ccp-in-pytorch

This repository implements Contrastive Credibility Propagation (CCP) in PyTorch [1]. CCP is an iterative semi-supervised learning framework that applies soft pseudolabels to unlabeled data. CCP unifies semi-supervised learning and noisy label learning for the goal of reliably outperforming a supervised baseline in any data scenario.

CCP has two stages. The first stage trains a neural network that predicts real-valued "q-vectors" for each unlabeled sample. Q-vectors characterize the extent to which that sample uniquely reflects each class. This stage is trained iteratively, with increasing amounts of originally-unlabeled data incorporated into learning the mapping from data to q-vectors. Each iteration reflects a full model training. In the second stage, CCP trains a classifier for the true task. For classification model input, rather than use the original encoding for each sample (e.g., image RGB-channel flattening or text embedding), CCP uses an interim representation that is produced through the q-vector prediction task.

This repository implements CCP from the v1 paper description as a set of PyTorch classes.

Getting started

As a user

See the usage-examples directory for examples of using this codebase.

The implementation automatically runs on a single GPU device named cuda:0 if CUDA is available. If unavailable, it defaults to using the CPU.

As a developer

One method: 1. conda create -n ccp python=3.9 -y 2. conda activate ccp 3. pip install -e .[typing,test,examples] (for package) 4. pip install -r requirements-dev.txt (for development environment) 5. pre-commit install (to get the pre-commit hooks active so you don't push unlinted/unformatted code)

And then use:

• make lint (make -i lint if you want every line to proceed)
• make format
• make test

Authors

Xavier Mignot and Pamela Toman (Palo Alto Networks)

References

[1] Brody Kutt, Pamela Toman, Xavier Mignot, Sujit Rokka Chhetri, Shan Huang, Nandini Ramanan, Min Du, William Hewlett. Contrastive Credibility Propagation for Reliable Semi-Supervised Learning. https://arxiv.org/abs/2211.09929v1