WARNING: There are some mistakes in the code available here, please do not use it as a benchmark or component. I will try to fix the project as soon as possibe. Please see the issues page to know the error.
Feature Generating Networks for ZSL in Pytorch
PyTorch implementation of paper: Feature Generating Networks for Zero-Shot Learning
4 datasets are currently supported: SUN, CUB, AWA1 & AWA2. All datasets can be downloaded here.
IMPORTANT:
The downloaded zip will have many files for each dataset, but we only require 2 files res101.mat & att_splits.mat. Move these 2 files per dataset to the appropriate folder in this repo before starting to train/test.
- For training the model, use:
python3 main.py --n_epochs 20 --use_cls_loss
All trainable parameters are saved in a folder named saved_models at the end of every epoch.