Template using hyperOpt and Mlflow with mm-series computer vision

MLflow

An open source platform for the end-to-end machine learning lifecycle. You can setup your own repository and dashboard and keep tracking the performance and metrics of different models

Hyperopt

HyperOpt is an open-source Python library for Bayesian optimization developed by James Bergstra. It is designed for large-scale optimization for models with hundreds of parameters and allows the optimization procedure to be scaled across multiple cores and multiple machines.

OpenMMLab

OpenMMLab builds the most influential open-source computer vision algorithm system in the deep learning era.

This repo I give a simple demo how to use MLflow & Hyperopt & MMpretrain. So, you can build your MLops pipeline solely using opensource.

Step 1. Setup custom MLflowHook
Step 2. Clone this repo on top of your MMpretrain repo (it can be any MM-series).

git clone https://github.com/ccomkhj/hyperoptmm.git hyperoptmm
mv hyperoptmm path/to/mmpretrain
rm -rf hyperoptmm

Step 3. Write down relevant environments onto run_train.sh
Step 4. Setup your config file.
Step 5. sh run_train.sh

In this repo, optimizer is tuned. You can setup anything you like.