有人试过在arm或者android上用ncnn跑sphereface20吗？

[ctgushiwei]

No description provided.

[Pinnh]

I have run sphereface20-on android with ncnn last year,any problem you met?

[ctgushiwei]

@Pinnh i want to known the time you cost on android

[Ivan6666]

@Pinnh run sphereface on android with ncnn load_model(bin_files.data()) faile ,but run sphereface on ubuntu is ok;

[Pinnh]

sorry, I saw you comment email on phone but forgot to reply.
sphereface20 run android is about 360±20ms
I not met this issue when run sphereface20 on android, android code is not the same as pc , and init params should from asset

pc:
ncnn::Net net;
net.load_param_bin("ncnn.param.bin");
net.load_model("ncnn.bin");

android:
JNIEXPORT jboolean JNICALL Java_com_tencent_ncnn_Ncnn_1Init(JNIEnv* env, jobject thiz, jbyteArray param, jbyteArray bin){
{
int len = env->GetArrayLength(param);
param_vec.resize(len);
env->GetByteArrayRegion(param, 0, len, (jbyte*)param_vec.data());
int ret = net_model.load_param(param_vec.data());
}
{
int len = env->GetArrayLength(bin);
bin_vec.resize(len);
env->GetByteArrayRegion(bin, 0, len, (jbyte*)bin_vec.data());
int ret = net_model.load_model(bin_vec.data());
}
return JNI_TRUE;
}
android java:
private void initNcnn() throws IOException
{
byte[] param = null;
byte[] bin = null;
{
InputStream assetsInputStream = getAssets().open("ncnn.param.bin");
int available = assetsInputStream.available();
param = new byte[available];
int byteCode = assetsInputStream.read(param);
assetsInputStream.close();
}
{
InputStream assetsInputStream = getAssets().open("ncnn.bin");
int available = assetsInputStream.available();
bin = new byte[available];
int byteCode = assetsInputStream.read(bin);
assetsInputStream.close();
}
Ncnn_Init(param, bin);
}

[EngineeringBoy]

兄弟，你搞定了没有，我现在在做这一块，从caffe转成ncnn有需要注意的吗？我feature提取出来的数值不对啊，你有遇到这样的问题吗？

[ctgushiwei]

搞定了，直接调用作者给的那个函数就行了，检查下param是否正确

[EngineeringBoy]

我现在遇到的是这样的问题，用特定的一张图片，在matlab中测试的时候提取出feature，然后我把这个feature保存，然后把网络转换成ncnn，再写c++代码提取出feature，两者对比，不一样，请问你又遇到这样的问题吗？

[Pinnh]

cosine similarity?