This is the PyTorch implementation of paper: Hyperbolic Visual Embedding Learning for Zero-Shot Recognition (CVPR 2020)

git clone git@github.com:ShaoTengLiu/Hyperbolic_ZSL.git
cd Hyperbolic_ZSL/code

Hyperbolic_ZSL/code is the default root directory.

• anaconda is recommended
• PyTorch version 1.0 or higher
• At least one GPU with adequate memory
• python 3.7 (python 2 may be possible with modifications)

Use environment.yml for quick start.

conda env create -f environment.yml

Generate data and put them in these folders:

• ../data/train/img: pretrained feature from ResNet of ImageNet
• ../data/val/img: pretrained feature from ResNet of ImageNet
• ../data/train/label: pretrained word embedding from chosen word model
• ../data/val/label: pretrained word embedding from chosen word model

One can run our code via

CUDA_VISIBLE_DEVICES=1 python devise_hype.py \
    --word_model poincare \
    --model_folder model_path \
    --loss_path loss_path \
    --dimension 300

• model_path: the folder to save models
• loss_path: the .jpg tp save loss curve

To run DeViSE in Euclidean space via

CUDA_VISIBLE_DEVICES=2 python devise.py \
    --word_model glove \
    --model_folder model_path \
    --loss_path loss_path \
    --dimension 600

Tools for computation in Riemannian geometry are not developed by us. One can extend his model to hyperbolic space with these tools. Thanks to these good repos.

• hyrnn can be found in hyrnn

• geoopt can be found in geoopt

• poincare_glove can be found in poincare_glove

• poincare-embeddings can be found in poincare-embeddings

Code refactoring is still continuing