UPDATE: Moved to, https://github.com/quickgrid/ai-resources/blob/master/setup-guide.md, to reduce fragmentation.
A build and setup guide for various libraries, tools, cloud usage etc.
Download the installer or zip version of cmake. Dlib 19.19 from official website had bug with opencv cv_image which is fixed on github source.
- CMake, https://cmake.org/download/
- Visual Studio 2019 Community Edition
- Vulkan SDK, https://vulkan.lunarg.com/sdk/home
- Cuda and cuDNN with
CUDA_PATHenvironment variable set - Dlib source code from github, https://github.com/davisking/dlib
- Anaconda, https://www.anaconda.com/products/individual
- NCNN source, https://github.com/Tencent/ncnn
- OpenCV source, https://github.com/opencv/opencv
- OpenCV contrib source, https://github.com/opencv/opencv_contrib
These are some of the projects I have tested. Most or all of these can be implemented in both pc and android.
| Name | Link |
|---|---|
| Very fast and accurate rotation-invariant face detector | https://github.com/MagicJackStone/FaceKit/tree/master/PCN |
| Ultra light face detector OpenCV DNN | https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/opencv_dnn |
| Ultra light face detector NCNN | https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB/tree/master/ncnn/src |
| Object Tracking | https://www.learnopencv.com/object-tracking-using-opencv-cpp-python/ |
| Facial Landmarks with Dlib | http://dlib.net/face_landmark_detection_ex.cpp.html |
| Facial Landmarks with OpenCV Contrib | https://docs.opencv.org/master/de/d27/tutorial_table_of_content_face.html https://github.com/kurnianggoro/GSOC2017 |
| YoloFace face detection | https://github.com/sthanhng/yoloface |
| OpenCV DNN Object Detection and Localizaiton | Pretrained darknet yolo (cfg, weights), Caffe (prototxt, caffemodel) etc. |
Note in order to make latest version of all three working in same project the easiest way for beginner is to setup opencv contrib/opencv, next setup dlib after changing target project name, cmake, ndk paths, finally setup ncnn.
NOTE: Must create C++ android project to make these all work together.
Build instructions for opencv contrib,
https://medium.com/beesightsoft/build-opencv-opencv-contrib-for-android-on-windows-9894b4fe6386
But easiest way at the moment is download prebuilt version of latest opencv contrib from,
https://pullrequest.opencv.org/buildbot/export/opencv_releases/master-contrib_pack-contrib-android/
Use the following link and template to setup native opencv project for android. Instead of using opencv android from their site just replace it with opencv contrib from the link above.
https://github.com/VlSomers/native-opencv-android-template
https://github.com/ivangrov/Android-Deep-Learning-with-OpenCV
Download dlib from github repo and extract to C: drive such that extracted path is C:\dlib\dlib-master\.
Use this dlib powershell script to copy dlib directly into the project. Must change the absolute paths before running the script. Change $CMAKE_BIN_PATH, $NDK, $PROJECT_PATH, $DLIB_PATH.
Paste the script in C:\dlib\ and run powershell as administrator. Paste set-location C:\dlib\dlib-master\ and press enter to go into that folder. Run Set-ExecutionPolicy RemoteSigned and press A.
Next, paste & "C:\dlib\dlib-android-setup.ps1" and enter to start compiling and copying files to your project. Must change paths accordingly in script before execution.
After all dlib files are copied to android project, paste Set-ExecutionPolicy Restricted to powershell and press A.
Further detail on script and execution procedure,
https://stackoverflow.com/questions/60548479/setting-up-dlib-for-android-studio/60550358#60550358
Original script source,
https://github.com/Luca96/dlib-for-android
https://github.com/Tencent/ncnn/wiki/how-to-build
Easiest option for android is download prebuilt releases,
https://github.com/Tencent/ncnn/releases
Download vulkan library for GPU support.
