-
Notifications
You must be signed in to change notification settings - Fork 34
/
detecter.py
213 lines (167 loc) · 7.92 KB
/
detecter.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
from __future__ import print_function
import argparse
import os
import time
import torch
import torch.backends.cudnn as cudnn
import numpy as np
import skimage
import torch.cuda as ct
from net_builder import SUPPORT_NETS, build_net
from losses.multiscaleloss import multiscaleloss
import torch.nn.functional as F
import torch.nn as nn
from dataloader.StereoLoader import StereoDataset
from dataloader.SceneFlowLoader import SceneFlowDataset
from utils.preprocess import scale_disp, save_pfm
from utils.common import count_parameters
from torch.utils.data import DataLoader
from torchvision import transforms
import psutil
from torch2trt import torch2trt
from networks.submodules import build_corr, channel_length
process = psutil.Process(os.getpid())
cudnn.benchmark = True
def load_model_trained_with_DP(net, state_dict):
own_state = net.state_dict()
for name, param in state_dict.items():
own_state[name[7:]].copy_(param)
def check_tensorrt(y, y_trt):
print(torch.max(torch.abs(y - y_trt)))
def trt_transform(net):
x = torch.rand((1, 6, 576, 960)).cuda()
net.extract_network = torch2trt(net.extract_network, [x])
# extract features
conv1_l, conv2_l, conv3a_l, conv3a_r = net.extract_network(x)
# build corr
out_corr = build_corr(conv3a_l, conv3a_r, max_disp=40)
# generate first-stage flows
net.dispcunet = torch2trt(net.dispcunet, [x, conv1_l, conv2_l, conv3a_l, out_corr])
dispnetc_flows = net.dispcunet(x, conv1_l, conv2_l, conv3a_l, out_corr)
dispnetc_final_flow = dispnetc_flows[0]
diff_img0 = x[:, :3, :, :] - x[:, 3:, :, :]
norm_diff_img0 = channel_length(diff_img0)
# concat img0, img1, img1->img0, flow, diff-mag
inputs_net2 = torch.cat((x, x[:, 3:, :, :], dispnetc_final_flow, norm_diff_img0), dim = 1)
net.dispnetres = torch2trt(net.dispnetres, [inputs_net2, dispnetc_final_flow])
return net
def detect(opt):
net_name = opt.net
model = opt.model
result_path = opt.rp
file_list = opt.filelist
filepath = opt.filepath
if not os.path.exists(result_path):
os.makedirs(result_path)
devices = [int(item) for item in opt.devices.split(',')]
ngpu = len(devices)
# build net according to the net name
if net_name == "psmnet" or net_name == "ganet":
net = build_net(net_name)(192)
elif net_name in ["fadnet", "dispnetc"]:
net = build_net(net_name)(batchNorm=False, lastRelu=True)
elif net_name == "mobilefadnet":
#B, max_disp, H, W = (wopt.batchSize, 40, 72, 120)
shape = (opt.batchSize, 40, 72, 120) #TODO: Should consider how to dynamically use
warp_size = (opt.batchSize, 3, 576, 960)
net = build_net(net_name)(batchNorm=False, lastRelu=True, input_img_shape=shape, warp_size=warp_size)
if ngpu > 1:
net = torch.nn.DataParallel(net, device_ids=devices)
model_data = torch.load(model)
print(model_data.keys())
if 'state_dict' in model_data.keys():
#net.load_state_dict(model_data['state_dict'])
load_model_trained_with_DP(net, model_data['state_dict'])
else:
net.load_state_dict(model_data)
num_of_parameters = count_parameters(net)
print('Model: %s, # of parameters: %d' % (net_name, num_of_parameters))
batch_size = int(opt.batchSize)
test_dataset = StereoDataset(txt_file=file_list, root_dir=filepath, phase='detect')
test_loader = DataLoader(test_dataset, batch_size = batch_size, \
shuffle = False, num_workers = 1, \
pin_memory = False)
net.eval()
#net.dispnetc.eval()
#net.dispnetres.eval()
net = net.cuda()
#for i, sample_batched in enumerate(test_loader):
