Full Installation :
On Desktop PC having High Power Performace, you should use NCSDK Full Installation.
Full Installation of NCSDK can generate graph file from prototxt and caffemodel.
Full Installation of NCSDK includes API-Library for NCS, and install Caffe and Tensorflow Frameworks and Some pre-trained DNN models like TinyYolo, SSD, Inception etc.
Installation of Frameworks and some pre-trained models depend on other web site contents, so you may fail installation.
Install Ubuntu 16.04 onto VirtualBox over 50GB vdi and add USB-Filter discription as 03e7 and 040e for NCS with VirtualBox Manager.
Booting Ubuntu on VirtualBox.
First step is that upgrade ubuntu.
$ apt update
$ apt upgrade
Download NCSDK.
$ git clone https://github.com/k5iogura/ncsdk_v2
Install ncsdk full
$ cd ncsdk_v2
$ vi ncsdk.conf
// edit INSTALL_TENSORFLOW=No
$ make install ; make install;
$ tail -1 ~/.bashrc
export "PYTHONPATH=${PYTHONPATH}:/opt/movidius/caffe/python"
# pip3 list | grep tensorflow
tensorflow(1.9)
// reinstall 1.5
# pip3 uninstall tensorflow
# pip3 install tensorflow==1.5
$ bash
// check
$ which mvNCCompile
/usr/local/bin/mvNCCompile
You need twice "make install".
You must be into new "bash" to update PYTHONPATH.
You must use "tensorflow"==1.5 because tesnsorflow>1.5 cause Illegal Instruction Error(Seg.V) since having AVX Optimizaion Code.
// Some module will be needed
$ pip3 install graphviz
$ pip3 scikit-image
// and more...
$ make examples
If you're lucky man then no error. If not, but leaving error messages in this phase is OK.
Say hello to NCS device
$ python3 examples/apps/hello_ncs_py/hello_ncs.py
- generate graph file by mvNCCompile tool on Desktop PC.
- python3 script predict using image and graph files show result.
python3 script include steps,
- open NCS devicess
- load graph file and meke graph opject
- load tensor image into graph object
- issue graph.GetResult()
- show result of prediction
$ cd ncsdk_v2/v2.08.01.02/ssd_camera/
$ python3 cam_object_mthread.py
"graph" file is generated from SSD_MobileNet caffemodel for NCSDK_V2. Graph file format of NCSDK_v1 and NCSDK_v2 are not Compatible, so that need to compile model and prototxt by mvNCCompile for NCSDK_v2
NCS recognizes my Ocsillo Scope as "TVmonitor";-)
- asynchronous prediction and loading images.
- Single NCS 3.79FPS
- Video playback 11.26FPS
- actual camera resolution: 640.0 x 480.0
- 300x300 Neural Network input size
Tested sample bellow,
$ cd ncsdk_v2/v2.08.01.02/ssd_video/
$ python3 video_objects_mthread.py
- asynchronous prediction and loading images.
- Single NCS 5.83FPS
- Video playback 17.63FPS
- actual video resolution: 960.0 x 540.0
- 300x300 Neural Network input size
Nov.02.2018