How tomouellette/herb works
Component Details
Training Infrastructure (Left):
• FolderDataset
• Optional WebDataset support
• To data/ directory
• Below that, create a "Training Models" section containing:
• DINO (Self-supervised learning)
• MAE (Masked Autoencoder)
• MBT (Masked Barlow Twins)
Model Architectures (Center):
• ConvNext
• MLP
• MLP Mixer
• NAViT
• ViT
• Parameter scaling variants:
• nano (2.5M params)
• micro (5M params)
• tiny (10M params)
• small (20M params)
• base (80M params)
• large (200M+ params)
Inference Infrastructure (Right):
• Candle ConvNext
• Candle MLP Mixer
• Candle ViT
• "Model Format" component showing:
• PyTorch (.pth)
• SafeTensors format
Connections and Data Flow
• Data flow from input to training models
• Model architecture selection for training
• Model saving paths to both PyTorch and SafeTensors formats
• SafeTensors loading path to Rust inference components
Additional Elements
• Configuration parameters
• Resource requirements
• Input/Output specifications
• Image size requirements (224x224)
Technical Details Layer
• Training hyperparameters
• Model-specific parameters
• I/O specifications
• GPU/CPU requirements
The dual-language implementation (Python/Rust)
Remember to maintain a balance between showing the technical depth of the ML components while keeping the overall system architecture clear and understandable.