Whats this about ?

Detailed code on applying a novel few shot clustering technique (EM style) to cluster images using very few actual labels (few shot clustering) State of the art accuracy acheieved in ImageNet 5-way 5-shot Demonstrated application in few shot building occupancy detection

Published paper here -> https://arxiv.org/abs/2008.05654

Complete Dataset See Additional_datasets.txt

To run

python main.py

(all the parameters of the code present in the top few lines of main.py, explained with comments)

Dependencies:

pytorch

opencv

numpy

Dataset directory structure

Directories need to be created :

data/labeled/0

    /1

...

...

    /n

(the few labeled images that you have)

(depending upon number of classes present in the few shot learning problem)

data/unlabeled/

(dump all the unlabeled images you want to cluster here)

data/validation/0

...

       /n

(same structure as data/labaled, this folder images used by the algorithm to track convergence progress if you dont have enough annotations for this folder, just comment out the validate() function in main.py)

data/model_pred/0

...

       /n

(same structure as data/labeled, here the model will store the clustering results in respective folders as the EM algorithm progresses)