-
Ubuntu下,通过CMake编译成功,在ncnn-master/build/src下生成了libncnn.a,并能够正确运行squeezenet实例。
现在将ncnn-master/example/squeezencnn/jni下android.mk中路径改为我的路径,如下:
NCNN_INSTALL_PATH := /home/workspace/ncnn/ncnn-master
L…
-
在ubuntu下的android studio中开发调试调用ncnn模型的程序,加载模型的代码为:
`int code;`
`code = net_.load_param(resnet_param.data());`
`__android_log_print(ANDROID_LOG_DEBUG, "resnet_param", "%d %d",resnet_param.size(),code…
-
**Output of 'strings libarm_compute.so | grep arm_compute_version':**
arm_compute_version=v0.0-unreleased Build options: {'arch': 'arm64-v8a', 'debug': '0', 'benchmark': '1', 'benchmark_tests':…
-
i made ncnn with default option and use your benchmark test of squeezenet on 8*A53 CPU in ubuntu 16.04.
i got about 960ms with about 700% CPU usage which means nearly full use of 8 core.
this re…
-
Hi ,
我使用的是caffe模型转为ncnn的模型,使用example里的squeezencnn,可以正常加载模型。
在运行时,也正常输入in Mat,可是extract最后的FC层的top blob确是空的,如果往前面几层是有数据的,请问,这种问题应该如何排查,谢谢。
-
Android平台下
layer {
name: "fc1"
type: "InnerProduct"
bottom: "pool4"
top: "fc1"
param {
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 2
decay_mult: 0
}
inn…
-
只看到在IOS GPU上,squeezenet能跑到30ms。能否提供在android上的性能,这样可以对比跟其他框架的性能。
从代码上来看,ncnn使用neon指令实现了convolution,感觉要比这里直接使用gemm要快一些。
-
用的libncnn.a文件是直接从github上下的ncnn-android-lib/arm64-v8a/libncnn.a
然后和armeabi-v7a一样的方法,进入到工程的jni文件夹下ndk-build。
然后报错:
h3c-macbook:jni H3c$ ndk-build
[armeabi-v7a] SharedLibrary : libsqueezencnn.so
…
-
我想将squeezeNet替换为mobileNet,已经生成好mobileNet.param .bin,接下来应该怎么做呢
ghost updated
6 years ago
-
I created a simple 1 op network with a param file
```
7767517
2 2
Input data 0 1 data 0=277 1=277 2=3
BinaryOp AddOne 1 1 data out 0=0 1=1 2=0.0
```
I used ncnn2mem to cr…