How ssraghuvanshi/panet works
Project Type and Purpose
• This is a deep‐learning research codebase for 3D medical image segmentation (BraTS 2020) using a Prior Attention Network (PANet).
• It includes training (train.py), inference (inference.py), model definitions (models/), configuration files (configs/), and data utilities (utils/).
File Structure Analysis
Top-level:
• configs/ — optimizer configuration files (adam.cfg, sgd.cfg)
• train.py — main training entry point
• inference.py — main inference entry point
• models/ — all network architectures and building blocks
• utils/ — data loading, augmentation, iterator logic
Patterns:
• Modular separation of concerns: configurations, data pipeline, model definitions, scripts
• Plug-and-play optimizers via configs
• Clear training vs. inference scripts
Key Components
a. Data Ingestion & Preprocessing
– External Data Store (./data) containing raw MRI volumes
– utils/data_pipeline.py: patch extraction, normalization, augmentation
– utils/iterator.py: batching and sampling (training vs. validation)
b. Configuration Management
– configs/*.cfg loaded at startup by train.py and inference.py to set hyperparameters, optimizer type, scheduler, mixed precision flags
c. Model Stack
– models/basic_modules.py: core convolutional layers, normalization, activation modules
– models/attention_blocks.py: prior attention mechanisms, parameterized skip connections
– models/networks.py: end‐to‐end PANet architecture, deep supervision heads
d. Training Loop (train.py)
– Reads configs → initializes data iterator → builds model → sets optimizer/scheduler → runs epochs with patch-based data feeding
– Implements deep supervision losses and intermediate attention guidance
– Checkpointing (model + optimizer + scheduler states)
e. Inference Pipeline (inference.py)
– Loads checkpoint → uses same patch-based pipeline → stitches 3D patches to full volume → outputs segmentation maps
Relationships & Data Flow
1. External Data → Data Pipeline → Iterator → Model Forward Pass
2. Model Outputs + Ground Truth → Loss Computation (deep supervision) → Backprop → Optimizer Update
3. Model Checkpoint → Saved to disk → Re‐loaded by inference.py
4. Inference: Iterator → Model → Post-processing → Segmentation Output
Architectural Patterns and Principles
• Modular separation: scripts (train/inference), configs, utils, models
• Pipeline pattern for data flow (chain of responsibility)
• Plug-in configuration (config files determine optimizer, mixed precision, data handling)
• Layered model definition (basic modules → attention blocks → full network)