Download opencv, opencv_contrib, cmake. Extract and paste both into a folder for path, ex: opencv-compilation. Use cmake-gui browse source to point to opencv folder and click configure. Find OPENCV_EXTRA_MODULES_PATH and point to opencv_contrib modules folder. Point PYTHON3_ variables to correct conda environment path. Untick BUILD_SHARED_LIBS, tick BUILD_opencv_world, click configure and generate. It may be necessary to have numpy installed beforehand. Unchecking unncessary modules will speed up process if python is the only target.
This will generate build folder inside opencv-compilation. Open OpenCV.sln, choose release configuration x64. In solution explorer open CMakeTargets, right click on ALL_BUILD and click build. After compilation done copy cv2.****.pyd file from opencv-compilation\build\lib\python3\Release to Anaconda or Miniconda custom enviroment. For examples, I used miniconda3 and the copy location for pyd file was C:\Users\...\miniconda3\envs\mycustomenv\Lib\site-packages.
Check installation by activating correct environment and running,
import cv2
print(cv2.__version__)
For Visual Studio 2019 Use opencv_worldXXX.lib, Ex: opencv_world430.lib for release configuration and opencv_world430d.lib for debug configuration. If opencv is extracted to C: then it can be fould in C:\opencv\build\x64\vc15\lib.
Depending on CPU cores CUDA compilation can be very slow. OpenCV CUDA build,
OpenCV python CUDA static build,
Compile instructions,
For use cmake to configure and Visual Studio compile ALL_BUILD dlib in release x64 to generate dlib_(...).lib.
In desired project set vc++ include directories, include path to dlib master downloaded from master to extracted folder. Set library directoies to generated dlib.lib. Copy dlib/all/source.cpp to visual studio source files folder where other c++ codes reside.
To enable jpg support set C/C++ > General > Additional Include Directories, to extracted dlib folder something like, C:\dlib-master\dlib\external\libjpeg. Set C/C++ > Preprocessor > Preprocessor definitions, to CRT_SECURE_NO_WARNINGS and DLIB_JPEG_SUPPORT for jpg support.
Example for testing, http://dlib.net/dnn_face_recognition_ex.cpp.html
https://github.com/Tencent/ncnn/wiki/how-to-build
Ninja Windows Release, https://github.com/ninja-build/ninja/releases
LLVM Download, https://releases.llvm.org/download.html
Make a build folder to store generated files. In cmake set this path. Also set path for source C++ folder. Click configure and select generator. Choose Ninja and tick specify native compilers. Set LLVM bin clang.exe and clang++.exe for C and C++.
There will be error that CMAKE_MAKE_PROGRAM-NOTFOUND. Browse to directory where ninja release was extracted and select ninja.exe. It should also be possible (not tested) to use ninja.exe that comes with android studio setup in android cmake folder.
Now press configure and again configure if all looks correct. Next, click generate. Open cmd navigate to project source folder and run ninja.exe. If not is system path then specify location as, "C:\ninja-win\ninja.exe" on command prompt.
cmake_minimum_required(VERSION 3.12)
project(hello_world)
add_executable(${PROJECT_NAME} main.cpp)
#include <iostream>
using namespace std;
int main(){
cout << "HELLO" << "\n";
cin.get();
return 0;
}Instructions for setting up c++ build and debugging on windows. For MinGW gcc, g++, gdb etc. I used Codeblocks MinGW version (may cause problem with std::thread) where in bin folder they are available.
Sample path is, {base_path}\codeblocks\MinGW\bin. Need to have Microsoft C/C++ extensions installed in vscodium.
https://code.visualstudio.com/docs/cpp/config-mingw
To create workspace add vscodium to path and use codium .. Create desired cpp file in the workspace. Next, choose Terminal > Configure Tasks > Shell:g++.exe build active file. This will create a build.json file and can be run from Terminal > Run Build Task... Next, from Run > Add Configurations choose C++ (GDB/LLDB) > g++.exe build and debug active file.
Now, set breakpoints and use Run > Start Debugging. If there is any error with preLaunchTask in launch.json, then matching with task.json label should fix it.
