rustytorch/demos

RustyTorch++ Demo Platform

A comprehensive suite of GPU-accelerated demos powered by RustyTorch++, packaged as a single-binary Tauri desktop application. Features 18 interactive demos across medical imaging, scientific computing, AI/ML, finance, and computer vision.

Overview

All demos use the unified rtx-tensor backend for GPU acceleration, supporting:

• NVIDIA CUDA (Windows/Linux)
• Apple Metal (M1/M2/M3/M4/M5)
• CPU fallback (all platforms)

Demo Categories

Medical / Science (6 demos)

Demo	Description	Physics/Algorithm
Virtual Catheter	Real-time hemodynamics PINN	Inverse Navier-Stokes
MRE Elastography	Tissue stiffness estimation	Inverse Helmholtz
Thermal Ablation	3D thermal therapy simulation	Pennes Bioheat Equation
SlideScope Pathology	Stain separation for digital pathology	Non-negative Matrix Factorization
RustyNeuro MEG/EEG	Neuroimaging analysis platform	Signal processing, source localization
Medical Digital Twin	Patient-specific ablation planning	Interactive physics simulation

AI / ML (6 demos)

Demo	Description	Model/Method
Neural Operator PDE	Real-time PDE solving	Fourier Neural Operator (FNO)
Physics-Informed Diffusion	Generative PDE solving	PIDDM (diffusion + physics loss)
FNO Benchmark	Compare solvers	FNO vs FEM vs FDM
PINN Benchmark	Multi-PDE benchmark	Heat, Burgers, Poisson equations
Inference Profiler	Model performance analysis	Latency, throughput, memory
Model Zoo	Pre-trained model gallery	Classification, detection, NLP

Finance (3 demos)

Demo	Description	Method
Portfolio Optimizer	Mean-variance optimization	Markowitz efficient frontier
Risk Analyzer	VaR and stress testing	Monte Carlo simulation
Time Series Forecast	Multi-model forecasting	ARIMA, Prophet, Transformer

Vision (3 demos)

Demo	Description	Model
Image Classifier	Real-time classification	ViT, ConvNeXt (ImageNet)
Object Detector	Bounding box detection	YOLOv8 (COCO 80 classes)
Segmentation	Pixel-wise segmentation	DeepLabV3, SegFormer, UNet

Architecture

demos/
├── shared/                     # Hemodynamics IPC types
├── mre-shared/                 # MRE elastography types
├── bioheat-shared/             # Bioheat transfer types
├── slidescope-shared/          # SlideScope types
├── neuro-shared/               # RustyNeuro types
├── neural-operator-shared/     # FNO demo types
├── piddm-shared/               # Physics-informed diffusion types
├── digital-twin-shared/        # Digital twin types
├── fno-benchmark-shared/       # FNO benchmark types
├── pinn-benchmark-shared/      # PINN benchmark types
├── inference-profiler-shared/  # Inference profiler types
├── model-zoo-shared/           # Model zoo types
├── portfolio-shared/           # Portfolio optimizer types
├── risk-shared/                # Risk analyzer types
├── timeseries-shared/          # Time series types
├── image-classifier-shared/    # Image classifier types
├── object-detector-shared/     # Object detector types
├── segmentation-shared/        # Segmentation types
│
├── rtx-hemodynamics/           # Virtual Catheter backend
├── rtx-mre/                    # MRE elastography backend
├── rtx-bioheat/                # Bioheat simulation backend
├── rtx-slidescope/             # SlideScope NMF backend
├── rtx-neuro/                  # RustyNeuro backend
├── rtx-neural-operator-demo/   # FNO demo backend
├── rtx-piddm-demo/             # PIDDM backend
├── rtx-digital-twin-demo/      # Digital twin backend
├── rtx-fno-benchmark/          # FNO benchmark backend
├── rtx-pinn-benchmark/         # PINN benchmark backend
├── rtx-inference-profiler/     # Inference profiler backend
├── rtx-model-zoo/              # Model zoo backend
├── rtx-portfolio/              # Portfolio optimizer backend
├── rtx-risk-analyzer/          # Risk analyzer backend
├── rtx-timeseries/             # Time series backend
├── rtx-image-classifier/       # Image classifier backend
├── rtx-object-detector/        # Object detector backend
├── rtx-segmentation-demo/      # Segmentation backend
│
├── server/                     # Service layer (all demos)
└── ui/                         # Tauri + React frontend

Quick Start

# Install dependencies
cd demos/ui
pnpm install

# Development mode
pnpm tauri dev

# Production build
pnpm tauri build

Demo Details

Medical / Science

1. Virtual Catheter (Hemodynamics)

Solves the inverse Navier-Stokes problem - inferring pressure fields from velocity data for non-invasive virtual catheterization.

Features:

• Real-time PINN inference with GPU acceleration
• Interactive geometry editing (stenosis, aneurysm, stent)
• Pulsatile flow animation with cardiac waveforms
• Multi-field visualization (velocity, pressure, WSS)

Physics:

Momentum: ρ(∂u/∂t + u·∇u) = -∇p + μ∇²u
Continuity: ∇·u = 0

2. MRE Elastography

Estimates tissue stiffness from MRI wave propagation using the Helmholtz equation.

