How idata-ora/hiadapter works
HiAdapter is a pathology-image classification research system for adapting pathology foundation models to downstream datasets. Its principal workflow prepares patch datasets, runs a UNI-based model with stain- and morphology-oriented adaptation, trains with classification and optional prototypical contrastive loss, and evaluates predictions. The README also documents five-fold data splitting and broader benchmark, unseen-stain/cancer, interpretability, and validation capabilities. Adapter wiring is only partly visible in the excerpts, so unsampled implementation details are not expanded.
Main entry points: function.py Training loop · train.py Training entry · vision_transformer.py UNI transformer · myTileDataset.py Patch dataset
Components 17
Data preparation
- Patch dataset
myDatasets/myTileDataset.py - Fold splitting
myDatasets/split_5fold.py - Prompt generation
myDatasets/get_prompt.py
Adapted model
- UNI transformer
model/uni/vision_transformer.py - Transformer layers
model/uni/layers - Stain adaptation
utils/cModule.py - Morphology fusion
utils/cfusion.py
Training and evaluation
- Training entry
train.py - Training loop
core/function.py - Classification losses
core/losses.py - PPCL objective
core/PPCL.py
Training support
- Metrics and reporting
utils/tools.py - Runtime utilities
utils/utils.py - Run parameters
utils/params.py - Optimizer configuration
utils/optConfig.py - Learning-rate schedule
utils/lrSchedulearConfig.py
Other components
- Researcher
Connections 19
- Researcher → Training entry: starts training
README.md - Researcher → Fold splitting: prepares folds
README.md - Fold splitting → Patch dataset: provides splits
- Training entry → Run parameters: loads options
train.py - Training entry → Patch dataset: builds datasets
train.py - Training entry → UNI transformer: selects model
train.py - Training entry → Training loop: runs training
train.py - Patch dataset → Runtime utilities: uses utilities
myDatasets/myTileDataset.py - Patch dataset → Prompt generation: uses prompts
- UNI transformer → Transformer layers: uses layers
model/uni/vision_transformer.py - UNI transformer → Stain adaptation: uses module
model/uni/vision_transformer.py - UNI transformer → Morphology fusion: uses fusion
model/uni/vision_transformer.py - Training loop → Classification losses: computes loss
core/function.py - Training loop → PPCL objective: updates proxies
core/function.py - Training loop → Metrics and reporting: uses metrics
core/function.py - Training loop → Runtime utilities: uses utilities
- Training loop → Optimizer configuration: configures optimizer
- Training loop → Learning-rate schedule: configures schedule
- Training entry → Runtime utilities: uses utilities
train.py