This method with cl.exe works for build, but debugging does not work.
https://code.visualstudio.com/docs/cpp/config-msvc
This setup is quite complex.
First read these carefully,
Ensure CUDA, cuDNN is setup in system properly. In my case CUDA 10.2 and cuDNN 7.6.5 for 10.2 x64 works.
Download the source from github as ZIP, https://github.com/CMU-Perceptual-Computing-Lab/openpose. Maybe git recursive clone is better. Extract it a folder. Ex: C:\openpose-master.
If 3rd party dependencies are not in extracted folder. Then download those project zip separately and extract to approprite folder. Ex: pybind11, caffe.
In openpose-master\3rdparty\windows use the bat files to download dependencies. If download is slow, open the batch files and manually download each zip files with given url and extract to folders caffe, caffe3rdparty, opencv, freeglut. For manual downloads the downloaded files may have extra characters added to them, so change their name as above.
Create a new folder inside openpose-master named build. Point to source folder(C:\openpose-master) and build(C:\openpose-master\build) folder in CMake GUI. Press configure and choose visual studio 2019 as generator and platform as x64 then press finish. Wait for cmake to configure. CMake should automatically find everything. Next, press configure again and press generate. It should not give any errors.
If there is any error, it has to with path to caffe, caffe3rdparty, opencv, freeglut, or they were not downloaded properly. If they are in custom path set each or their value until generate button no longer gives any error.
Open build folder now and open the *.sln file with visual studio. Change configuration to Release and in solution explorer run ALL_BUILD target or run Local Windows Debugger directly.
In order to run openpose samples create a new c++ empty project in VS2019. Set configuration to Release. Add desired sample cpp code to run from, https://github.com/CMU-Perceptual-Computing-Lab/openpose/tree/master/examples/tutorial_api_cpp to Source Files folder in project.
Next step in to add *.dll, *.lib files and necessary include folders via project Properties. In VC++ include directories, set path to all include folder in openpose-master source folders. Ex:
C:\openpose-master\include
C:\openpose-master\3rdparty\windows\opencv\include
Do this for all include folder in 3rdparty folder. Next, set Library directories,
C:\openpose-master\build\src\openpose\Release
C:\openpose-master\3rdparty\windows\opencv\x64\vc15\lib
Do same for all lib folders in 3rdparty directory. Next, write name of each *.lib files in Linker > Input > Additional Dependencies such as,
freeglut.lib
glog.lib
gflags.lib
caffe.lib
caffeproto.lib
opencv_world420.lib
openpose.lib
(Optional) Use the d suffix versions for debug config if debugging config is set, otherwise use above for release.
Copy models folder from openpose source to VS2019 solution project folder. It will have caffemodels and prototxt but caffemodels are placeholder. If batch was used to run then it should have everything ready. Otherwise paste the downloaded caffe models here for different tasks which were downloaded with *.bat links manually.
In the demo cpp code set path to image file correctly before running. Run the code and it will give error that *.dll not found. In Visual Studio solution x64 folder in explorer where *.exe is found, paste the dll files from here,
C:\openpose-master\build\bin
Also paste, openpose.dll from build folder. Search for it not found. Now each sample codes should run directly with image/video path change.
https://github.com/quickgrid/Build-Guide/blob/master/openpose_face_keypoints_sample.jpg
TightVNC Viewer, https://www.tightvnc.com/download.php
RealVNC Viewer (Alternative to TightVNC), https://www.realvnc.com/en/
Putty, https://www.putty.org/
WinSCP , https://winscp.net/eng/index.php
Use Shift + Insert key to paste copied text from windows clipboard to Putty ubuntu terminal vim. Use Shift + : for inserting commands, :q! for exiting without saving. Use AWS Public IPv4 DNS in Putty hostname. Use puttygen to convert pem file to ppk to be used by putty, click save private key. In putty SSH > Auth section browse to ppk file and click open to connect.
