Machine learning for 3D graphics applications on AMD platforms
Machine Learning for 3D Graphics is a separate domain that is becoming the mainstream in the recent advancements of computer graphics. It replaces analytical algorithms in solving complex lighting, temporal stability and aliasing problems with ML based statistically approximated methods that are both faster and have better visual quality compared with the former methods.
The workshop has the goal of providing hands on experience on the ML methods and especially in ML training conducted with the use of AMD architecture of AMD Threadripper workstation systems equipped with ADM Radeon W6800 Pro GPU cards.
We expect the participants to be familiar with pytorch, basic concepts of machine learning and neural networks, and some principles of 3D graphics.
You will have a unique opportunity to collaborate with experts in ML based solutions for modern 3D Graphics
Organizers:
- Tomasz Madajczak, PhD, Manager Software Development, AMD
- Tomasz Galaj, PhD, 3D Graphics engineer, AMD
- Michal Znalezniak, ML engineer, AMD
- Jakub Poła, GPGPU Compute engineer, AMD
- Damian Andrysiak, ML Engineer, AMD
Setup of AMD GPU for ML or GPGPU
Jakub Poła and Michal Znalezniak will present setup of ROCm (Radeon-Open-Computing) libraries that are comprising the GPU backend for pytorch machine learning. We will start with the brief introduction to ROCm platform and HIP C++ runtime API and kernel language.
The participants are invited to read:
- Introduction to ROCm:
- Docker setup guidelines for AMD Radeon
Basic ML models for Image Processing
Michal Znalezniak will present a workshop on training CNNs for simple image processing with showing various loss function in action.
Recommended reading before the session:
Single Image Upscaling for 3D Graphics
Damian Andrysiak will present a workshop on upscaling from 1080p into 4k resolution game re
Recommended reading before the session:
- What is a UNET
- What is a VGG Loss
- Deconvolutions (Transpose Conv2D) and their checkerboard artifact
- Why L1 norm is better for SR than L2 norm? and this
Generative AI in action with Diffusion Model
Michal Znalezniak and Damian Andrysiak will present a workshop on Generative AI that shows 3D context generation from text with Stable Diffusion.
Recommended reading before the session:
miniNN NRC AI in action
Tomasz Galaj will present a workshop on MLP network for image reconstruction, that have endless number of applications in 3D graphics.
Recommended reading before the session: