Discriminability as an Affinity Measure for Large-Scale Hierarchical Image Clustering

This reposority contains the experiments that I have conducted for my master thesis at University of Passau.

To run this code you need to install the following dependencies:

• Python 3.6
• PyTorch >=1.0.0
• faiss >= 1.5.0
• scikit-learn >=0.20.0
• Jupyter Notebook

Further you need a copy of the ImageNet dataset, which must be accessible at datasets/imagenet-full.

The experiments which depend on DeepCluster embeddings also require the download of the DeepCluster pretrained model. To download the pretrained model simply execute:

cd deepcluster
download_model.sh

Discriminability vs. DeepCluster

To run the first experiment series, which compares Discriminability against the DeepCluster embeddings run the following notebooks:

• create-imagenet-subsets.ipynb
• imagenet-subsets.ipynb
• imagenet-subsets-eval.ipynb

Artificial Initial Clustering

The second experiment, which explores the influence of noise and initial cluster size on the quality of the affinity measure, can be executed with the following notebooks:

• imagenet-largescale.ipynb
• imagenet-largescale-eval.ipynb

This experiment does not require the pretrained DeepCluster model.

Benefit of Retraining

The third experiment, which explores whether their is a benefit from recomputing the affinity measure by retraining the neural network after a merge decision can be reproduced with the notebooks:

• retraining.ipynb
• retraining-eval.ipynb