Deep Clustering with Incomplete Noisy Pairwise Annotations: A Geometric Regularization Approach

This repo provides Python implementation of the work "Deep Clustering with Incomplete Noisy Pairwise Annotations: A Geometric Regularization Approach", accepted to ICML 2023.

Setup

• Clone this project: git clone https://github.com/ductri/VolMaxDCC
• cd VolMaxDCC
• (Optional but strongly suggest) Create a separate virtual environment using python3: python3 -m venv localenv, then activate it source localenv/bin/activate
• Install all package listed in requirements.txt: pip install -r requirements.txt

Training

We have set a default configution. Without any change, you should able to perform:

• Training the experiment on ImageNet10 in noiseless pairwise setting:

python src/our_model__train_demo_imagenet10.py

This will load a pairwise labels dataset that have been drawn randomly and stored to datasets/. This dataset containing 10k pairs drawn randomly from the training part of ImageNet10.

• Evaluate performance in terms of ACC, NMI, and ARI:

python src/our_model__eval.py

This will evaluate the learned mapping f using 2k test dataset.

Other options:

• You can create different pairwise dataset by inspecting file imagenet10_create_pair.py, similarly for stl10 and cifar10.

Datasets

Real pairwise annotations

• Create ImageNet-10 from ImageNet with the following classes:

n02056570
n02085936
n02128757
n02690373
n02692877
n03095699
n04254680
n04285008
n04467665
n07747607

Download all images of these class and store them in some directory, for example imagenet10/raw. The folder should look like

imagenet10/raw
           |---n02056570/
           |---n02085936/
           |---n02128757/
           |---n02690373/
           |---n02692877/
           |---n03095699/
           |---n04254680/
           |---n04285008/
           |---n04467665/
           |---n07747607/

• Load the dataset using

dataset = torchvision.datasets.ImageFolder('imagenet10/raw')

The order of this dataset matters.

• Load pairwise indices

import pickle as pkl
with open('datasets/imagenet10/pairs/pair_8994_real_imagenet10-cc-10k_trial_0.pkl', 'rb') as i_f:
    pairwise_data = pkl.load(i_f)

pairwise_data is a dict with following keys:

• shuffle_inds: a permutation of range(13000). This permutation is used to shuffle order of the dataset.
• ind_pairs: list of 8994 pairs of indices with respect to shuffle_inds.
• label_pairs: a list of 8994 pairwise labels of ind_pairs, annotated by AMT workers.
• true_label_pairs: a list of 8994 pairwise labels of ind_pairs, but inferred from ground truth class label.
• X: feature vector extracted from a pretrained unsupervised method.
• true_y: corresponding class label of the data X.