-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
141 lines (103 loc) · 4.8 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
import copy
import torch
import argparse
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.data import DataLoader
from nets.PAAD import PAAD
from utils import loss_fn, model_evaluation
from custom_dataset import InterventionDataset
def main(args):
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
if torch.cuda.is_available():
torch.cuda.manual_seed(args.seed)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
train_set = InterventionDataset(args.train_image_path, args.train_csv_path, 'train')
test_set = InterventionDataset(args.test_image_path, args.test_csv_path, 'test')
train_loader = DataLoader(
dataset=train_set, batch_size=args.train_batch_size, shuffle=True, num_workers=2)
test_loader = DataLoader(
dataset=test_set, batch_size=args.test_batch_size, shuffle=False, num_workers=2)
print("Dataset is ready. Start training...")
paad = PAAD(
device=device,
freeze_features=args.freeze_features,
pretrained_file=args.pretrained_file,
horizon=args.horizon).to(device)
parameters = filter(lambda p: p.requires_grad, paad.parameters())
optimizer = torch.optim.Adam(parameters, lr=args.learning_rate,
weight_decay=args.weight_decay)
alpha = 0.01 * len(train_loader)
# start training
best_ap = 0.0
train_loss_over_epochs = []
test_ap_over_epochs = []
#scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=50, gamma=0.2)
for epoch in range(args.epochs):
#print("Current learning rate: ", optimizer.param_groups[0]['lr'])
paad.train()
running_loss = 0.0
for iteration, (img, pred_traj_img, lidar_scan, label) in enumerate(train_loader):
img, pred_traj_img = img.to(device), pred_traj_img.to(device)
lidar_scan, label = lidar_scan.to(device), label.to(device)
recon_lidar, mean, log_var, pred_inv_score = paad(img, pred_traj_img, lidar_scan)
loss = loss_fn(recon_lidar, lidar_scan, mean, log_var, pred_inv_score, label, alpha)
optimizer.zero_grad()
loss.backward()
optimizer.step()
running_loss += loss.item()
running_loss /= len(train_loader)
print("Epoch {:02d}/{:02d}, Loss {:9.4f}".format(
epoch+1, args.epochs, running_loss))
# evaluate the model on the test set
ap = model_evaluation(test_loader, paad, device)
print("Average precision on the test set: {:.4f}".format(ap))
train_loss_over_epochs.append(running_loss)
test_ap_over_epochs.append(ap)
# save the best model
if ap > best_ap:
PATH = './nets/paad.pth'
torch.save(paad.state_dict(), PATH)
best_ap = copy.deepcopy(ap)
#scheduler.step()
# plot training set loss and test set average precision over epochs
fig = plt.figure()
plt.subplot(2,1,1)
plt.ylabel("Train loss")
plt.plot(np.arange(args.epochs)+1, train_loss_over_epochs, 'k-')
plt.title("Train loss and test average precision")
plt.xlim(1, args.epochs)
plt.grid(True)
plt.subplot(2,1,2)
plt.ylabel("Test average precision")
plt.plot(np.arange(args.epochs)+1, test_ap_over_epochs, 'b-')
plt.xlabel("Epochs")
plt.xlim(1, args.epochs)
plt.grid(True)
plt.savefig("learning_curve.png")
plt.close(fig)
print("Finished training.")
print("The best test average precision: {:.4f}".format(best_ap))
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# dataset parameters
parser.add_argument("--train_image_path", type=str, default='train_set/images_train/')
parser.add_argument("--train_csv_path", type=str, default='train_set/labeled_data_train.csv')
parser.add_argument("--test_image_path", type=str, default='test_set/images_test/')
parser.add_argument("--test_csv_path", type=str, default='test_set/labeled_data_test.csv')
# training parameters
parser.add_argument("--seed", type=int, default=230)
parser.add_argument("--epochs", type=int, default=50)
parser.add_argument("--train_batch_size", type=int, default=16)
parser.add_argument("--learning_rate", type=float, default=0.0005)
parser.add_argument("--weight_decay", type=float, default=0.00015)
parser.add_argument("--test_batch_size", type=int, default=64)
# model parameters
parser.add_argument("--freeze_features", type=bool, default=True)
parser.add_argument("--pretrained_file", type=str,
default="nets/VisionNavNet_state_hd.pth.tar")
parser.add_argument("--horizon", type=int, default=10)
args = parser.parse_args()
main(args)