# input = torch.cat((sample_batched['img_left'], sample_batched['img_right']), 1)
# num_of_samples = input.size(0)
# input = input.cuda()
# x = input
# break
#net_trt = trt_transform(net)
net_trt = net.get_tensorrt_model()
#net_trt = net
torch.save(net_trt.state_dict(), 'models/mobilefadnet_trt.pth')
s = time.time()
avg_time = []
display = 50
warmup = 10
for i, sample_batched in enumerate(test_loader):
#if i > 215:
# break
stime = time.time()
input = torch.cat((sample_batched['img_left'], sample_batched['img_right']), 1)
print('input Shape: {}'.format(input.size()))
num_of_samples = input.size(0)
#output, input_var = detect_batch(net, sample_batched, opt.net, (540, 960))
input = input.cuda()
input_var = input #torch.autograd.Variable(input, volatile=True)
input_var = F.interpolate(input_var, (576, 960), mode='bilinear')
iotime = time.time()
print('[{}] IO time:{}'.format(i, iotime-stime))
if i > warmup:
ss = time.time()
with torch.no_grad():
if opt.net == "psmnet" or opt.net == "ganet":
output = net_trt(input_var)
output = output.unsqueeze(1)
elif opt.net == "dispnetc":
output = net_trt(input_var)[0]
else:
output = net_trt(input_var)[-1]
itime = time.time()
print('[{}] Inference time:{}'.format(i, itime-iotime))
if i > warmup:
avg_time.append((time.time() - ss))
if (i - warmup) % display == 0:
print('Average inference time: %f' % np.mean(avg_time))
mbytes = 1024.*1024
print('GPU memory usage memory_allocated: %d MBytes, max_memory_allocated: %d MBytes, memory_cached: %d MBytes, max_memory_cached: %d MBytes, CPU memory usage: %d MBytes' % \
(ct.memory_allocated()/mbytes, ct.max_memory_allocated()/mbytes, ct.memory_cached()/mbytes, ct.max_memory_cached()/mbytes, process.memory_info().rss/mbytes))
avg_time = []
print('[%d] output shape:' % i, output.size())
output = scale_disp(output, (output.size()[0], 540, 960))
disp = output[:, 0, :, :]
ptime = time.time()
print('[{}] Post-processing time:{}'.format(i, ptime-itime))
for j in range(num_of_samples):
name_items = sample_batched['img_names'][0][j].split('/')
# write disparity to file
output_disp = disp[j]
np_disp = disp[j].float().cpu().numpy()
print('Batch[{}]: {}, average disp: {}({}-{}).'.format(i, j, np.mean(np_disp), np.min(np_disp), np.max(np_disp)))
save_name = '_'.join(name_items).replace(".png", "_d.png")# for girl02 dataset
print('Name: {}'.format(save_name))
skimage.io.imsave(os.path.join(result_path, save_name),(np_disp*256).astype('uint16'))
print('Current batch time used:: {}'.format(time.time()-stime))
#save_name = '_'.join(name_items).replace("png", "pfm")# for girl02 dataset
#print('Name: {}'.format(save_name))
#np_disp = np.flip(np_disp, axis=0)
#save_pfm('{}/{}'.format(result_path, save_name), np_disp)
print('Evaluation time used: {}'.format(time.time()-s))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--net', type=str, help='indicate the name of net', default='fadnet', choices=SUPPORT_NETS)
parser.add_argument('--model', type=str, help='model to load', default='best.pth')
parser.add_argument('--filelist', type=str, help='file list', default='FlyingThings3D_release_TEST.list')
parser.add_argument('--filepath', type=str, help='file path', default='./data')
parser.add_argument('--devices', type=str, help='devices', default='0')
parser.add_argument('--display', type=int, help='Num of samples to print', default=10)
parser.add_argument('--rp', type=str, help='result path', default='./result')
parser.add_argument('--flowDiv', type=float, help='flow division', default='1.0')
parser.add_argument('--batchSize', type=int, help='mini batch size', default=1)
opt = parser.parse_args()
detect(opt)