How jmhb0/o2vae works
Main Components
a. Data Processing Layer:
• Input data handling (data/datasets.py)
• Data preprocessing components (prealignment/)
• Dataset-specific modules (O2MNIST and MEFS datasets)
b. Model Architecture:
• Core VAE implementation (models/vae.py)
• Encoder components:
• O2-invariant encoders (models/encoders_o2/)
• Vanilla encoders (models/encoders_vanilla/)
• Decoder components (models/decoders/)
• Registration module (registration/)
c. Training Infrastructure:
• Training loops (train_loops.py)
• Configuration management (configs/)
• Logging and monitoring (Weights & Biases integration)
d. Utility Layer:
• Evaluation utilities (utils/eval_utils.py)
• Plotting utilities (utils/plotting_utils.py)
• Cluster utilities (utils/cluster_utils.py)
[Input Data] → [Preprocessing] → [Model Architecture] → [Output Representations]
• ↑
• [Training Infrastructure]
• ↑
• [Configuration & Utils]
Key Relationships to Highlight
a. Data Flow:
• How data moves from raw input through preprocessing to the model
• VAE's encoder-decoder architecture
• Demonstrate how the registration module interfaces with the training process
b. Training Flow:
• Configuration management flow
• Training loop interaction with model components
• Logging and monitoring pathways
External Dependencies
• PyTorch integration
• e2cnn library connection
• Weights & Biases logging
Deployment Components
• Model saving/loading paths
• Pretrained model handling
• Example notebook integration
• Any special symbols used for O2-invariance features
• Configuration options and their impacts
This design should emphasize the project's unique focus on orientation-invariant representation learning while maintaining clarity about the overall system architecture.