AlexLens is an Image Classification and Transfer Learning program for heterogeneous platforms with the Neural-Net-Classification built from scratch. It was developed during a project at KIT. The team consisted of Viet Pham, Dmitrii Seletkov, Aleksandr Eismont, Jakub Trzciński and Friedemann Claus.
The program is designed to run on a specific PC, which has Ubuntu 18.04 as the operating system, 8 GB of RAM, a Core-i5 4600 CPU and its onboard graphics as the GPU. But with exception to the GPU classification of AlexNet it should work on most other hardware (not software) as well - unsupportedly. It also takes advantage of up to four Intel Movidius Neural Compute Sticks (Gen 1 is officially supported, but Gen 2 also works according to our tests).
For it to run you need to follow three steps:
- Download "resources" here
- Unzip it and place it in the AlexLens folder
- Install the necessary libraries, if you don't have them running already:
Optionally, we also provide a training dataset as an example here.
The necessary libraries are:
| Library | Usecase | Download/ Tutorial | Comments |
|---|---|---|---|
| Qt5 | GUI | Download Page | You can skip the registration. In the installer, select under "Qt 5.13.0": Desktop gcc 64-bit, Sources, Qt Charts, Qt Data Visualization, Qt Debug Information Files and under "Developer and Designer Tools": Qt 3D Studio 2.4.0 |
| Eigen | Matrices and vectors | sudo apt-get install libeigen3-dev |
that's all |
| Torch | Transfer Learning and CPU-Classification of other Neural Networks than AlexNet | Download | Unzip and move libtorch folder to AlexLens/thirdparty |
| OpenVINO | ASIC-Sticks and the included OpenCV | Tutorial | Go through steps 1,2,3,5,7 and 9. After Step 9 you should have a folder named "inference_enginge_samples_build" in your /home/"username" directory |
| OpenCL | Low-level access to the GPU | sudo apt update sudo apt install ocl-icd-opencl-dev sudo apt-get install beignet |
Confirmed to work on Intel HD Graphics of 4th, 7th and 8th generation |
| HDF5 | To read the weights file with a good performance | sudo apt-get install libhdf5-dev |
that's all |
| Libusb | For dynamically detecting the amount of USB-devices used | sudo apt-get install libusb-1.0-0-dev |
that's all |