search

parya-j / team-rope

This repository contains materials of final project for data science practicum offered at University of Georgia by Dr. Quinn Spring 2018
MIT License
0 stars 1 forks source link

clustering model #2

Open omid-s opened 6 years ago

omid-s commented 6 years ago

we need a clustring model we can assume 1000 classes for now

omid-s commented 6 years ago

used a Keras Implemntation for DEC algorithm results : Horrible! https://github.com/XifengGuo/DEC-keras

raun1 commented 6 years ago

Well you are trying to cluster 15k classes :)

omid-s commented 6 years ago

implemented another clustering using KMeans :) to took a long time to train :D but it's done. I chose 1000 as # of classes for now

raun1 commented 6 years ago

Maybe fewer clusters would be better. I mean the 15k will be cus of different types. Monuments = monuments , doesnt matter if the rome thing or .... Yea i dont know much monuments LOL. But the point is maybe reduce it to 100 classes :? I mean 1k classes might be good too. I have no idea lol. We need to get the ensemble up asap then we can tweak. On an average how long does the clustering algorithm take? Dense net will take a day or so to train.

omid-s commented 6 years ago

took around 10h for a 120K sample size

raun1 commented 6 years ago

Nice then we dont have to worry about time thing. I will have the updated dense net trained by tomorrow night.

raun1 commented 6 years ago

BTW We can simply use the old dense net pretrained weights :) to tweak the clustering algorithm. Do we have some result from the pretrained networks @parya-j . In that case we can fix those slots and tweak the ensemble for the clustering bit.

omid-s commented 6 years ago

just rerun it for 100 classes