sbharadwajj / embarrassingly-simple-zero-shot-learning

This is a started/demo code for Zero-Shot-Learning via implementation of Embarrassingly simple ZSL (ICML 2015)

Embarrsingly simple zero-shot learning

This is the implementation of the paper "An embarrassingly simple approach to zero-shot learning." (EsZsl) ICML, [pdf].

The file demo_eszsl is a jupyter notebook which contains a walk through of EsZsl.

Dataset

The dataset splits can be downloaded here, please download the Proposed Split and place it in the same folder.

Find additional details about the dataset in the README.md of the Proposed split.

Training and Testing

If you want to skip the demo and just run training and testing for different dataset splits use:

python main.py --dataset SUN --dataset_path xlsa17/data/ --alpha 3 --gamma 1

Setting the hyperparameters alpha and gamma is optional. If the values are not given, the code will evaluate on the train and validation set to find the best hyperparameters.

Results

This version does not have the kernel implementation used in the paper. Hence the results fluctuate by a margin of 1-4%.

The results are taken from the paper Zero-Shot Learning - A Comprehensive Evaluation of the Good, the Bad and the Ugly and are evaluated for features extracted from ResNet-50 for the Proposed split.

Dataset	Paper - (top-1 accuracy in %)	Respository Results	Hyper-params(trainval & test)
CUB	53.9	51.31	Alpha=2, Gamma=0
AWA1	58.2	56.19	Alpha=3, Gamma=0
AWA2	58.6	54.50	Alpha=3, Gamma=0
aPY	38.3	38.47	Alpha=3, Gamma=-1
SUN	54.5	55.62	Alpha=2, Gamma=2

References

If this repository was useful for your research, please cite.

@misc{chichilicious,
  author = {Bharadwaj, Shrisha},
  title = {embarrsingly-simple-zero-shot-learning},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/chichilicious/embarrsingly-simple-zero-shot-learning}},
}