前年年前做一个手机移动端图像识别项目的时候，
先后尝试了mxnet,thnets,caffe,tensorflow.
当时的情况是,mxnet内存管理奇差，内存经常由于模型运算分配不足，app挂掉。
后来调研了下caffe发现也很不友好。
最后发现thnets相对比较轻巧，
经过算法调优之后，性能还不错，
特别是在ios下启用了Accelerate加速库。
后来tensorflow快速发展，就切到tensorflow上了。
最近看了下thnets，作者 mvitez 看来不怎么上心了。
至今为止，改进的不多。
thnets的移动端样例代码，可以参考:
https://github.com/e-lab/apps-iOs
https://github.com/e-lab/apps-android
有一段时间nnpack加速库起来了，就想着把thnets给patch一下nnpack.
但是由于项目太赶，没那个时间去做。
后来也因为切换到tensorflow上了。
thnets就被雪藏了。
向作者提交了两个建议，1，改用stb_image加载图片 2, 支持windows平台
这个两个工作，我都做了。
作者合了1。
2 我关闭了。几天前去看历史记录，作者当时问我关闭的原因,我没回。
真正的原因是。。。thnets被我遗忘了，而windows 版本的存在意义并不大。
今天稍微花了点时间，在windows写个thnets的demo样例，给有需要的网友~
项目地址:
https://github.com/cpuimage/thnets
代码示例见:demo.c
#include <string.h>
#include <stdlib.h>
#include <stdio.h>
#include "thnets.h"
// http://cpuimage.cnblogs.com/
THNETWORK * net; char * labels[] = { "lamp"," bottle"," watch"," pen"," grass"," shoe"," wall"," chair"," mug"," fork"," table"," book"," tablet"," bookcase"," pencil"," door"," face"," ceiling"," sofa"," bicycle"," aluminum - can"," window"," road"," stairs"," floor"," painting"," toy"," remote"," computer"," plant"," television"," dog"," laptop"," microwave"," cat"," tree"," knife"," car"," motorcycle"," person"," cup"," sidewalk"," telephone"," spoon"," hand"," sofabed" }; int main(int argc, char ** argv) {
img_t image = { };
//test.jpg
char * pic_file = argv[];
//model
char * model_path = argv[]; int dropclassifier = ;
loadimage(pic_file, &image);
THInit();
printf("init_thnets.");
net = THLoadNetwork(model_path);
if (net) {
THUseSpatialConvolutionMM(net, );
if (dropclassifier == ) {
if (net->net->modules[net->net->nelem - ].type == MT_SoftMax)
net->net->nelem--;
if (net->net->modules[net->net->nelem - ].type == MT_Linear)
net->net->nelem--;
if (net->net->modules[net->net->nelem - ].type == MT_View)
net->net->nelem--;
}
}
else {
printf("Shiiiiit went down.");
return -;
}
float * percentages = ;
int outwidth, outheight; if (!net) {
return ;
}
int i = ;
int nbatch = ;
unsigned char *bitmaps[];
for (i = ; i < nbatch; i++)
bitmaps[i] = image.bitmap;
int size = THProcessImages(net, bitmaps, nbatch, image.width, image.height, image.cp * image.width, &percentages, &outwidth, &outheight, );
for (i = ; i < nbatch; i++)
if (bitmaps[i])
free(bitmaps[i]);
if (percentages)
{
float max[] = { };
int maxPos[] = { };
for (int j = ; j < size; j++) {
if (percentages[j] > max[]) {
maxPos[] = j;
max[] = percentages[j];
}
else if (percentages[j] > max[]) {
maxPos[] = j;
max[] = percentages[j];
}
else if (percentages[j] > max[]) {
maxPos[] = j;
max[] = percentages[j];
}
}
for (int index = ; index < ; index += ) {
const float predictionValue = max[index];
if (predictionValue > 0.05f) {
printf(" %f %s \n", predictionValue, labels[maxPos[index] % size]);
}
}
free(percentages);
}
getchar();
}
采用vs编译的话，需要下载openblas:
https://jaist.dl.sourceforge.net/project/openblas/v0.2.19/
lib库链接文件：libopenblas.dll.a 即可编译。
附带的模型 来自e-lab的项目。
示例输出得分最高并且高于0.05的三个结果。
示例图片：test.jpg
示例模型: model
对应的标签，见代码或模型文件夹下的categories.txt
以上，待有精力再对thnets进行性能调优。
对于前向传播而言，最好的代码学习资料莫过于：darknet
https://github.com/pjreddie/darknet
darknet 代码写得十分耐看，逻辑清晰。
darknet+nnpack:
https://github.com/digitalbrain79/darknet-nnpack
若有其他相关问题或者需求也可以邮件联系俺探讨。
邮箱地址是:
gaozhihan@vip.qq.com