Use ubuntu as username for default ubuntu image. Instructions for installing and connecting to GUI is available here.
https://ubuntu.com/tutorials/ubuntu-desktop-aws
It will also work for windows host. Paste all the required commands in putty terminal. If all went correctly use the IPv4 DNS address appended with :1 to connect to ubuntu GUI from tightVNC. Make sure to set up inbound rule for VNC access.
Use df -BM for checking free space.
Run these commands one by one on putty terminal. Tested on ubuntu 18.04 ec2 instance. Full instructions here, https://phoenixnap.com/kb/install-docker-on-ubuntu-20-04.
sudo apt update
sudo apt-get install apt-transport-https ca-certificates curl software-properties-common
curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo apt-key add -
sudo add-apt-repository "deb [arch=amd64] https://download.docker.com/linux/ubuntu $(lsb_release -cs) stable"
sudo apt update
sudo apt-get install docker-ce
docker --version
sudo systemctl start docker
sudo systemctl enable docker
sudo systemctl status docker
Now, docker is ready for android tensorflow lite build.
Build instructions here, https://www.tensorflow.org/lite/guide/build_android.
To fix issue with Got permission denied while trying to connect to the docker daemon socket at unix:///var/run/docker.sock: ..... use the command sudo chmod 666 /var/run/docker.sock.
Within docker type ls to get folder list and navigate to tensorflow_src with cd tensorflow_src. Here, type ./configure to configure tensorflow build.
Enter an exited docker container with, docker start 7263550e2520 && docker attach 7263550e2520. To start the last one use, docker start `docker ps -q -l` && docker attach `docker ps -q -l` .
To get list of stopped containers with information use, docker ps -a.
Use WinSCP to download files from aws ec2 to local system. Download the android tensorflow aar files with this to put into android project.
Use ppk generated with puttygen in WinSCP advanced authentication. Log in to server with user name.
WinSCP AWS EC2 instructions, https://winscp.net/eng/docs/guide_amazon_ec2.
Use screen to detach and run commands. Install with,
apt-get install screen
Enter screen on terminal and ctrl + a + d to detach. Use screen -ls for a screen list and screen -r to resume single screen. For multiple screens use screen -r <number> to attach.
Source, https://www.interserver.net/tips/kb/using-screen-to-attach-and-detach-console-sessions/
https://github.com/tensorflow/tensorflow/blob/master/.bazelrc
https://docs.bazel.build/versions/master/user-manual.html
https://www.tensorflow.org/lite/guide/ops_select
https://www.tensorflow.org/lite/guide/reduce_binary_size
https://www.tensorflow.org/lite/guide/build_android
Source code link for the part, https://github.com/AlexeyAB/darknet/blob/master/src/image.c
Here, as of current date the code is as below on the draw_detections_v3 function. Commenting out these two lines will prevent accuracy string from being drawn on predictions.jpg image. Similarly class output or both can also be removed but why do so?
sprintf(prob_str, ": %.2f", selected_detections[i].det.prob[selected_detections[i].best_class]);
strcat(labelstr, prob_str);
This should be done before doing make. After make command there will be a file named darknet within darknet folder.
For GPU support use a gpu instance for example, ml.p3.2xlarge, is 16gb Nvidia V100 gpu. If vpc not required then select no vpc. Start the instance and run Open JupyterLab. The main reason for writing this is there is an error with make when makefile contains OPENCV=1. A working but poor solution is set OPENCV=0 before make, but this will slow down training and inference as pointed out in output log. For kernel conda_python3 will work.
TODO: Use yum to install opencv on system to see if opencv problem is solved.
Easiest option in windows is to download libraqm from here, https://www.lfd.uci.edu/~gohlke/pythonlibs/#pillow. I have tested on miniconda python 3.7 custom environment. To use this put *.dll from above link in same folder as python.exe in miniconda/anaconda envs/MY_ENV_NAME folder. If the code below gives True then it will render Bangla fonts properly.
from PIL import features
print(features.check("raqm"))
Windows install instructions,
Based on answer from,
https://stackoverflow.com/questions/66184573/how-do-i-install-libraqm-library-in-google-colab.
