Examples and use cases for darkon
pip install darkon
Influence score calculation example It first trains CNN model for MNIST data, and calculate influence score to find most helpful and harmful training samples. Then, retrain the model after removing 100 most helpful training samples and 100 most harmful training samples, respectively.
- dataset: cifar10
- network: resnet110
- Upweight influence function
- Mislabel detection with all of layers
- Mislabel detection with one top layer
- model: used pre-trained model in TF slim
- network: resnet v1 50 in TF slim
- Gradcam & Guided Gradcam
- model: used pre-trained model
- network: cnn text classification by dennybritz
- Gradcam for text sentiment classification