deepclustering 0.0.3

No project description provided

deep-clustering-toolbox

PyTorch Vision toolbox not only for deep-clustering

Introduction

I still use this repo for research propose. I update some modules frequently to make the framework flexible enough.

This repo contains the base code for a deep learning framework using PyTorch, to benchmark algorithms for various dataset. The current version supports MNIST, CIFAR10, SVHN and STL-10 for semisupervised and unsupervised learning. ACDC, Promise12, WMH and so on are supported as segmentation counterpart.

Features:

• Powerful cmd parser using yaml module, providing flexible input formats without predefined argparser.
• Automatic checkpoint management adapting to various settings
• Automatic meter recording and experimental status plotting using matplotlib and threads
• Various build-in loss functions and help tricks and assert statements frequently used in PyTorch Framework, such as disable_tracking_bn, ema, vat, etc.
• Various post-processing tools such as Viewer for Medical image segmentations, multislice_viwers for 3D dataset real-time debug and report script for experimental summaries.
• Extendable modules for rapid development.

Several projects are benefited from this scalable framework, builing top on this, including:

• DeepClustering implemented for

• Invariant Information Clustering for Unsupervised Image Classification and Segmentation,
• Learning Discrete Representations via Information Maximizing Self-Augmented Training,
• Information based Deep Clustering: An experimental study

• SemiSupervised classification for

• Semi-Supervised Learning by Augmented Distribution Alignment,
• Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning,
• Temporal Ensembling for Semi-Supervised Learning,
• Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

• SemiSupervised Segmentation for

• Adversarial Learning for Semi-Supervised Semantic Segmentation,
• Semi-Supervised and Task-Driven Data Augmentation,
• Deep Co-Training for Semi-Supervised Image Segmentation

• Discretely-constrained CNN for

• Discretely-constrained deep network for weakly-supervised segmentation,
• Mutual information based segmentation on medical imaging

They are examples how to develop research framework with the assistance of our proposed deep-clustering-toolbox.

---

Playground

Several papers have been implemented based on this framework. I store them in the playground folder. The papers include:

• Auto-Encoding Variational Bayes
• mixup: BEYOND EMPIRICAL RISK MINIMIZATION
• MINE: Mutual Information Neural Estimation
• Averaging Weights Leads to Wider Optima and Better Generalization
• THERE ARE MANY CONSISTENT EXPLANATIONS OF UNLABELED DATA: WHY YOU SHOULD AVERAGE
• Prior-aware Neural Network for Partially-Supervised Multi-Organ Segmentation

---

Installation

git clone https://github.com/jizongFox/deep-clustering-toolbox.git
cd deep-clustering-toolbox  
python setup install # for those who do not want to make changes immediately.
# or
python setup develop # for those who want to modify the code and make the impact immediate.

Or very simply

pip install deepclustering

Citation

If you feel useful for your project, please consider citing this work.

@article{peng2019deep,
  title={Deep Co-Training for Semi-Supervised Image Segmentation},
  author={Peng, Jizong and Estradab, Guillermo and Pedersoli, Marco and Desrosiers, Christian},
  journal={arXiv preprint arXiv:1903.11233},
  year={2019}
}