转自:https://blog.csdn.net/lxk2017/article/details/88921302
参考:https://blog.csdn.net/l7H9JA4/article/details/104624333
YOLO-v3合并卷积层与BN层
批量归一化-BN层(Batch Normalization)
BN计算公式:
合并卷积层与BN层:
部分代码实现
实验结果
具体代码实现
批量归一化-BN层(Batch Normalization)
随机梯度下降法(SGD)对于训练深度网络简单高效,但是它有个毛病,就是需要我们人为的去选择参数,比如学习率、参数初始化、权重衰减系数、Drop out比例等。这些参数的选择对训练结果至关重要,以至于我们很多时间都浪费在这些的调参上。那么使用BN层之后,你可以不需要那么刻意的慢慢调整参数。(详见论文《Batch Normalization_ Accelerating Deep Network Training by Reducing Internal Covariate Shift》 )。
在神经网络训练网络模型时,BN层能够加速网络收敛,并且能够控制过拟合现象的发生,一般放在卷积层之后,激活层之前。BN层将数据归一化后,能够有效解决梯度消失与梯度爆炸问题。虽然BN层在训练时起到了积极作用,然而,在网络Inference时多了一些层的运算,影响了模型的性能,且占用了更多的内存或者显存空间。因此,有必要将 BN 层的参数合并到卷积层,减少计算来提升模型Inference的速度。
BN计算公式:
在yolo-v3中,BN计算过程如下:
其中x_out为BN计算结果,x_conv为BN前面的卷积计算结果,其余的参数都保存在.weights文件中。
合并卷积层与BN层:
卷积+BN:
此时.weights文件中的参数只剩下权值w和偏置b,合并后的参数要写到新的.weights文件中,注意再次运行Inference代码时,记得将.cfg文件中所有的batch_normalize=1改为batch_normalize=0。
部分代码实现
保存合并后的参数,【文件parser.c中增加代码】
//保存convolutional_weights
void save_convolutional_weights_nobn(layer l, FILE *fp)
{
if(l.binary){
//save_convolutional_weights_binary(l, fp);
//return;
}
#ifdef GPU
if(gpu_index >= 0){
pull_convolutional_layer(l);
}
#endif
int num = l.nweights;
//fwrite(l.biases, sizeof(float), l.n, fp);
/*if (l.batch_normalize){
fwrite(l.scales, sizeof(float), l.n, fp);
fwrite(l.rolling_mean, sizeof(float), l.n, fp);
fwrite(l.rolling_variance, sizeof(float), l.n, fp);
}*/
if (l.batch_normalize) {
for (int j = 0; j < l.n; j++) {
l.biases[j] = l.biases[j] - l.scales[j] * l.rolling_mean[j] / (sqrt(l.rolling_variance[j]) + 0.000001f);
for (int k = 0; k < l.size*l.size*l.c; k++) {
l.weights[j*l.size*l.size*l.c + k] = l.scales[j] * l.weights[j*l.size*l.size*l.c + k] / (sqrt(l.rolling_variance[j]) + 0.000001f);
}
}
}
fwrite(l.biases, sizeof(float), l.n, fp);
fwrite(l.weights, sizeof(float), num, fp);
}
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
Inference时加载更改后的.weights文件:【文件parser.c中增加代码】
void load_convolutional_weights_nobn(layer l, FILE *fp)
{
if(l.binary){
//load_convolutional_weights_binary(l, fp);
//return;
}
if(l.numload) l.n = l.numload;
int num = l.c/l.groups*l.n*l.size*l.size;
fread(l.biases, sizeof(float), l.n, fp);
//fprintf(stderr, "Loading l.biases num:%d,size:%d*%d\n", l.n, l.n, sizeof(float));
fread(l.weights, sizeof(float), num, fp);
//fprintf(stderr, "Loading weights num:%d,size:%d*%d\n", num, num,sizeof(float));
if(l.c == 3) scal_cpu(num, 1./256, l.weights, 1);
if (l.flipped) {
transpose_matrix(l.weights, l.c*l.size*l.size, l.n);
}
if (l.binary) binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.weights);
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
增加配置参数代码:在文件detector.c中
//detector.c
//void run_detector(int argc, char **argv)中增加部分代码
void run_detector(int argc, char **argv)
{
......
if(0==strcmp(argv[2], "test")) test_detector(datacfg, cfg, weights, filename, thresh, hier_thresh, outfile, fullscreen);
else if(0==strcmp(argv[2], "train")) train_detector(datacfg, cfg, weights, gpus, ngpus, clear);
else if(0==strcmp(argv[2], "valid")) validate_detector(datacfg, cfg, weights, outfile);
else if(0==strcmp(argv[2], "valid2")) validate_detector_flip(datacfg, cfg, weights, outfile);
else if(0==strcmp(argv[2], "recall")) validate_detector_recall(cfg, weights);
else if(0==strcmp(argv[2], "demo")) {
list *options = read_data_cfg(datacfg);
int classes = option_find_int(options, "classes", 20);
char *name_list = option_find_str(options, "names", "data/names.list");
char **names = get_labels(name_list);
demo(cfg, weights, thresh, cam_index, filename, names, classes, frame_skip, prefix, avg, hier_thresh, width, height, fps, fullscreen);
}
//add here
else if(0==strcmp(argv[2], "combineBN")) test_detector_comBN(datacfg, cfg, weights, filename, weightname,thresh, hier_thresh, outfile, fullscreen);
}
//增加test_detector_comBN函数
void test_detector_comBN(char *datacfg, char *cfgfile, char *weightfile, char *filename,char *weightname ,float thresh, float hier_thresh, char *outfile, int fullscreen)
{
list *options = read_data_cfg(datacfg);
char *name_list = option_find_str(options, "names", "data/names.list");
char **names = get_labels(name_list);
image **alphabet = load_alphabet();
network *net = load_network(cfgfile, weightfile, 0);
// 定点化保存参数
save_weights_nobn(net, weightname);
}
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
实验结果
自己训练了一个模型,模型结构如下:
执行combineBN 命令:
./darknet detector combineBN cfg/2024.data cfg/2024-test.cfg 2024_140000.weights data/1.jpg save.weights
1
其中.data、.cfg、.weights为自己的参数文件,合并后的权值存储为save.weights。
使用合并后的权值进行Inference:
./darknet detector test cfg/2024.data cfg/2024-test-nobn.cfg save.weights data/1.jpg
1
结果:
实验是在权值和输入参数量化为8bit后进行测试的,平台为笔记本(i7-6700HQ CPU),提升约10.7%。
实验结果:
合并前/ms 合并后/ms
1001 894
具体代码实现
代码在https://github.com/XiaokangLei/darknet-nobn已经把测试用的yolov3-tiny.weights和yolov3-tiny-nobn.cfg放到代码里面了。
【注意!!】记得将.cfg文件中所有的batch_normalize=1改为batch_normalize=0
————————————————
版权声明:本文为CSDN博主「lxk2017」的原创文章,遵循CC 4.0 BY-SA版权协议,转载请附上原文出处链接及本声明。
原文链接:https://blog.csdn.net/lxk2017/article/details/88921302
来源:oschina
链接:https://my.oschina.net/u/4352811/blog/4416680