Features:

• 2D/3D stiffness map reconstruction
• Wave field visualization
• Configurable frequency and wavelength
• Real-time GPU-accelerated solver

Physics:

∇²u + k²u = 0  (Helmholtz)
G = ρω²/k²     (Shear modulus estimation)

3. Thermal Ablation (Bioheat)

Simulates thermal therapy using the Pennes bioheat equation.

Features:

• 3D temperature field simulation
• Heat source placement and configuration
• Thermal dose (CEM43) calculation
• Treatment planning visualization

Physics:

ρc(∂T/∂t) = k∇²T + ρ_b·c_b·ω_b(T_a - T) + Q

4. SlideScope Pathology

GPU-accelerated stain separation using Non-negative Matrix Factorization.

Features:

• H&E and IHC stain unmixing
• Deep zoom tile viewing
• Real-time processing with progress
• Unified rtx-tensor backend

Algorithm:

V ≈ W × H  (NMF decomposition)

5. RustyNeuro MEG/EEG

Comprehensive neuroimaging analysis platform.

Features:

• Signal visualization and filtering
• Epoching and averaging
• Source localization
• Connectivity analysis
• Inspired by Brainstorm

6. Medical Digital Twin

Patient-specific organ simulation for treatment planning.

Features:

• Organ geometry selection
• Probe placement configuration
• Interactive bioheat simulation
• What-if analysis

AI / ML

7. Neural Operator PDE

Real-time PDE solving with Fourier Neural Operators.

Features:

• Draw boundary conditions interactively
• Instant solutions for Darcy flow, heat equation
• 1000x faster than traditional methods
• Train custom models

8. Physics-Informed Diffusion (PIDDM)

Generative PDE solving with physics-consistent diffusion.

Features:

• Train diffusion models with physics loss
• Sample PDE solutions
• Supports Poisson, Heat, Darcy, Burgers
• Physics consistency visualization

9. FNO Benchmark

Compare Neural Operators vs classical solvers.

Features:

• Side-by-side: FNO vs FEM vs FDM
• Resolution sweep (32x32 to 256x256)
• Metrics: time, accuracy, memory
• CSV export

10. PINN Benchmark

Interactive Physics-Informed Neural Network benchmark.

Features:

• Heat, Burgers, Poisson equations
• Training convergence visualization
• Architecture comparison
• Loss curve charts

11. Inference Profiler

Model performance analysis tool.

Features:

• Latency metrics (p50, p95, p99)
• Throughput measurement
• Memory profiling
• GPU utilization
• Batch size sweeps

12. Model Zoo

Pre-trained model gallery for common ML tasks.

Features:

• Image classification (ResNet, ViT, ConvNeXt)
• Object detection (YOLO)
• Segmentation (DeepLab, SegFormer)
• Text generation, NLP, audio
• One-click inference

Finance

13. Portfolio Optimizer

Mean-variance portfolio optimization.

Features:

• Efficient frontier visualization
• Asset correlation matrix
• Risk/return trade-off analysis
• GPU-accelerated covariance estimation

14. Risk Analyzer

Value-at-Risk and stress testing.

Features:

• Historical and parametric VaR
• Monte Carlo simulation
• Stress scenarios
• Risk decomposition

15. Time Series Forecast

Multi-model time series forecasting.

Features:

• ARIMA/SARIMA models
• Prophet integration
• Transformer architectures
• Confidence intervals
• Sample datasets

Vision

16. Image Classifier

Real-time image classification.

Features:

• ViT and ConvNeXt models
• 1000 ImageNet classes
• GPU acceleration
• Confidence visualization

17. Object Detector

YOLO-style object detection.

Features:

• YOLOv8 variants (Nano to XLarge)
• 80 COCO classes
• Adjustable confidence threshold
• NMS configuration
• Bounding box visualization

18. Segmentation

Pixel-wise semantic segmentation.

Features:

• DeepLabV3, SegFormer, UNet, FCN
• PASCAL VOC (21 classes)
• ADE20K (150 classes)
• Cityscapes (19 classes)
• Mask overlay visualization

GPU Backend

All demos use rtx-tensor for GPU operations:

[features]
default = ["cpu"]
cpu = ["rtx-tensor/cpu"]
cuda = ["rtx-tensor/cuda"]
metal = ["rtx-tensor/metal"]

Supported Operations

• Matrix multiplication (matmul)
• Element-wise operations (add, sub, mul, div)
• Activation functions (relu, sigmoid, tanh, gelu, silu)
• Trigonometric functions (sin, cos) for Fourier features
• Reduction operations (sum, mean, max, min)
• Transpose, reshape, convolution

Testing

# Run all backend tests
cargo test --workspace -p "*-shared" -p "rtx-*"

# Run specific demo tests
cargo test -p rtx-hemodynamics
cargo test -p inference-profiler-shared
cargo test -p segmentation-shared

# Frontend tests
cd demos/ui && pnpm test

Test Coverage

Category	Crates	Tests
Medical/Science	12	600+
AI/ML	12	500+
Finance	6	200+
Vision	6	250+
Total	36	1,550+

Performance Targets

Metric	Target
Inference latency	<50ms for 10k points
UI responsiveness	30+ FPS
GPU memory	<2GB per simulation
Build time	<5 min full rebuild

License

Part of RustyTorch++. See repository root for license information.