/contrastivelearning

Implementation of SimCLR model using pytorch for using contrastive learning over shapenet dataset

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

Contrastive Learning using Point Cloud

Model

This repository explores the task of contrastive learning applied to 3d point clouds using SimCLR model.

  • SimCLR MODEL:
    simclr-general-architecture
  • NTXENT Loss: image The above image is the mathematical formula for NTXENT Loss introduced by SimCLR model. The numerator comprises of taking consine similarity of the the positive pairs which would lie between -1 to 1. It is then divided by a constant T or temperature and exponential of the value is taken to make it positive. The denominator comprises of summision of exponential of cosine similarity of negative pairs from 0 to 2N ( 2N for augmented pairs ). We dont sum the denominator for k=i ( i.e same as numerator ). Logarithm of whole thing is taken to soften the loss followed by a negative sign.

Results

image image image image