search

shahsohil / DCC

This repository contains the source code and data for reproducing results of Deep Continuous Clustering paper
MIT License
209 stars 53 forks source link

Clarification of 'Z' and 'U' #10

Closed scottfleming closed 5 years ago

scottfleming commented 5 years ago

Just to clarify, is 'Z' the representations after SDAE but before fine tuning with DCC and 'U' is after fine tuning with DCC? Got a little confused because in the paper it appears that 'Y' is the representations after SDAE and 'Z' are the representations after fine-tuning...

shahsohil commented 5 years ago

NO. Both are the representation obtained after fine-tuning. 'Z' are the latent representation of auto-encoder. While 'U' are obtained as byproduct of application of robust loss in eq(3). 'U' more or less follows underlying topology and hence it is considered to be actual clustering representation (output).

LemonPi commented 5 years ago

@shahsohil I'm also confused about this. Is there an equation relating Z and U? The paper exclusively has equations involving Z, but the loss computation in the code only uses U

shahsohil commented 5 years ago

@LemonPi In the code 'Z' are denoted by 'enc' variable. enc, dec = net(inputs_Var)

'enc' is used for the computation of 'U'. loss2 = criterion2(enc, sampweights, pairweights, pairs, index, _sigma1, _sigma2, _lambda)