SeonghoBaek / CPC

CPC based Representation Learning

Contrastive Predictive Coding

Tensorflow Implementation of CPC v2: Data Efficient Image Recognition with CPC

Reference Paper: https://arxiv.org/abs/1905.09272

Unsupervised pretraining based on CPC encoding.

Model Summary

Input: [96 x 96 x 3] Patch: [24 x 24 x 3] Overlap size 12. 1 input image is composed with 49 patches.

Flow: Input Image -> Make Patches -> Encoding -> Pixel CNN -> CPC

Encoder: ResDense Block + Global AVG Pool, No Pooling Layer(Conv Only), Batch Norm(Fine Tune Only) and Self Attention.

ResDense Block is dense convolution block with residual connection(See JointCenterLoss).

Contact: [e-mail]