How thienhocdata/quality-aware-gan-augmentation works
This is a research project skeleton for quality-aware GAN augmentation of imbalanced image datasets. Its documented workflow trains a GAN using training data, generates minority-class images, filters them for quality, class correctness, and diversity, then evaluates a classifier trained on a balanced set. The README also identifies baseline, oversampling, augmentation, and class-weighting comparisons. Source excerpts show package boundaries but not implementations or runtime wiring, so the graph reflects documented capabilities and workflow rather than verified calls.
Main entry points: data Preprocessing and loaders · gan Minority GAN training · evaluation Metrics and comparison · baseline.yaml Experiment settings
Components 14
Data preparation
- Imbalanced images
data/raw - Preprocessing and loaders
src/data - Experiment settings
configs/baseline.yaml
GAN generation
- Minority GAN training
src/gan - Generated images
outputs/generated - Model checkpoints
outputs/checkpoints
Filtering and training
- Quality and diversity filtering
src/filtering - Balanced training set
data/processed - Image classifier
src/classifier - Baseline and balancing methods
src/classifier
Evaluation
- Metrics and comparison
src/evaluation - Results and reports
reports - Run logs
outputs/logs
Other components
- Researcher
Connections 18
- Researcher → Experiment settings: sets experiment
- Imbalanced images → Preprocessing and loaders: provides images
- Experiment settings → Preprocessing and loaders: configures
- Preprocessing and loaders → Minority GAN training: provides train data
- Experiment settings → Minority GAN training: configures
- Minority GAN training → Generated images: writes images
- Minority GAN training → Model checkpoints: saves model
- Generated images → Quality and diversity filtering: scores and filters
- Preprocessing and loaders → Quality and diversity filtering: provides class data
- Quality and diversity filtering → Balanced training set: writes selected images
- Preprocessing and loaders → Balanced training set: provides original data
- Balanced training set → Image classifier: trains on
- Preprocessing and loaders → Image classifier: provides evaluation data
- Preprocessing and loaders → Baseline and balancing methods: provides data
- Baseline and balancing methods → Metrics and comparison: provides comparison results
- Image classifier → Metrics and comparison: provides predictions
- Metrics and comparison → Results and reports: writes results
- Metrics and comparison → Run logs: records metrics