Semantically Aligned Bias Reducing Zero-Shot Learning

Implementation of Alignment GAN using Spectral Normalization (AGSN) for Generalized Inductive Zero-Shot Learning in PyTorch.

AGSN takes motivation from the models SABR by Paul et al., Spectral Normalization by Miyato et al., CADA-VAE by Schonfeld et al., GDAN by Huang et al. and many others.

Prerequisites

Compatible with Python 3.x, PyTorch 1.5.

Training the model

1. Download datasets for zero shot learning from http://datasets.d2.mpi-inf.mpg.de/xian/xlsa17.zip.
2. Place the datasets in some folder and change the data_root string accordingly.
3. If the pretrained flag is set to True, then the trained models will be loaded from the path specified by model_path.

The default dataset used is CUB.

Here are all the pretrained models.