PlaidML-Multi-GPU

A demo application of deep learning framework based on OpenCL for multi-GPUs

Young-Woon Lee, Young-Jin Heo, Chang-Sik Cho, Byung-Gyu Kim*, Open-CL Based Multi GPU Acceleration for Video Object Detection, IEEE International Conference on Consumer Electronics (ICCE2021) (IEEE), Jan. 2021.

1. Environment

- CPU: AMD Ryzen Threadripper 2970wx 24
- GPU: AMD Radeon RX Vega 64 8GB (x2)

- OS: Ubuntu Linux 18.04.3 LTS (linux kernel 5.1)

- python==3.6.9
- numpy==1.17.3
- keras==2.2.4
- tensorflow==1.13.1 (NOT!!! tensorflow-gpu)
- pyqt==5.9.2
- opencv==3.4.2
- plaidml-keras==0.6.4  

2. Usage

Create Virtual Environment using conda (anaconda3 or miniconda3)

at root directory of PlaidML-Multi-GPU repository where has environment.yml

$ conda env create
$ conda activate plaidml_env
(plaidml_env) $

Important:

1. This demo source code assumes that the currently running environment is equipped with two GPUs.

2. Use the modified plaidml's keras backend code accordingly, which is src/backend.py.

3. You must replace the original backend.py file of plaidml installed in the current environment with src/backend.py to operate normally.

Example:

$ pwd
/home/USER_HOME/PlaidML-Multi-GPU/

move where plaidml-keras package installed

$ cd /home/USER_HOME/miniconda3/envs/PLAIDML_ENV/lib/python3.6/site-packages/plaidml/keras/ 

change the original source file

$ mv backend.py backend.py.old

copy the customized backend.py

$ cp /home/USER_HOME/PlaidML-Multi-GPU/src/backend.py backend.py

Download pre-trained weight file to weight/VGG_VOC0712_SSD_300x300_iter_120000.h5

https://drive.google.com/open?id=121-kCXaOHOkJE_Kf5lKcJvC_5q1fYb_q

also you can find details here:

https://github.com/pierluigiferrari/ssd_keras

Start demo program

(plaidml_env) $ python main.py

3. Summary of demo system

The developed application contains the following components and functions.

1. window view: This is the part where the video is output. The resolution of the video being played is adjusted.

2. Single button: That is, it uses only 1 GPU out of 2 GPUs in action. The Stop button is activated during operation.

3. Multi button: Both GPUs are used. The Stop button is activated during operation.

demo GUI shows like this:

The screen that executed the program by inputting a video file is as follows:

when using only 1-GPU:

when both GPUs are used:

Reference

We highly inspired from the following repositories for this application development:

https://github.com/pierluigiferrari/ssd_keras

https://github.com/tanakataiki/ssd_kerasV2