-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_openlane.py
165 lines (151 loc) · 7.29 KB
/
train_openlane.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
import sys
sys.path.append('/mnt/ve_perception/wangruihao/code/BEV-LaneDet')
import torch
import time
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch.utils.data import DataLoader
import torch.nn as nn
from models.util.load_model import load_checkpoint, resume_training
from models.util.save_model import save_model_dp
from models.loss import IoULoss, NDPushPullLoss
from utils.config_util import load_config_module
from sklearn.metrics import f1_score
import numpy as np
class Combine_Model_and_Loss(torch.nn.Module):
def __init__(self, model):
super(Combine_Model_and_Loss, self).__init__()
self.model = model
self.bce = torch.nn.BCEWithLogitsLoss(pos_weight=torch.tensor([10.0]))
self.iou_loss = IoULoss()
self.poopoo = NDPushPullLoss(1.0, 1., 1.0, 5.0, 200)
self.mse_loss = nn.MSELoss()
self.bce_loss = nn.BCELoss()
# self.sigmoid = nn.Sigmoid()
def forward(self, inputs, gt_seg=None, gt_instance=None, gt_offset_y=None, gt_z=None, image_gt_segment=None,
image_gt_instance=None, train=True):
res = self.model(inputs)
pred, emb, offset_y, z = res[0]
pred_2d, emb_2d = res[1]
#weight = res[2]
#gamma_32 = res[2]
#gamma_64 = res[3]
if train:
## 3d
loss_seg = self.bce(pred, gt_seg) + self.iou_loss(torch.sigmoid(pred), gt_seg)
loss_emb = self.poopoo(emb, gt_instance)
loss_offset = self.bce_loss(gt_seg * torch.sigmoid(offset_y), gt_offset_y)
loss_z = self.mse_loss(gt_seg * z, gt_z)
loss_total = 3 * loss_seg + 0.5 * loss_emb
loss_total = loss_total.unsqueeze(0)
loss_offset = 60 * loss_offset.unsqueeze(0)
loss_z = 30 * loss_z.unsqueeze(0)
## 2d
loss_seg_2d = self.bce(pred_2d, image_gt_segment) + self.iou_loss(torch.sigmoid(pred_2d), image_gt_segment)
loss_emb_2d = self.poopoo(emb_2d, image_gt_instance)
loss_total_2d = 3 * loss_seg_2d + 0.5 * loss_emb_2d
loss_total_2d = loss_total_2d.unsqueeze(0)
return pred, loss_total, loss_total_2d, loss_offset, loss_z#, weight, gamma_32, gamma_64
else:
return pred
def train_epoch(model, dataset, optimizer, configs, epoch):
# Last iter as mean loss of whole epoch
model.train()
time_start = time.time()
losses_avg = {}
'''image,image_gt_segment,image_gt_instance,ipm_gt_segment,ipm_gt_instance'''
for idx, (
input_data, gt_seg_data, gt_emb_data, offset_y_data, z_data, image_gt_segment, image_gt_instance) in enumerate(
dataset):
# loss_back, loss_iter = forward_on_cuda(gpu, gt_data, input_data, loss, models)
input_data = input_data.cuda()
gt_seg_data = gt_seg_data.cuda()
gt_emb_data = gt_emb_data.cuda()
offset_y_data = offset_y_data.cuda()
z_data = z_data.cuda()
image_gt_segment = image_gt_segment.cuda()
image_gt_instance = image_gt_instance.cuda()
prediction, loss_total_bev, loss_total_2d, loss_offset, loss_z = model(input_data,
gt_seg_data,
gt_emb_data,
offset_y_data, z_data,
image_gt_segment,
image_gt_instance)
loss_back_bev = loss_total_bev.mean()
loss_back_2d = loss_total_2d.mean()
loss_offset = loss_offset.mean()
loss_z = loss_z.mean()
loss_back_total = loss_back_bev + 0.5 * loss_back_2d + loss_offset + loss_z
''' caclute loss '''
optimizer.zero_grad()
loss_back_total.backward()
optimizer.step()
if idx % 50 == 0:
print(idx, loss_back_bev.item(), '*' * 10)
if idx % 300 == 0:
target = gt_seg_data.detach().cpu().numpy().ravel()
pred = torch.sigmoid(prediction).detach().cpu().numpy().ravel()
f1_bev_seg = f1_score((target > 0.5).astype(np.int64), (pred > 0.5).astype(np.int64), zero_division=1)
loss_iter = {"BEV Loss": loss_back_bev.item(), 'offset loss': loss_offset.item(), 'z loss': loss_z.item(),
"F1_BEV_seg": f1_bev_seg}
# losses_show = loss_iter
time_epoch = time.time()
print(idx, loss_iter)
#print('weight = ', weight)
#print('weight_32 = ', weight_32)
#print('gamma_32 = ', gamma_32)
#print('weight_64 = ', weight_64)
#print('gamma_64 = ', gamma_64)
print('time: ', time_epoch - time_start)
#print('weight = ', weight)
#print('weight_32 = ', weight_32)
#print('gamma_32 = ', gamma_32)
#print('weight_64 = ', weight_64)
#print('gamma_64 = ', gamma_64)
def worker_function(config_file, gpu_id, checkpoint_path=None):
print('use gpu ids is '+','.join([str(i) for i in gpu_id]))
configs = load_config_module(config_file)
''' models and optimizer '''
model = configs.model()
model = Combine_Model_and_Loss(model)
if torch.cuda.is_available():
model = model.cuda()
model = torch.nn.DataParallel(model)
optimizer = configs.optimizer(filter(lambda p: p.requires_grad, model.parameters()), **configs.optimizer_params)
scheduler = getattr(configs, "scheduler", CosineAnnealingLR)(optimizer, configs.epochs)
if checkpoint_path:
if getattr(configs, "load_optimizer", True):
resume_training(checkpoint_path, model.module, optimizer, scheduler)
else:
load_checkpoint(checkpoint_path, model.module, None)
''' dataset '''
Dataset = getattr(configs, "train_dataset", None)
if Dataset is None:
Dataset = configs.training_dataset
train_loader = DataLoader(Dataset(), **configs.loader_args, pin_memory=True)
''' get validation '''
# if configs.with_validation:
# val_dataset = Dataset(**configs.val_dataset_args)
# val_loader = DataLoader(val_dataset, **configs.val_loader_args, pin_memory=True)
# val_loss = getattr(configs, "val_loss", loss)
# if eval_only:
# loss_mean = val_dp(model, val_loader, val_loss)
# print(loss_mean)
# return
time_start = time.time()
for epoch in range(configs.epochs):
print('*' * 100, epoch)
train_epoch(model, train_loader, optimizer, configs, epoch)
scheduler.step()
save_model_dp(model, optimizer, configs.model_save_path, 'ep%03d.pth' % epoch)
save_model_dp(model, None, configs.model_save_path, 'latest.pth')
time_end = time.time()
time_spent = time_end - time_start
hour_spent = time_spent // 3600
min_spent = (time_spent - hour_spent * 3600) // 60
sec_spent = (time_spent - hour_spent * 3600 - min_spent * 60) // 1
print('total time :', hour_spent, 'hours', min_spent, 'minutes', sec_spent, 'seconds')
# TODO template config file.
if __name__ == '__main__':
import warnings
warnings.filterwarnings("ignore")
worker_function('./openlane_config.py', gpu_id=[0, 1])