Running Deep Mosaic in GPU Mode on $40 NVidia Tesla K80 & Tesla M60 - Success! #149
Description
NVidia Telsa K80 cards are dirt cheap on eBay now, as datacenters are shedding them now!
https://www.ebay.com/sch/i.html?_nkw=tesla+k80+24gb
So, I decided pick one up and stick it in a old ASUS TUF B350M-PLUS GAMING motherboard with Ryzen 5 Pro 2400GE processor.
These are the steps I followed to get this working:
K80 is a space heater and draws a lot of power, I am using a 650W PS>
Tesla K80 does not have any video output ports, so you need a motherboard with IGFX or another video card to connect your monitor.
Enable above 4G decoding in BIOS, usually in the PCI Subsystem Tab, this has to be done prior to putting the card in the motherboard. Otherwise the card/machine will not boot.
Install the card in the first (CPU) PCI slot.
Windows will recognize the card and install default drivers from 2015. Use device manager to check if the card has been recognized by your system.
Download & Install NVidia Datacenter Driver (Windows 10, this driver works on Windows 11)
Download & Install NVidia Cuda 11.0
Download & Install ffmpeg binaries - I used the latest version. Make sure your PATH points to the installation directory/bin or wherever you put ff*.exe-s
https://www.gyan.dev/ffmpeg/builds/
Download Python 3.7
Create a virtual environment
py -m venv D:\DMvEnv
Activate virtual environment
cd D:\DMvEnv
Scripts\Activate
Ensure that the correct and most recent version of pip is installed (command to invoke Python 3.7 seems to be py and python)
py -m pip install --upgrade pip
Dependencies for Requiremetns.TXT - I had to install them one by one
pip install numpy==1.19.3
pip install opencv_python==4.5.1.48
pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
pip install matplotlib==3.3.2
pip install tensorboardX==2.2
pip install scikit-image==0.17.2
Now download the source code from github
git clone https://github.com/HypoX64/DeepMosaics.git
Testing the installation
(1) Copy the pretrained models to ./pretrained_models
Make sure you are in the root folder of the DeepMosaic source tree
cd D:\DMvEnv\DeepMosaics
(2)Test Adding Mosaic
py deepmosaic.py --media_path ./imgs/ruoruo.jpg --model_path ./pretrained_models/add_face.pth --gpu_id 0
(3) Test Removing Mosaic
py deepmosaic.py --media_path ./result/ruoruo_add.jpg --model_path ./pretrained_models/clean_face_HD.pth --gpu_id 0
P.S. My biggest struggle was to figure out why the dependencies were not installing, once I figured out pip version was the issue, rest of the installation was a breeze. Hope this helps someone out there trying to run in GPU mode. The steps are valid for all NVidia cards (assuming CUDA 11 support is present).
You can pick the pip command for pytorch here:
https://pytorch.org/get-started/previous-versions/