psmmettes/hpn

Hyperspherical Prototype Networks

Hyperspherical Prototype Networks

This repository contains the PyTorch code for the NeurIPS 2019 paper "Hyperspherical Prototype Networks".
The paper is available here: https://arxiv.org/abs/1901.10514



The repository includes:

• Download link for pre-computed prototypes.
• Classification scripts for CIFAR-100, ImageNet-200, and CUB Birds.
• Script to construct your own prototypes.
• Joint classification and regression script for OmniArt.

Downloading and constructing hyperspherical prototypes

To obtain prototypes pre-computed for the paper, perform the following steps:

cd prototypes/
wget -r -nH --cut-dirs=3 --no-parent --reject="index.html*" http://isis-data.science.uva.nl/mettes/hpn/prototypes/
cd ..

The folder 'sgd' denotes the prototypes without semantic priors, 'sgd-sem' with semantic priors. The folders 'sem' and 'simplex' denote the baseline prototypes of Table 1.

To create your own prototypes, use the prototypes.py script. An example run for 100 classes and 50 dimensions:

python prototypes.py -c 100 -d 50 -r prototypes/sgd/

In case you want to construct prototypes on CIFAR-100 or ImageNet-200 with word2vec representations, please download the wtv files as follows:

mkdir -p wtv
cd wtv/
wget -r -nH --cut-dirs=3 --no-parent --reject="index.html*" http://isis-data.science.uva.nl/mettes/hpn/wtv/
cd ..

Running hyperspherical prototype networks

To perform classification and joint optimization with Hyperspherical Prototype Networks, use the scripts that start with 'hpn_'.
For CIFAR-100 using 50-dimensional prototypes without semantic priors (akin to column 4 of Table 1 of the paper), run the following:

python hpn_cifar.py --datadir data/ --resdir res/ --hpnfile prototypes/sgd/prototypes-50d-100c.npy --seed 100

All the other scripts work precisely the same.

The CUB Birds dataset can be obtained from the original dataset: http://www.vision.caltech.edu/visipedia/CUB-200-2011.html
The prepared ImageNet-200 and OmniArt datasets can be obtained as follows:

cd data/
wget -r -nH --cut-dirs=3 --no-parent --reject="index.html*" http://isis-data.science.uva.nl/mettes/hpn/data/imagenet200/
wget -r -nH --cut-dirs=3 --no-parent --reject="index.html*" http://isis-data.science.uva.nl/mettes/hpn/data/omniart/
cd ..

Please cite the paper accordingly:

@inproceedings{mettes2019hyperspherical,
  title={Hyperspherical Prototype Networks},
  author={Mettes, Pascal and van der Pol, Elise and Snoek, Cees G M},
  booktitle={Advances in Neural Information Processing Systems},
  year={2019}
}

hpn's Issues

Some qusetions about the implementation of the sematic ranking loss in this code

Q1:

I find the elements in nns and others is useless in the triplets collection
`

for i in xrange(wtvv.shape[0]):

for j in xrange(len(nns[i])):      
 
    for k in xrange(len(others[i])):     
          
        triplets.append([i,j,i,k])        

triplets = np.array(triplets).astype(int)

`

Q2:
The implementation of prototype_loss_sem is different from L_{PI} in the paper. Is this a trick? Could you explain how it works?

.

Hi!Can you post pre-trained code on github? The download link only contains npy files. Thank you.

The question about the implement environment.

What version of python and torch should we use?

About args.wtvfile in prototypes.py

Hi. If I want to use the wtvfile, where can I get it? Thanks a lot!

Missing argument when computing prototypes with word embeddings

Hi, thanks for the nice repo and very interesting paper!

I am trying to recompute the prototypes on CIFAR100 with privileged information, using the following command:
python prototypes.py -c 100 -d 50 -r prototypes/sgd-sem/ -w wtv/cifar100_wtvvectors.npy

but I get this error:
Traceback (most recent call last): File "prototypes.py", line 95, in <module> nns.append(sorder[:args.nn]) AttributeError: 'Namespace' object has no attribute 'nn'

And it seems indeed that argument nn is not defined anywhere...