c] += len(idxs)
centroids = centroids/class_counts.unsqueeze(1)
centroids = centroids.to(device)
# evaluation using test set
for i, batch in enumerate(dataset.test_dl):
images = batch[0].to(device)
labels = batch[1].to(device)
scores, _, _, _ = pred_model(images, centroids=centroids)
acc, idx = calc_accuracy(scores, labels, return_idx=True)
wrong_idx.extend(idx+(i*args.TEST_BATCH_SIZE))
test_acc += acc
else:
centroids = None
for i, batch in enumerate(dataset.test_dl):
images = batch[0].to(device)
labels = batch[1].to(device)
scores, _, _, _ = pred_model(images)
acc, idx = calc_accuracy(scores, labels, return_idx=True)
wrong_idx.extend(idx+(i*args.TEST_BATCH_SIZE))
test_acc += acc
wrong_idx = [idx.tolist() for idx in wrong_idx]
test_acc = test_acc/(len(dataset.test_dl))
metric_history['test_acc'].append(test_acc)
if scheduler is not None:
scheduler.step()
print('Epoch: %d || Softmax Metric: %d || Loss: %.3f || Acc: %.3f || Test_Acc: %.3f || Time: %.3f '
%(epoch, sm_metric, train_loss, train_acc, test_acc, time.time()-st))
return pred_model, optimizer, scheduler, metric_history, centroids, wrong_idx
if __name__ == "__main__":
# get settings
args = get_args()
# set directory
path = 'models/{args.DATASET}'
if args.REGULAR:
path += '_reg'
else:
path += '_twin'
if args.FREEZE:
path += '_freeze'
path += f'_{args.NET}'
path += f'_{args.TRAIN_BATCH_SIZE}'
path += f'_{args.LATENT_DIM}'
path += f'_{args.LEARNING_RATE}'
if not os.path.exists(path):
os.makedirs(path)
np.savez(path+f'/train_settings.npz', **args.__dict__)
max_test_acc = []
for t in range(args.NUM_REPEATS):
# get dataset
print('Loading dataset')
dataset = Dataset(args.DATASET)
args.CLASSES = dataset.n_classes
dataset.set_train_dataloader(args.TRAIN_BATCH_SIZE, t)
dataset.set_test_dataloader(args.TEST_BATCH_SIZE)
dataset.set_centroid_dataloader(args.TRAIN_BATCH_SIZE, t)
# get model
print('Loading model')
nets = {'ResNet50': [ResNet50(), 2048, 200], 'ResNet34': [ResNet34(), 512, 200], 'ResNet18': [ResNet18(), 512, 200], 'ConvNet': [ConvNet(), 512, 75], 'FCNet': [FCNet(), 256, 20]}
cnn_model = nets[args.NET][0]
dist_model = DistNet(args.LATENT_DIM, args.CLASSES, args.INIT_WEIGHT, args.REGULAR)
pred_model = PredictionNet(cnn_model, dist_model, nets[args.NET][1], args.LATENT_DIM, args.CLASSES, args.REGULAR)
pred_model = pred_model.to(device)
if args.FREEZE:
# freeze W matrix
for param in pred_model.distnet.parameters():
param.requires_grad = False
# set optimizer and scheduler
optimizer = torch.optim.SGD(pred_model.parameters(), lr=args.LEARNING_RATE, momentum=0.9, weight_decay=5e-4)
if args.NET == 'ConvNet':
milestones = [25, 50]
gamma = 0.1
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=milestones, gamma=gamma)
elif args.NET == 'ResNet18':
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=200)
elif args.NET == 'ResNet34' or args.NET == 'ResNet50':
try:
joint_optimizer_specs = [{'params': pred_model.convnet.parameters(), 'lr': args.LEARNING_RATE, 'weight_decay': 5e-4},
{'params': pred_model.distnet.parameters(), 'lr': args.LEARNING_RATE*10}, {'params': pred_model.fc1.parameters(), 'lr': args.LEARNING_RATE*10}, {'params': pred_model.bn1.parameters(), 'lr': args.LEARNING_RATE*10}]
except:
joint_optimizer_specs = [{'params': pred_model.convnet.parameters(), 'lr': args.LEARNING_RATE, 'weight_decay': 5e-4},
{'params': pred_model.fc1.parameters(), 'lr': args.LEARNING_RATE*10}]
optimizer = torch.optim.Adam(joint_optimizer_specs)
scheduler = None
else:
scheduler = None
# train model
print('Training model')
best_acc = 0.0
metric_history = {'train_acc': [], 'test_acc': []}
for epoch in range(nets[args.NET][2]):
dataset.reset_centroid_dataloader_iter()
pred_model, optimizer, scheduler, metric_history, centroids, wrong_idx = train_step(args, epoch, dataset, pred_model, optimizer, scheduler, metric_history)
current_acc = metric_history['test_acc'][-1]
if current_acc > best_acc:
save_model(epoch, pred_model, optimizer, scheduler, path+f'/model_{t}.pt')
np.save(path+f"/wrong_indices_{t}.npy", wrong_idx)
if centroids is not None:
np.save(path+f'/centroids_{t}.npy', centroids.cpu().detach().numpy())
best_acc = current_acc
# save metrics
np.savez(path+f'/train_metrics_{t}.npz', **metric_history)
max_test_acc.append(best_acc)
np.save(path+f'/test_accuracies.npy', max_test_acc)