This repository contains the code for the new loss function found by using the Genetic Programming (GP) approach.
The full paper can be downloded here.
The repository contains an example of how NGL can be used to train the InceptionV3 model on CIFAR-100 dataset.
The repository contains two implementations of the NGL loss: in pytorch and in tensorflow. Both versions can be used to train deep learning models for image classification.
NGL was evaluated on seven datasets, which differed by the number of images, classes, by the type of images (grayscale and RGB), and their sizes.
| Loss | Malaria | Pcam | Colorectal Histology | CIFAR-10 | Fashion-MNIST | CIFAR-100 | Caltech 101 | Mean |
|---|---|---|---|---|---|---|---|---|
| CE | 94.0 | 69.4 | 88.9 | 92.8 | 94.0 | 68.2 | 72.5 | ±0 |
| SCE | -0.06 | +0.62 | -0.45 | -3.66 | -2.42 | -0.80 | +2.71 | -0.58 |
| Focal | +0.34 | +1.03 | +2.89 | -0.64 | -0.02 | -2.14 | -0.78 | +0.10 |
| CE + L2 | +0.11 | -0.64 | +0.88 | +0.32 | +0.09 | +0.38 | +3.67 | +0.69 |
| NGL | -0.27 | +7.07 | +1.77 | +0.12 | +0.07 | +1.01 | +5.00 | +2.11 |
See paper for details.
