Neural Compute SDK version 2.04 is now available! The NCAPPZOO has also been updated for NCSDK 2!

[idata]

What’s new in NCSDK 2.04?

Multiple graph allocations on a single NCS device is now possible! This feature allows for users to use multiple graph files from multiple models and queue them up for inference in a sequential manner.

NCSDK 1 can still be found on the master branch of the NCSDK repository, although NCSDK 1 will likely be replaced by NCSDK 2 in the near future. At the moment, NCSDK version 2.xx will be on the ncsdk2 branch of the github repo. To install NCSDK version 2.xx, you can use the following command:

git clone -b ncsdk2 https://github.com/movidius/ncsdk.git

More details and information on the differences between NCSDK 1 and NCSDK 2 can be found at https://movidius.github.io/ncsdk/ncapi/readme.html. Additionally NCSDK version 2.04.00.06 release notes can be found at https://movidius.github.io/ncsdk/release_notes.html.

The NCAPPZOO has also been updated for NCSDK 2!

To view the NCAPPZOO examples for NCSDK 2, please visit https://github.com/movidius/ncappzoo/ and switch to the ncsdk2 branch of the repository. The NCSDK v 1.12 version of the NCAPPZOO will stay in the master branch of the NCAPPZOO.

[idata]

@Tome_at_Intel , does Neural Compute SDK version 2.04 support 16 bit floating point?

I tried to generate graph by mvNCCompile tool from tensorflow model(.pb, file which is 32bit), and I did not find the there is a parameter which could specify 16bit or 32bit floating point. And the output graph works well. And before I used same .pb file to generate graph by NCSDK 1.2, the output is 16bit floating point, and it works as well.So my question is whether 2.04 still supports 16bit float point, maybe according to the input model(.pb) or with specific parameter? Or it only supports 32bit floating point now?Thanks for your answer in advance.

[idata]

@minnowspirit Yes the NCSDK does support fp16. But now it takes fp32 and does the conversion internally now. You can read more about it at https://movidius.github.io/ncsdk/ncapi/python_api_migration.html. Scroll down to the input/output data types section.

[idata]

Got it, thank you @Tome_at_Intel

[idata]

I have install the NCSDK 1. , how could I upgrade to the NCSDK 2. I use the uninstall commant to uninstall the NCSDK1 and then install the NCSDK2. But I meet error and I don't know how to fix it. Could you pls give a help .Thanks！@Tome_at_Intel

[idata]

@ideallyworld What is the error you are seeing?

[idata]

The log info is just as follow（install error 131）：

root@ubuntu:~/workspace# cd ncsdk

root@ubuntu:~/workspace/ncsdk# lsapi examples install.sh Makefile requirements_apt_raspbian.txt uninstall-opencv.shATTRIBUTIONS extras install-utilities.sh ncsdk.conf requirements_apt.txt uninstall.shdocs install-opencv.sh LICENSE README.md requirements.txt version.txtroot@ubuntu:~/workspace/ncsdk# make install./uninstall-opencv.shOpenCV uninstall startingReading installer configuration variables from ./ncsdk.confYou are using pip version 8.1.1, however version 10.0.1 is available.You should consider upgrading via the 'pip install --upgrade pip' command.OpenCV uninstall finished

make install starting.

./install.shMovidius Neural Compute Toolkit InstallationInstaller NCSDK version: 2.04.00.06

Reading installer configuration variables from ./ncsdk.conf

Installing on Ubuntu 16.04

print_ncsdk_config()

INSTALL_DIR - Main installation directoryINSTALL_CAFFE - Flag to install TensorFlowCAFFE_FLAVOR - Specific 'flavor' of caffe to installCAFFE_USE_CUDA - Use CUDA enabled version of caffeINSTALL_TENSORFLOW - Flag to install TensorFlowINSTALL_TOOLKIT - Flag to install Neural Compute ToolkitPIP_SYSTEM_INSTALL - Globally install pip packages via sudo -HVERBOSE - Flag to enable more verbose installationUSE_VIRTUALENV - Flag to enable python virtualenvMAKE_NJOBS - Number of processes to use for parallel build (i.e. make -j MAKE_NJOBS)

INSTALL_DIR=/opt/movidius

INSTALL_CAFFE=yesCAFFE_FLAVOR=ssdCAFFE_USE_CUDA=noINSTALL_TENSORFLOW=yesINSTALL_TOOLKIT=yesPIP_SYSTEM_INSTALL=yesVERBOSE=yesUSE_VIRTUALENV=noMAKE_NJOBS=2