This script contains example in google colab,
Black and white image can be converted to mesh by first using Inkscape to trace the image bitmap, then saving as plain svg. Which can be imported into blender and conveted to mesh. Blender 2.92 contain Grease Pencil tracing which works for image masks too. Drag and drop image in blender then using Object > Trace Image to Grease Pencil will convert to grease pencil strokes which can converted to path, curve, mesh. The mesh can be extruded to create 3D text object.
Further details here,
I am not sure if internet on router is necessary to setup the camera with router first time. It can be used without internet on the router after setup. Admin password on router should be 16 characters or less. There is a problem of auto truncating a longer password on certain router models which can be missed easily causing login problem.
If wired internet is not available mobile hospot can be used to get internet in the router. If router has WDS bridge then it can be used to connect to mobile hotspot on 2.4GHz or 5GHz with password. This will allow other devices connected to router to get internet access of the mobile hotspot.
This camera has two setup process wifi or ethernet. Setup camera through Tapo mobile app by connecting to camera first and connecting to router after camera is connected to it. When using wifi camera will mention connected to wifi through loud speaker.
rtsp streaming can be done via methods below for high and low resolution feed,
rtsp://username:password@camera_ip:554/stream1
rtsp://username:password@camera_ip:554/stream2
This can used to stream via VLC or OpenCV video capture to run deep learning models. Username and password needs to be setup via tapo app for above to work.
- Try INT8, FP16 quantization.
- Use ONNX runtime, OpenCV DNN, TFLite.
- Use model compression, prunning.
- Run drawing, other operations on separate threads.
- In OpenCV use threaded non blocking webcam, file video capture.
- Detect in a frame and in next N frames run tracking only then again detect. Frame dropping and skipping.
- Empty frame buffer/queue to always get the latest frame.
- Do cpu specific compilation of code take advantage of cpu architectures such as AVX2, AVX512 etc.
- Compile with oneDNN, oneMKL, CUDA, oneTBB, gstreamer etc.
- Use vectorized implementations instead of naive approaches. Try Numba jit for cpu/gpu, CuPy.
- Use C++ instead of python for whole code.
- Move slower part of python code to C++ and call as library.
- Use a smaller input resolution for the model at the cost of accuracy.
- Convert model to onnx, openvino optimized format.
- Use better data structures and algorithms, tweak code to speedup.
First follow this video by Dr. Jeff Heaton heaton to get started with the process.
https://www.youtube.com/watch?v=WNfFD1MSj44
Various AMI images available in aws marketplace one of which is NVIDIA Deep Learning AMI. Other interesting ones are AWS Deep Learning AMI (Ubuntu 18.04) and AWS Deep Learning Base AMI (Ubuntu 18.04). These seem to have pytorch, tensorflow setup by default. Linux version may be newer in future. Links to these AMI,
https://aws.amazon.com/marketplace/pp/prodview-e7zxdqduz4cbs
https://aws.amazon.com/marketplace/pp/prodview-dxk3xpeg6znhm
https://aws.amazon.com/marketplace/pp/prodview-x5nivojpquy6y
AWS marketplace AMI with aws, nvidia deep learning containers,
Nvidia NGC website has various containers such Tensorflow, Pytorch. For pytorch, tensorflow the docker commands available here,
https://ngc.nvidia.com/catalog/containers/nvidia:pytorch
https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow
https://ngc.nvidia.com/catalog
Release notes on AWS and Nvidia AMI,
https://docs.aws.amazon.com/dlami/latest/devguide/appendix-ami-release-notes.html
https://docs.nvidia.com/ngc/ngc-ami-release-notes/
More information on Deep Learning AMI and EC2 P3 instance,
https://docs.aws.amazon.com/dlami/latest/devguide/gs.html
https://docs.aws.amazon.com/dlami/latest/devguide/overview-conda.html
https://aws.amazon.com/ec2/instance-types/p3/
To setup python container first setup tunneling port, ppk, ip via putty. These commands can be found on NGC containers link above.
