MiOC Official Implementation

The official implementattion of the paper "One-Class SVM-guided Negative Mixing for Enhanced Contrastive Learning". The paper is availabe at OpenReview.

• This repository is based on MoCo Official Implementation

Requirements

pip install numpy torch blobfile tqdm pyYaml pillow jaxtyping beartype pytorch-lightning omegaconf hydra-core wandb

Mixing One-Class SVM Negative Samples - MiOC

• We use One-Class SVM to find the negative samples that are most similar to the query.

• we generate $S_n$, synthetic negative samples by mixing a random query $z^q_i$ with random negative samples $z^-_i$.

• we generate the $S_o$, synthetic negative samples by mixing a random query $z^q_i$ with inner One-Class SVM negative samples $z^-_i$.

Please refer to the paper for more details.

The code is present in the Code folder. Please go through the Instructions for the implementation details.

Here are the linear eval results for the Imagenet-100 dataset along with the TSNE plot of the features in Cifar-10 dataset.

See Also

We would recommend to read the

• MoCo Article.
• MoCHI Article.

License

This repository is released under the MIT license.

Contributing to MiOC

Please see this link - How to CONTRIBUTE for more details.

Code of Conduct

Please see the CODE OF CONDUCT for more details.