Didn't find a valid NCSDK installation

Searching and removing previous NCSDK installation from the systemSearching and removing NCS python API…You are using pip version 8.1.1, however version 10.0.1 is available.You should consider upgrading via the 'pip install --upgrade pip' command.Uninstalling mvnc-1.12.0.1:Successfully uninstalled mvnc-1.12.0.1You are using pip version 8.1.1, however version 10.0.1 is available.You should consider upgrading via the 'pip install --upgrade pip' command.Uninstalling mvnc-1.12.0.1:Successfully uninstalled mvnc-1.12.0.1doneSearching and removing udev rules…doneRunning ldconfig…doneUpdating udev rules…doneSuccessfully uninstalled NCSDK from the systemDownloading NCSDK-2.04.00.06.tar.gz--2018-06-12 21:11:38-- https://downloadmirror.intel.com/27738/eng/NCSDK-2.04.00.06.tar.gzResolving downloadmirror.intel.com (downloadmirror.intel.com)… 23.48.22.30Connecting to downloadmirror.intel.com (downloadmirror.intel.com)|23.48.22.30|:443… connected.HTTP request sent, awaiting response… 200 OKLength: 1383386 (1.3M) [application/x-gzip]Saving to: ‘NCSDK-2.04.00.06.tar.gz’

。。。。。。

http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main DEP-11 64x64 Icons

Err:9 http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main Sources404 Not FoundIgn:12 http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main amd64 PackagesIgn:15 http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main i386 PackagesIgn:16 http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main all PackagesIgn:17 http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main Translation-en_USIgn:18 http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main Translation-enIgn:19 http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main amd64 DEP-11 MetadataIgn:22 http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial/main DEP-11 64x64 IconsFetched 4,610 kB in 46s (99.7 kB/s)Reading package lists…W: The repository 'http://ppa.launchpad.net/fcitx-team/nightly/ubuntu xenial Release' does not have a Release file.E: Failed to fetch http://ppa.launchpad.net/fcitx-team/nightly/ubuntu/dists/xenial/main/source/Sources 404 Not FoundE: Some index files failed to download. They have been ignored, or old ones used instead.Installation failed: Command ' apt-get update' return code=100. Error on line 308 in ./install-utilities.sh. Will exit Makefile:47: recipe for target 'install' failedmake: *** [install] Error 131@Tome_at_Intel

[idata]

Facing the exact same issue with installation of NCSDK 2 as mentioned by @ideallyworld in the above post.

Could someone please help in resolving this issue?

Thanks in advance.

[idata]

@ideallyworld , @abhi

I do not know if it will work or not.Various packages seem to be old.$ sudo apt update $ sudo apt upgrade $ cd <new ncsdk downloaded path> $ make install

[idata]

I need some help, since I cannot find the repository for NCSDK 2

http://github.com/Movidius/ncsdk

[idata]

https://ncsforum.movidius.com/discussion/1053/ncsdk-on-github-is-going-opensource-current-software-releases-moved-temporarily#latest

[idata]

Is there an example about how to load multi model in on stick device?Such as MTCNN,there are three model for three stages,how can this be done in version 2.0?

[idata]

Is there an example about how to load multi model in on stick device?Such as MTCNN,there are three model for three stages,how can this be done in version 2.0?

[idata]

Is there an example about how to load multi model in on stick device?Such as MTCNN,there are three model for three stages,how can this be done in version 2.0?

[idata]

@shartoo

I do not know if it will be helpful, but it is the repository I created.I am implementing Multi Graph.Load two or three graphs simultaneously.Because it is logic based on Multi Stick, the part of "Try ~ Except ~" may be a bit confusing.https://github.com/PINTO0309/MobileNet-SSD-RealSense.gitMultiStickSSDwithRealSense.py # 1:= Enabled MobileNet-SSD Model if ssd_detection_mode == 1: with open(join(graph_folder, "graph"), mode="rb") as f: graph_buffers.append(f.read()) graphs.append(mvnc.Graph('MobileNet-SSD')) # 1:= Enabled Fullweight FaceDetection Model if face_detection_mode == 1: with open(join(graph_folder, "graph.fullfacedetection"), mode="rb") as f: graph_buffers.append(f.read()) graphs.append(mvnc.Graph('FullFaceDetection')) # 2:= Enabled Lightweight FaceDetection Model if face_detection_mode == 2: with open(join(graph_folder, "graph.shortfacedetection"), mode="rb") as f: graph_buffers.append(f.read()) graphs.append(mvnc.Graph('ShortFaceDetection')) devopen = False for device in devices: try: device = mvnc.Device(device) device.open() for (graph, graph_buffer) in zip(graphs, graph_buffers): graphHandles.append(graph.allocate_with_fifos(device, graph_buffer)) devopen = True break except: continue

By the way, MTCNN has multiple output problems, it is impossible to smartly implement with NCSDK v2.