In putty hostname set ip address of the instance. Convert *.pem to *.ppk via puttygen with passphrase. Set *.ppk to connection > SSH > Auth > Private key file for authentication. Set connection > SSH > Tunnels source port to 8888, destination to localhost:8888 and click add. Save session settings with a name.
From putty terminal first run,
docker pull nvcr.io/nvidia/pytorch:21.05-py3
Afterward run below. The pytorch container version may be different and should be adjusted in both places. Mounting, container directory setup is also available in above NGC container link which I did not try here.
docker run --gpus all -it --rm -p 8888:8888 nvcr.io/nvidia/pytorch:21.05-py3
Following, https://stackoverflow.com/a/51851298/1689698, using the command for jupyter notebook gives the token which is used run jupyter notebook.
jupyter notebook --ip 0.0.0.0 --port 8888 --no-browser --allow-root
This will give somethis like, http://hostname:8888/?token=<ALPHANUMERTIC_TOKEN>. In the browser running, http://localhost:8888/?token=<ALPHANUMERTIC_TOKEN> will give accesss to jupyter notebook. It may take some time to load in browser for the first time.
For running jupyter-lab on aws using putty follow above method but use modified commands below.
docker run --gpus all -it --rm -p 8888:8888 -e JUPYTER_ENABLE_LAB=yes nvcr.io/nvidia/pytorch:21.05-py3
jupyter lab --ip 0.0.0.0 --port 8888 --no-browser --allow-root
For mouting local directory try this command, docker run --gpus all -it --rm -p 8888:8888 -e JUPYTER_ENABLE_LAB=yes -v /home/ubuntu/data/ml:/data/ml nvcr.io/nvidia/pytorch:21.05-py3, here, -v local_dir:container_dir.
Copy file from docker container to host via, docker cp <CONTAINER_ID>:<FILE_TO_COPY_PATH> <HOST_PATH>. For example, docker cp 22d5ce7be75a:/path/to/weight.pt /home/ubuntu/.
OpenCV ImportError: libGL.so.1: cannot open shared object file: No such file or directory is solved by !pip install opencv-python-headless.
This way multiple terminals can be opened and anyone with ip address, ppk file will be able to access and modify jupyter notebook.
Get API token from kaggle account > API > Create New Token. This will provide kaggle.json file. Install kaggle api via pip, !pip install kaggle. Upload the json file and copy/move to ~/.kaggle/kaggle.json, example, cp /workspace/kaggle.json ~/.kaggle/kaggle.json.
Delete the file if uploaded to workspace. Run, chmod 600 /root/.kaggle/kaggle.json.
Download dataset following kaggle API. For example celebA dataset on kaggle can be downloaded by, !kaggle datasets download -d jessicali9530/celeba-dataset.
Setting up ip camera requires having router and pc, but does not need NVR. It can record video via the tool below. Also can do RTSP streaming via, VLC, OpenCV or others to see video or run deep learning models.
SADP Tool - Set gateway, ip, password etc.
IVMS 4200 - Video recording, tools, configurations.
If an embedded python package is required by some python or another language, GUI or console app in windows then it can be used. It can be downloaded here, https://www.python.org/downloads/windows/. An use case here is if a python app installer is created by pyinstaller and it requires embedded python to call some scripts.
But default embedded python is lacking most libraries and does not have pip. In order to get pip first an Windows embeddable package (64-bit) zip file must be downloaded for desired python version. Then get-pip.py should be downloaded and put into where the embeddable python zip was extracted. It can be downloaded following this link, https://pip.pypa.io/en/stable/installation/.
Running, python get-pip.py from terminal/cmd in the python directory will get pip. Now, packages can be installed normally. For example, python -m pip install pyqt5.
In order to test it from a python script it can be called via subprocess. For example,
p = subprocess.Popen(
[os.path.join(os.getcwd(), 'embedded_python', 'python.exe'), 'script_that_requies_embedded_python.py'],
stdout=subprocess.PIPE, stderr=subprocess.PIPE
)