Skip to content

Latest commit

 

History

History
38 lines (31 loc) · 2.2 KB

README.md

File metadata and controls

38 lines (31 loc) · 2.2 KB
Raw

Cluster-loss-Tensorflow

This repository is an implementation of Deep Metric Learning via Facility Location on tensorflow. We build this on Cifar100 and Densenet-40. This paper is available here. For the loss layer implementation, look at here. For the Densenet implementation, look at here.

@inproceedings{songCVPR17,
  Author    = {Hyun Oh Song and Stefanie Jegelka and Vivek Rathod and Kevin Murphy},
  Title     = {Deep Metric Learning via Facility Location},
  Booktitle = {Computer Vision and Pattern Recognition (CVPR)},
  Year      = {2017}
}

Installation

  1. Install prerequsites for tensorflow (see: tensorflow-gpu installation instructions).
  2. Run pip install -r requirements.txt get required support.

Training Procedure

  1. Modify metric_learning_densenet.py for training-params and densenet-params. We pick Cifar100 as our training data, because it's tiny, save GPU-memory (when batch size 64, it cost about 4.6G GPU-Memory) and good for doing research.
  2. Run python metric_learning_densenet.py, the data_provider with automaticlly handle data download and process. After that, start Densenet-Cluster-loss training.
  3. Download Downsampled Imagenet with size 32x32 from here. Modify metric_learning_densenet.py train on Imagenet.

Feature Extraction after Training

  1. Modify metric_learning_densenet.py extract feature embeddings on cifar test set, the embeddings is saved with .npy format used for evaluation process.

Clustering and Retrieval Evaluation

  1. Run python visualization/tsne.py can plot and save the cluster result on Cifar database. tSNE

Repository Information

  • Densenet tensorflow training code
  • Deep metric learning cluster loss code
  • Evaludation
    • NMI, Recall@K code
    • feature extraction code
    • feature visulization code (tSNE)
  • Dataset support
    • cifar-10
    • cifar-100
    • imagenet-32x32