https://github.com/arkaung/dropout_bayesian_nn

Demonstration for using dropout as a means of bayesian approximation in Deep Neural Network
https://github.com/arkaung/dropout_bayesian_nn

bayesian deep-neural-networks

Last synced: 21 days ago
JSON representation

Demonstration for using dropout as a means of bayesian approximation in Deep Neural Network

• Host: GitHub
• URL: https://github.com/arkaung/dropout_bayesian_nn
• Owner: ArkAung
• Created: 2020-05-12T02:58:35.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2020-05-13T07:28:09.000Z (over 4 years ago)
• Last Synced: 2024-11-17T23:10:13.584Z (3 months ago)
• Topics: bayesian, deep-neural-networks
• Language: Jupyter Notebook
• Size: 10.6 MB
• Stars: 3
• Watchers: 2
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Dropout as Bayesian Approximation for Deep Neural Network

[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ArkAung/dropout_bayesian_nn)

You can run the whole notebook in Colab without issue. 

Getting helper scripts from Github repo, downloading datasets,

 and install required packages are all handled. 

### Running on local machine

* Clone repository and change directory to local copy

* Create a conda environment `conda create --name bayesian_nn`

* Activate conda environment `conda activate bayesian_nn`

* Install required libraries `pip install -r requirements.txt`

* Run jupyter lab `jupyter lab`

* Open `Dropout as Bayesian Approximation.ipynb`

## Visualizing Model Uncertainty

Data is passed through the model with `Dropout` turned on 

during the inference step for multiple times. The resulting 

probability distributions are what the model produce over 

multiple passes.

When the model is uncertain:

![Uncertain](assets/uncertain.png)

When the model is certain:

![Certain](assets/certain.png)

High certainty:

![High_Certain](assets/high_certainty.png)

Extremely high certainty:

![Ext_Certain](assets/extremely_high_certainty.png)