-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_backbone.py
170 lines (143 loc) · 5.71 KB
/
train_backbone.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
import argparse
import os
from datetime import datetime
import socket
import timeit
import torchvision
from torch.utils.tensorboard import SummaryWriter
import numpy as np
import torch
import torch.optim as optim
from torchvision import transforms
from torch.utils.data import DataLoader
import argparse
import matplotlib.pyplot as plt
import torch.nn as nn
import torch.nn.functional as F
from network.backbone import ResNet
from network.seghead import SegHead
from network.fpn import FPN101
from dataloaders.backbone_dataloader import BackboneDataset
from network.GeneralizedRCNN import GeneralizedRCNN
import cv2
def get_img_size(sequence):
if sequence==5:
return [640,480]
else:
return [1920,1080]
def main(cfg):
gpu_id = cfg.gpu_id
device = torch.device("cuda:" + str(gpu_id) if torch.cuda.is_available() else "cpu")
torch.cuda.set_device(gpu_id)
# # Setting other parameters
resume_epoch = 0 # Default is 0, change if want to resume
nEpochs = 1000 # Number of epochs for training (500.000/2079)
batch_size = 1
snapshot = 10 # Store a model every snapshot epochs
beta = 0.001
margin = 0.3
lr_B = 0.0001
lr_S = 0.0001
wd = 0.0002
save_root_dir = "models"
# save_dir = os.path.join(save_root_dir,"{:04}".format(sequence))
save_dir = "models"
if not os.path.exists(save_dir):
os.makedirs(os.path.join(save_dir))
backbone = GeneralizedRCNN()
seghead=SegHead([2048,1024])
BackBoneName = "GeneralizedRCNN"
SegHeadName = "seghead"
# backbone.load_state_dict(
# torch.load(os.path.join(save_dir, BackBoneName + '_epoch-' + str(24) + '.pth'),
# map_location=lambda storage, loc: storage))
# seghead.load_state_dict(
# torch.load(os.path.join(save_dir, SegHeadName + '_epoch-' + str(24) + '.pth'),
# map_location=lambda storage, loc: storage))
# Logging into Tensorboard
log_dir = os.path.join(save_dir, 'runs', datetime.now().strftime('%b%d_%H-%M-%S') + '_' + socket.gethostname())
writer = SummaryWriter(log_dir=log_dir, comment='-parent')
backbone=backbone.cuda(device)
seghead=seghead.cuda(device)
# Use the following optimizer
optimizerB = optim.Adam(backbone.parameters(), lr=lr_B, weight_decay=wd)
optimizerS = optim.Adam(seghead.parameters(), lr=lr_S, weight_decay=wd)
data_transforms = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomGrayscale(0.1),
transforms.ColorJitter()
])
# ms_train = [BackboneDataset(transform=data_transforms,level=0),BackboneDataset(transform=data_transforms,level=1),
# BackboneDataset(transform=data_transforms,level=2),BackboneDataset(transform=data_transforms,level=3)]
# loaders=[]
# for train_set in ms_train:
# loaders.append(DataLoader(train_set, batch_size=batch_size,num_workers=2))
trainSet = BackboneDataset(transform=data_transforms,level=3)
trainloader = DataLoader(trainSet, batch_size=batch_size)
ii=0
for epoch in range(resume_epoch, nEpochs):
start_time = timeit.default_timer()
# for lev,trainloader in enumerate(loaders):
for sample_batched in trainloader:
ii+=1
inputs,gts = sample_batched["img"],sample_batched["mask"]
inputs.requires_grad_()
inputs = inputs.cuda(device)
# print(gts)
gts = gts.cuda(device)
feature = backbone.forward(inputs)
out = seghead(feature)
# print(feature[0].shape)
# plt.imshow(feature[0].detach().cpu()[0][0])
# plt.show()
# plt.imshow(feature[1].detach().cpu()[0][0])
# plt.show()
# plt.imshow(feature[2].detach().cpu()[0][0])
# plt.show()
# plt.imshow(feature[3].detach().cpu()[0][0])
# plt.show()
plt.imshow(out.detach().cpu()[0][0])
plt.show()
output = out.detach().cpu().numpy()[0][0]
output[output<0.5]=0
print(output.shape)
output = cv2.resize(output, (2048, 1024))
plt.imshow(output)
plt.show()
plt.imshow(gts.cpu()[0][0])
plt.show()
loss = F.binary_cross_entropy_with_logits(out,gts)
# for pre,gt in zip(out,gts):
# losses.append()
# for pre, gt in zip(out, gts):
# gt=gt.cpu().detach().numpy()
# pre = pre.cpu().detach().numpy()
# plt.imshow(pre)
# plt.show()
# plt.imshow(gt)
# plt.show()
# loss = sum(losses)
backbone.zero_grad()
seghead.zero_grad()
loss.backward()
optimizerB.step()
optimizerS.step()
if ii % 1 == 0:
print(
"Iters: [%2d] time: %4.4f, loss: %.8f"
% (ii, timeit.default_timer() - start_time,loss.item())
)
if ii % 5 == 0:
writer.add_scalar('data/loss_iter', loss.item(), ii)
stop_time = timeit.default_timer()
print("Execution time: " + str(stop_time - start_time))
print("save models")
torch.save(backbone.state_dict(), os.path.join(save_dir, BackBoneName + '_epoch-' + str(epoch) + '.pth'))
torch.save(seghead.state_dict(), os.path.join(save_dir, SegHeadName + '_epoch-' + str(epoch) + '.pth'))
writer.close()
if __name__ == "__main__":
parser = argparse.ArgumentParser(prog='train.py')
parser.add_argument('--gpu_id', type=int, default=0, help='tracking buffer')
opt = parser.parse_args()
main(opt)