https://github.com/IcewineChen/mxnet-batch_hard_triplet_loss

mxnet version batch hard triplet loss
https://github.com/IcewineChen/mxnet-batch_hard_triplet_loss

deep-learning mxnet python3

Last synced: 2 months ago
JSON representation

mxnet version batch hard triplet loss

• Host: GitHub
• URL: https://github.com/IcewineChen/mxnet-batch_hard_triplet_loss
• Owner: IcewineChen
• License: mit
• Created: 2018-07-01T15:53:15.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2018-08-30T07:05:35.000Z (over 6 years ago)
• Last Synced: 2024-08-01T22:40:56.499Z (5 months ago)
• Topics: deep-learning, mxnet, python3
• Language: Python
• Homepage:
• Size: 51.8 KB
• Stars: 13
• Watchers: 2
• Forks: 3
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• Awesome-MXNet - batch_hard_triplet_loss

README

        
Mxnet-version batch hard triplet loss

---

Based on  (https://arxiv.org/abs/1703.07737) 

Based on some tricks from omoindrot's repository. (https://github.com/omoindrot/tensorflow-triplet-loss)

## Introduction

-  In this paper,authors propose a new format of triplet loss called batch hard. For more details of batch hard triplet loss, you can read  (https://arxiv.org/abs/1703.07737).

-  More efficient than triplet loss which proposed by facenet. More details in (https://arxiv.org/abs/1503.03832) 

-  Inplement the hard mining method and soft-margin

-  Can be used in many tasks. Firstly I code this to do some research on re-id and image retrieval tasks.

-  In the future maybe add batch all triplet loss. Compared to batch hard, sometimes it can make the experiment more efficient.

## Architecture

1. Using resnetV2 to get 128-dimension embeddings

1. Using triplet loss to train embeddings

1. the network is defined in resnet.py

## Requirements

The code has been tested with CUDA 8.0 and ubuntu 16.04.

- python3

- mxnet-cu80==1.3

how to train:\

See parsers in train.py. Then Set your dataset path and some params of based resnet network.\

The network has been defined in resnet.py.Batch_hard.py now has been deprecated.