-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
245 lines (196 loc) · 8.45 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
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
# import torch.nn.init
from colorama import Fore
from dataset import EurosatDataset
from model import Classifier
from sklearn.metrics import accuracy_score
from torch.autograd import Variable
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
import argparse
import numpy as np
import random
import torch
import torch.optim as optim
import warnings
import os
# suppress warnings
warnings.filterwarnings("ignore")
def parse_arguments():
"""
Object for parsing command line strings into Python objects.
"""
arg = argparse.ArgumentParser()
arg.add_argument("--epoch", type=int, default=10)
arg.add_argument(
"--device", type=int, default=torch.device("cuda" if torch.cuda.is_available() else "cpu"), choices=['cuda', 'cpu'])
arg.add_argument("--save_dir", type=str, default="./saved_models")
arg.add_argument("--data_dir", type=str, default="data/EuroSAT/")
arg.add_argument("--batch_size", type=int, default=128,
help="total number of batch size of labeled data")
arg.add_argument("--eval_batch_size", type=int, default=256,
help="batch size of evaluation data loader")
arg.add_argument("--criterion", type=str,
default="crossentropy", choices=['crossentropy'])
arg.add_argument("--optimizer", type=str,
default="sgd", choices=['sgd', 'adam'])
arg.add_argument("--learning_rate", type=float, default=0.01)
arg.add_argument("--momentum", type=float, default=0.9)
arg.add_argument("--weight_decay", type=float, default=5e-4)
arg.add_argument("--dropout", type=float, default=0.0,
choices=[0.0, 0.5, 0.7, 0.9, 0.99, 0.999])
arg.add_argument("--num_workers", type=int, default=3)
arg.add_argument("--seed", type=int, default=42)
return arg.parse_args()
def softmax(x):
"""
Sofmax function.
"""
return np.exp(x)/sum(np.exp(x))
def accuracy(gt_S, pred_S):
"""
Get the accuracy of the model.
Args:
gt_S : Ground truth.
pred_S : Prediction.
Returns:
Accuracy classification score.
"""
_, alp = torch.max(torch.from_numpy(pred_S), 1)
return accuracy_score(gt_S, np.asarray(alp))
def main(args):
"""
Image Classification Model Training.
"""
# pylint: disable=not-callable
# Seed for reproducibility
torch.manual_seed(args.seed)
np.random.seed(args.seed)
torch.cuda.manual_seed(args.seed)
random.seed(args.seed)
writer = SummaryWriter(
comment=f"Learn_{args.learning_rate}_Drop{args.dropout}")
# Construct Dataset
train_dataset = EurosatDataset(
is_train=True, seed=args.seed, root_dir=args.data_dir)
print(
f'Number of data on Train Dataset is {len(train_dataset)}')
test_dataset = EurosatDataset(
is_train=False, seed=args.seed, root_dir=args.data_dir)
print(
f'Number of data on Test Dataset is {len(test_dataset)}')
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
drop_last=True,
)
test_loader = torch.utils.data.DataLoader(
test_dataset,
batch_size=args.eval_batch_size,
shuffle=True,
num_workers=args.num_workers,
drop_last=True,
)
print(
f'The images split in train and test is based on the seed {args.seed}')
config = {
"input_size": train_dataset.size[0],
"labels": train_dataset.label_encoding,
"loss_function": args.criterion,
'dropout': args.dropout,
}
model = Classifier(config=config)
model = model.to(args.device)
if args.optimizer == 'adam':
optimizer = optim.Adam(model.parameters(), lr=args.learning_rate,
weight_decay=args.weight_decay)
elif args.optimizer == 'sgd':
optimizer = optim.SGD(model.parameters(), lr=args.learning_rate,
momentum=args.momentum, weight_decay=args.weight_decay)
else:
raise ValueError('Unknown optimizer')
test_losses = []
train_losses = []
test_acc = []
train_acc = []
best_acc = 0
correct_pred = {classname: 0 for classname in train_dataset.label_encoding}
total_pred = {classname: 0 for classname in train_dataset.label_encoding}
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir, exist_ok=True)
print(f'Start training with {args.epoch} epochs')
for e in range(1, args.epoch + 1):
with tqdm(train_loader, unit="batch", bar_format="{l_bar}%s{bar}%s{r_bar}" % (Fore.BLUE, Fore.RESET)) as train_epoch:
model.train()
for data, target in train_epoch:
# get the inputs; train_epoch is a list of [data, target]
data, target = (
Variable(data.to(args.device)),
Variable(target.to(args.device)),
)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
output = model(data)
loss = model.loss(output, target)
loss.backward()
optimizer.step()
train_losses = np.append(train_losses, loss.item())
pred = output.data.cpu().numpy() # [0]
pred = softmax(pred)
gt = target.data.cpu().numpy() # [0]
train_acc = np.append(train_acc, accuracy(gt, pred))
train_epoch.set_description(f"Epoch {e}")
train_epoch.set_postfix(
loss=loss.item(), acc=train_acc[-1], mean_loss=np.mean(train_losses), mean_acc=np.mean(train_acc))
writer.add_scalar('Loss/train', np.mean(train_losses), e)
writer.add_scalar('Accuracy/train', np.mean(train_acc), e)
with tqdm(test_loader, unit="batch", bar_format="{l_bar}%s{bar}%s{r_bar}" % (Fore.RED, Fore.RESET)) as test_epoch:
model.eval()
# since we're not training, we don't need to calculate the gradients for our outputs
with torch.no_grad():
for data, target in test_epoch:
data, target = (
Variable(data.to(args.device)),
Variable(target.to(args.device)),
)
# calculate outputs by running images through the network
output = model(data)
loss = model.loss(output, target)
test_losses = np.append(test_losses, loss.item())
# the class with the highest conf is what we choose as prediction
_, pred = torch.max(output, 1)
for label, prediction in zip(target, pred):
if label == prediction:
correct_pred[train_dataset.label_encoding[
label]] += 1
total_pred[train_dataset.label_encoding[label]] += 1
pred = output.data.cpu().numpy() # [0]
pred = softmax(pred)
gt = target.data.cpu().numpy() # [0]
test_acc = np.append(test_acc, accuracy(gt, pred))
test_epoch.set_description(f"Test")
test_epoch.set_postfix(
loss=loss.item(), acc=test_acc[-1], mean_loss=np.mean(test_losses), mean_acc=np.mean(test_acc))
writer.add_scalar('Loss/test', np.mean(test_losses), e)
writer.add_scalar('Accuracy/test', np.mean(test_acc), e)
state = {
'config': config,
'state_dict': model.state_dict(),
}
if np.mean(test_acc) > best_acc:
best_acc = np.mean(test_acc)
torch.save(state,
f"{args.save_dir}/model_best.pth")
torch.save(state,
f"{args.save_dir}/model_epoch{e}_acc{round(np.mean(test_acc),2)}.pth")
for classname, correct_count in correct_pred.items():
t_accuracy = 100 * float(correct_count) / total_pred[classname]
# print(f'Accuracy for class: {classname:5s} is {t_accuracy:.1f} %')
writer.add_scalar(f"Accuracy/{classname}", t_accuracy, e)
print('Finished Training')
print(f'Saving the model on {args.save_dir}')
if __name__ == "__main__":
args = parse_arguments()
main(args)