System Architecture

Ununennium is designed as a layered framework for high-performance geospatial machine learning.

graph TD
    User[User API] --> High[High-Level API]
    User --> Core[Core Primitives]

    subgraph "High-Level API"
        Models[Model Registry]
        Trainer[Trainer / Callbacks]
        Pipe[Pipelines]
    end

    subgraph "Core Primitives"
        GT[GeoTensor]
        GB[GeoBatch]
        IO[Streaming I/O]
    end

    subgraph "Backend"
        Torch[PyTorch]
        GDAL[GDAL/Rasterio]
        CUDA[CUDA Kernels]
    end

    High --> Core
    Core --> Backend

Core Components

1. GeoTensor

The fundamental data structure extending torch.Tensor with geospatial metadata: - Affine Transform: Maps pixel coordinates to map coordinates. - CRS: Coordinate Reference System (WKT/EPSG). - Bounding Box: Spatial extent management.

\text{Pixel}(col, row) \xrightarrow{M} \text{Map}(x, y)

2. Streaming I/O

Lazy loading architecture for files larger than memory (e.g., 100GB+ Sentinel-2 scenes). Uses windowed reading with caching.

3. Model Registry

Standardized interface for 15+ models (CNN, ViT, GAN, PINN), supporting: - Automatic weight downloading - Config-driven instantiation - Standardized forward/predict signatures

Data Flow

sequenceDiagram
    participant D as Data Storage (S3/Local)
    participant I as IO Layer
    participant T as Tiler
    participant M as Model

    D->>I: Stream Window
    I->>T: Raw Patch
    T->>T: Augment & Normalize
    T->>M: Forward Pass
    M->>T: Prediction
    T->>I: Write Result