Adaptive Confidence Smoothing for Generalized Zero-Shot Learning

Code for our paper: Atzmon & Chechik, "Adaptive Confidence Smoothing for Generalized Zero-Shot Learning", CVPR 2019

Installation

Code and Data

Download or clone the code in this repository.
cd to the project directory
Download the data (~500MB) by typing
wget http://chechiklab.biu.ac.il/~yuvval/COSMO/data.zip
Extract the data by typing
unzip -o data.zip
data.zip can be deleted

Anaconda Environment

Quick installation under Anaconda:

conda env create -f conda_requirements.yml

Alternatively, below are detailed installation instructions with Anaconda.

yes | conda create -n COSMO python=3.6

conda activate COSMO

yes | conda install pandas ipython jupyter nb_conda matplotlib 
yes | conda install scikit-learn=0.19.1
yes | pip install h5py seaborn

# Other packages that I find useful, but are unrelated to this project
yes | conda install -c conda-forge jupyter_contrib_nbextensions
yes | conda install -c conda-forge jupyter_nbextensions_configurator

Directory Structure

directory	file	description
`src/`	*	Sources files
`src/utils/`	*	Sources to useful utility procedures.
`data/`	{CUB, AWA1, SUN, FLO}/	Xian (CVPR, 2017) zero-shot data for CUB, AWA1, SUN, and FLOWER.
`data/LAGO_GZSL_predictions/*/`	pred_gzsl_val.npz	predictions of LAGO on GZSL-val set, when trained on train set.
`data/LAGO_GZSL_predictions/*/`	pred_gzsl_test.npz	predictions of LAGO on GZSL-test set, when trained on train+GZSLval set.
`data/XianGAN_predictions/*/`	pred_gzsl_val.npz	predictions of fCLSWGAN on GZSL-val set, when trained on train set.
`data/XianGAN_predictions/*/`	pred_gzsl_test.npz	predictions of fCLSWGAN on GZSL-test set, when trained on train+GZSLval set.
`output/`	*	Contains the outputs of the experimental framework (results & models).
`output/seen_expert_model/`	*	Contains cached models for seen experts.
`output/COSMO/`	*	Contains results of the experimental framework.

Execute COSMO

To execute COSMO+LAGO, submit the following in the main project dir.

PYTHONPATH="./" python src/main_cosmo.py --dataset_name=CUB --data_dir=data/CUB --zs_expert_name=LAGO
PYTHONPATH="./" python src/main_cosmo.py --dataset_name=AWA1 --data_dir=data/AWA1 --zs_expert_name=LAGO
PYTHONPATH="./" python src/main_cosmo.py --dataset_name=SUN --data_dir=data/SUN --zs_expert_name=LAGO

To execute COSMO+fCLSWGAN, submit the following in the main project dir.

PYTHONPATH="./" python src/main_cosmo.py --dataset_name=CUB --data_dir=data/CUB --zs_expert_name=XianGAN
PYTHONPATH="./" python src/main_cosmo.py --dataset_name=AWA1 --data_dir=data/AWA1 --zs_expert_name=XianGAN
PYTHONPATH="./" python src/main_cosmo.py --dataset_name=SUN --data_dir=data/SUN --zs_expert_name=XianGAN
PYTHONPATH="./" python src/main_cosmo.py --dataset_name=FLO --data_dir=data/FLO --zs_expert_name=XianGAN

NOTE:

You must run the code from the project root directory.

Cite our paper

If you use this code, please cite our paper.

@inproceedings{atzmon2019COSMO,
title={Adaptive Confidence Smoothing for Generalized Zero-Shot Learning},
author={Atzmon, Yuval and Chechik, Gal},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year={2019},
}

yuvalatzmon / COSMO