Parameter-ensemble-differential-evolution - Shows how to do parameter ensembling using differential evolution.

Last update: May 04, 2022

Overview

Ensembling parameters with differential evolution

This repository shows how to ensemble parameters of two trained neural networks using differential evolution. The steps followed are as follows:

Train two networks (architecturally same) on the same dataset (CIFAR-10 used here) but from two different random initializations.
Ensemble their weights using the following formulae:
```
w_t = w_o * ema + (1 - ema) * w_p
```
w_o and w_p represents the learned of a neural network.
Randomly initialize a network (same architecture as above) and populate its parameters w_t using the above formulae.

ema is usually chosen by the developer in an empirical manner. This project uses differential evolution to find it.

Below are the top-1 accuracies (on CIFAR-10 test set) of two individually trained two models along with their ensembled variant:

Model one: 63.23%
Model two: 63.42%
Ensembled: 63.35%

With the more conventional average prediction ensembling, I was able to get to 64.92%. This is way better than what I got by ensembling the parameters. Nevertheless, the purpose of this project was to just try out an idea.

Reproducing the results

Ensure the requirements.txt is satisfied. Then train two models with ensuring your working directory is at the root of this project:

$ git clone https://github.com/sayakpaul/parameter-ensemble-differential-evolution
$ cd parameter-ensemble-differential-evolution
$ pip install -qr requirements.txt
$ for i in `seq 1 2`; python train.py; done

Then just follow the ensemble-parameters.ipynb notebook. You can also use the networks I trained. Instructions are available inside the notebook.

Parameter-ensemble-differential-evolution - Shows how to do parameter ensembling using differential evolution.

Related tags

Overview

Ensembling parameters with differential evolution

Reproducing the results

References

You might also like...

Neural Ensemble Search for Performant and Calibrated Predictions

An Ensemble of CNN (Python 3.5.1 Tensorflow 1.3 numpy 1.13)

zeus is a Python implementation of the Ensemble Slice Sampling method.

Pytorch implementation of SenFormer: Efficient Self-Ensemble Framework for Semantic Segmentation

Ensemble Knowledge Guided Sub-network Search and Fine-tuning for Filter Pruning

This Jupyter notebook shows one way to implement a simple first-order low-pass filter on sampled data in discrete time.

A fast Evolution Strategy implementation in Python

Code for the paper Task Agnostic Morphology Evolution.

Pytorch implementation of FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks

Releases(v0.1.0)

v0.1.0(Jan 2, 2022)

Owner

Sayak Paul

Official repository for the ICLR 2021 paper Evaluating the Disentanglement of Deep Generative Models with Manifold Topology

[ICCV 2021] Excavating the Potential Capacity of Self-Supervised Monocular Depth Estimation

Generative Adversarial Networks for High Energy Physics extended to a multi-layer calorimeter simulation

Rl-quickstart - Reinforcement Learning Quickstart

Fuwa-http - The http client implementation for the fuwa eco-system

Implementations of CNNs, RNNs, GANs, etc

Self-Supervised Vision Transformers Learn Visual Concepts in Histopathology (LMRL Workshop, NeurIPS 2021)

A Game-Theoretic Perspective on Risk-Sensitive Reinforcement Learning

Implementation for On Provable Benefits of Depth in Training Graph Convolutional Networks

ruptures: change point detection in Python

So-ViT: Mind Visual Tokens for Vision Transformer

MOpt-AFL provided by the paper "MOPT: Optimized Mutation Scheduling for Fuzzers"

Source code for "Pack Together: Entity and Relation Extraction with Levitated Marker"

IGCN : Image-to-graph convolutional network

Jupyter Dock is a set of Jupyter Notebooks for performing molecular docking protocols interactively, as well as visualizing, converting file formats and analyzing the results.

A python implementation of Physics-informed Spline Learning for nonlinear dynamics discovery

[ICLR 2021] "CPT: Efficient Deep Neural Network Training via Cyclic Precision" by Yonggan Fu, Han Guo, Meng Li, Xin Yang, Yining Ding, Vikas Chandra, Yingyan Lin

Adaptive FNO transformer - official Pytorch implementation

DI-smartcross - Decision Intelligence Platform for Traffic Crossing Signal Control

Code for the Paper: Conditional Variational Capsule Network for Open Set Recognition