How tishansathruwan/maldam works
MaLDAM is a YOLOv5-based malaria detector trained for HCM-to-LCM transfer. Its main workflow pairs HCM and LCM views, adapts the HCM view with microscopy augmentations, and jointly trains detection and VICReg-based global/masked-local alignment; validation evaluates cross-domain detection. The graph preserves major training, evaluation, logging, and model-support modules. Some links reflect README descriptions or SOURCE INDEX references; unsampled implementations and their internal call paths are not asserted.
Main entry points: contrastive_trainer.py Training orchestrator · dataloaders_contrastive.py Paired HCM/LCM loader · val.py Validation and benchmarking · loss.py VICRegL alignment loss
Components 23
Paired Data
- Paired HCM/LCM loader
src/utils/dataloaders_contrastive.py - Detection data loader
src/utils/dataloaders.py - Microscopy augmentation
src/utils/augmentations.py
Joint Training
- Run configuration
src/config.py - Training orchestrator
src/contrastive_trainer.py - Detection loss
src/utils/loss.py - VICRegL projection head
src/utils/vicregl/model.py - VICRegL alignment loss
src/utils/vicregl/loss.py - Local feature matching
src/utils/vicregl/utils.py - Feature matching primitives
src/utils/feature_matching.py - Device and training utilities
src/utils/torch_utils.py - Shared training utilities
src/utils/general.py
Detection Model
- YOLO detector
src/models/yolo.py - Neural network layers
src/models/common.py - Model loading and ensembles
src/models/experimental.py - Anchor tuning
src/utils/autoanchor.py
Evaluation and Reporting
- Validation and benchmarking
src/utils/val.py - Detection metrics
src/utils/metrics.py - Plots and visualizations
src/utils/plots.py - Training callbacks
src/utils/callbacks.py - Experiment logging
src/utils/loggers/__init__.py - Checkpoints and results
src/contrastive_trainer.py
Other components
- Training operator
Connections 22
- Training operator → Training orchestrator: starts
- Training orchestrator → Run configuration: loads config
main.py - Training orchestrator → Paired HCM/LCM loader: loads paired views
src/contrastive_trainer.py - Paired HCM/LCM loader → Detection data loader: uses loader
src/utils/dataloaders_contrastive.py - Paired HCM/LCM loader → Microscopy augmentation: applies transforms
src/utils/dataloaders_contrastive.py - Training orchestrator → YOLO detector: trains model
src/contrastive_trainer.py - YOLO detector → Neural network layers: builds from
src/models/yolo.py - Training orchestrator → Detection loss: computes loss
src/contrastive_trainer.py - Training orchestrator → VICRegL projection head: creates head
src/contrastive_trainer.py - Training orchestrator → VICRegL alignment loss: computes alignment
src/contrastive_trainer.py - VICRegL alignment loss → Local feature matching: uses matching
src/utils/vicregl/loss.py - Training orchestrator → Feature matching primitives: uses primitives
- Training orchestrator → Anchor tuning: checks anchors
src/contrastive_trainer.py - Training orchestrator → Device and training utilities: uses device utilities
src/contrastive_trainer.py - Training orchestrator → Shared training utilities: uses shared utilities
src/contrastive_trainer.py - Training orchestrator → Validation and benchmarking: runs validation
src/contrastive_trainer.py - Validation and benchmarking → Detection metrics: computes metrics
src/utils/val.py - Validation and benchmarking → Plots and visualizations: produces plots
src/utils/val.py - Training orchestrator → Training callbacks: dispatches hooks
src/contrastive_trainer.py - Training orchestrator → Experiment logging: reports run
src/contrastive_trainer.py - Training orchestrator → Model loading and ensembles: loads checkpoints
src/contrastive_trainer.py - Training orchestrator → Checkpoints and results: saves checkpoints
src/contrastive_trainer.py