[idata]

Hi @PINTO i have tried loading three model of MTCNN's three stage in one stick but failed.Here is what i did:

def load_graph(graph_file1,graph_file2,graph_file3): #mvnc.SetGlobalOption(mvnc.GlobalOption.LOG_LEVEL, 0) devices = mvnc2.enumerate_devices() if len(devices) == 0: print('No devices found') return 1 device = mvnc2.Device(devices[0]) device.open() # Load graph from disk and allocate graph via API with open(graph_file1, mode='rb') as f: graph_from_disk1 = f.read() graph1 = mvnc2.Graph("graph1") fifo_in1, fifo_out1 = graph1.allocate_with_fifos(device, graph_from_disk1) with open(graph_file2, mode='rb') as f: graph_from_disk2 = f.read() graph2 = mvnc2.Graph("graph2") fifo_in2, fifo_out2 = graph2.allocate_with_fifos(device, graph_from_disk2) with open(graph_file3, mode='rb') as f: graph_from_disk3 = f.read() graph3 = mvnc2.Graph("graph3") fifo_in3, fifo_out3 = graph3.allocate_with_fifos(device, graph_from_disk3) return graph1,fifo_in1, fifo_out1,graph2,fifo_in2, fifo_out2,graph3,fifo_in3, fifo_out3

