Your one-stop hub for all Machine Learning related affairs 😍
MLhub is a low code machine learning tool that allows you to automatically train and benchmark machine learning modes without typing a single line of code.
Excited? Jump straight into the Examples!
Clone the repository
git clone git@github.com:Revuj/MLhub.git
Setup a virtual environment
python3 -m venv env
Install the requirements
pip install -r requirements.txt
python3 src/main.py --specs <models definition path> [--dockerize]
- models definition path - path to the JSON file containing the definition of the models to be executed.
- dockerize flag - boolean that enables a dockerized execution for the application (default is False)
Machine Learning is now more present than ever in our society. It completly changed the game for various industry sectors, and if you're not making any use of it, you're probably going to stay behind.
However, machine learning isn't easy to understand and hard to use as writting code for it can be a pain in the ass, but ideally, everyone should be able to benefit from this type of techonology, even if you don't know how to write a single line of code!
This is where MLhub comes into place.
MLhub allows you to train multiple machine learning models, for regression and classification problems, without writing a single line of code. In fact, the only thing you need is a JSON file where you define the location of your datasets and the models that you want to test. Then you just need to wait for some MLhub magic happening behind the scenes ✨.
MLhub gives you the freedom to fully customize your models, in case you know what you're actually doing, or you can also just pick your model and leave the rest to us (like in the example above)!
If you want to costumize them, you should be able to use any parameter that is used in the sklearn models. Here's an example for a Decision Tree Regressor:
MLhub supports multiple machine learning problems, such as regression problems, classification problems and more specific problems like image classification using Convolutional Neural Networks.
For Regression problems, MLhub supports the following algorithms:
For classification problems, MLhub supports the following algorithms:
Mlhub also supports image classification using Convolutional Neural Network, allowing you to define the set of layers that compose the model, the loss and the optimizer used to optimize the model and also the training settings such as the epochs and batch size to use during the train.
Our tool allows you to automatically train models and provide you results that allow you to evaluate their performance. For regression problems we measure metrics such as r2 score, mean squared error and mean absolute error. For classification problems we measure metrics such as accuracy, precision, recall and f1 score. In both cases we also measure the CPU time so that you also have some information about performance.
For each model that you have defined on your JSON file, we provide you two python notebooks (one that was executed already and one that wasn't), with everything that we ran behind the scenes, so you can understand what happened and modify something if you want.
We also generate a CSV file and special report that hold a statistical comparison between the different models.
The program is executed through the shell using just one argument which is the path for the JSON file containing the data paths and models' definitions.
This module is responsible for parsing and validating the JSON file. This is done using the python jsonschema library. After that, it guarantees asynchronous execution by creating a pool of processes (one for each module) and then passing it the necessary information for the Generator module.
Take a look at our Schema in more detail to get started with our tool.
Constructs and executes reports in the form of csv and python notebook files. Since most notebooks have a significant amount of common code, we utilized the jinja2 template engine and its inheritance capabilities as a way to render the final notebooks efficiently and with style. All of the notebooks are executed using nbconvert preprocessors.
Both the Parser and Generator modules can be executed inside a Docker container so that our application runs smoothly, quickly and securely in any system.
If the flag --dockerize is passed, the notebooks will be generated and run in a dockerized environment.
The used docker image was created specifically for the the project and contains all the dependencies of MLhub.
We believe this project fits the TACS course perfectly as it represents a nice example of Metaprogramming and Model-Driven Engineering for two different reasons:
- we have made use of a DSL (Domain Specific Language) in order to define the different models and its parameters using a JSON file
- our tool performs source-code generation to construct the different reports that are responsible for training and evaluating machine learning models.
MLhub can also be categorized as an Automatic Programming tool as it is capable of synthesizing machine learning programs with just a few simple inputs that anyone is capable of writting, even those with very limited machine learning knowledge.
- Support more types of problems (e.g. NLP)
- Increase CNN customization (e.g. support more layer and more parameters)
- Have more Auto ML features (e.g. perform grid search to find the best parameters for each model)
- Build a user interface to provide an alternative for composing the JSON input file
- Rafa Varela
- José Silva
- Vítor Barbosa
The project is MIT licensed.