error is below

... I1102 10:19:38.878919 3190 net.cpp:157] Top shape: 1 2 (2) I1102 10:19:38.878955 3190 net.cpp:165] Memory required for data: 940680 I1102 10:19:38.879007 3190 layer_factory.hpp:77] Creating layer conv6-2 I1102 10:19:38.879052 3190 net.cpp:100] Creating Layer conv6-2 I1102 10:19:38.879079 3190 net.cpp:434] conv6-2 <- conv5_prelu5_0_split_1 I1102 10:19:38.879118 3190 net.cpp:408] conv6-2 -> conv6-2 I1102 10:19:38.879233 3190 net.cpp:150] Setting up conv6-2 I1102 10:19:38.879261 3190 net.cpp:157] Top shape: 1 4 (4) I1102 10:19:38.879293 3190 net.cpp:165] Memory required for data: 940696 I1102 10:19:38.879334 3190 layer_factory.hpp:77] Creating layer conv6-3 I1102 10:19:38.879374 3190 net.cpp:100] Creating Layer conv6-3 I1102 10:19:38.879400 3190 net.cpp:434] conv6-3 <- conv5_prelu5_0_split_2 I1102 10:19:38.879444 3190 net.cpp:408] conv6-3 -> conv6-3 I1102 10:19:38.879634 3190 net.cpp:150] Setting up conv6-3 I1102 10:19:38.879663 3190 net.cpp:157] Top shape: 1 10 (10) I1102 10:19:38.879696 3190 net.cpp:165] Memory required for data: 940736 I1102 10:19:38.879737 3190 layer_factory.hpp:77] Creating layer prob1 I1102 10:19:38.879772 3190 net.cpp:100] Creating Layer prob1 I1102 10:19:38.879798 3190 net.cpp:434] prob1 <- conv6-1 I1102 10:19:38.879837 3190 net.cpp:408] prob1 -> prob1 I1102 10:19:38.879884 3190 net.cpp:150] Setting up prob1 I1102 10:19:38.879909 3190 net.cpp:157] Top shape: 1 2 (2) I1102 10:19:38.879940 3190 net.cpp:165] Memory required for data: 940744 I1102 10:19:38.879966 3190 net.cpp:228] prob1 does not need backward computation. I1102 10:19:38.879994 3190 net.cpp:228] conv6-3 does not need backward computation. I1102 10:19:38.880022 3190 net.cpp:228] conv6-2 does not need backward computation. I1102 10:19:38.880049 3190 net.cpp:228] conv6-1 does not need backward computation. I1102 10:19:38.880079 3190 net.cpp:228] conv5_prelu5_0_split does not need backward computation. I1102 10:19:38.880105 3190 net.cpp:228] prelu5 does not need backward computation. I1102 10:19:38.880131 3190 net.cpp:228] drop5 does not need backward computation. I1102 10:19:38.880157 3190 net.cpp:228] conv5 does not need backward computation. I1102 10:19:38.880185 3190 net.cpp:228] prelu4 does not need backward computation. I1102 10:19:38.880211 3190 net.cpp:228] conv4 does not need backward computation. I1102 10:19:38.880239 3190 net.cpp:228] pool3 does not need backward computation. I1102 10:19:38.880266 3190 net.cpp:228] prelu3 does not need backward computation. I1102 10:19:38.880292 3190 net.cpp:228] conv3 does not need backward computation. I1102 10:19:38.880321 3190 net.cpp:228] pool2 does not need backward computation. I1102 10:19:38.880347 3190 net.cpp:228] prelu2 does not need backward computation. I1102 10:19:38.880374 3190 net.cpp:228] conv2 does not need backward computation. I1102 10:19:38.880401 3190 net.cpp:228] pool1 does not need backward computation. I1102 10:19:38.880429 3190 net.cpp:228] prelu1 does not need backward computation. I1102 10:19:38.880455 3190 net.cpp:228] conv1 does not need backward computation. I1102 10:19:38.880482 3190 net.cpp:228] input does not need backward computation. I1102 10:19:38.880508 3190 net.cpp:270] This network produces output conv6-2 I1102 10:19:38.880535 3190 net.cpp:270] This network produces output conv6-3 I1102 10:19:38.880564 3190 net.cpp:270] This network produces output prob1 I1102 10:19:38.880630 3190 net.cpp:283] Network initialization done. I1102 10:19:38.886859 3190 net.cpp:761] Ignoring source layer data48 I1102 10:19:38.886955 3190 net.cpp:761] Ignoring source layer slicer_label I1102 10:19:38.886981 3190 net.cpp:761] Ignoring source layer label1_slicer_label_0_split I1102 10:19:38.888417 3190 net.cpp:761] Ignoring source layer conv6-1_conv6-1_0_split I1102 10:19:38.888537 3190 net.cpp:761] Ignoring source layer loss1 I1102 10:19:38.888568 3190 net.cpp:761] Ignoring source layer accuracy1 I1102 10:19:38.888595 3190 net.cpp:761] Ignoring source layer loss2 I1102 10:19:38.888620 3190 net.cpp:761] Ignoring source layer loss3 E: [ 0] XLinkOpenStream:1035 Max streamId reached deaddead! W: [ 0] ncGraphAllocate:1008 can't open stream for graphBuffer transmission Traceback (most recent call last): File "/home/pi/workspace/test/mtcnn/mtcnn_caffe/demo/test.py", line 223, in <module> main() File "/home/pi/workspace/test/mtcnn/mtcnn_caffe/demo/test.py", line 207, in main graph1, fifo_in1, fifo_out1, graph2, fifo_in2, fifo_out2, graph3, fifo_in3, fifo_out3 = load_graph(graph_file1, graph_file2, graph_file3) File "/home/pi/workspace/test/mtcnn/mtcnn_caffe/demo/test.py", line 197, in load_graph fifo_in3, fifo_out3 = graph3.allocate_with_fifos(device, graph_from_disk3) File "/usr/local/lib/python3.5/dist-packages/mvnc/mvncapi.py", line 613, in allocate_with_fifos raise Exception(Status(status)) Exception: Status.ERROR

It seems that load two graph is ok,while fail when loading a third graph.

[idata]

There is a caffe version of MTCNN here caffe-mtcnn github,what i did above was cloned the caffe-mtcnn and compiled the caffe model using nvccCompile of ncsdk for using accelerator of ncsdk

[idata]

@shartoo

I used the repository you suggested and tried to implement it in cooperation with other engineers, but it did not succeed.https://ncsforum.movidius.com/discussion/comment/3391/https://ncsforum.movidius.com/discussion/comment/3441/https://ncsforum.movidius.com/discussion/comment/3729/https://ncsforum.movidius.com/discussion/comment/3815/

Please show below.

Unfortunately, Intel's engineers say it is a bug in NCSDK.https://ncsforum.movidius.com/discussion/1259/cant-load-more-than-two-models#latest

I think your question is not appropriate for this topic.

(Because it is hard to touch the eyes of other expert engineers.)I think it would be better to create a new topic yourself.

[idata]

@PINTO Thank you for your advice .Now i have to quit this solution ,the nvccCompile seems work for only few model provide by intel official