HoNCAML (Holistic No Code Automated Machine Learning) is a tool aimed to run automated machine learning pipelines, and specifically focused on finding the best model and hyperparameters for the problem at hand.
Following the no code paradigm, no Python knowledge is needed. There are two ways to define pipelines:
- Through the Graphical User Interface
- Through YAML configuration files
There are three types of provided pipelines.
Train a specific model with the hyperparameters specified.
- Input: A dataset for the training.
- Output: The model object stored to disk.
Use a model to generate predictions for a specific dataset.
- Input: A dataset for the test, together with a model object.
- Output: A tabular file with the predictions.
Search for the best model and hyperparameters for the dataset at hand.
- Input: A dataset for the benchmark.
- Output: Main output is a configuration file with the best model and hyperparameters, and a tabular file with the results for all configurations tested.
HoNCAML has been designed having the following aspects in mind:
- Ease of use
- Modularity
- Extensibility
- Simpler is better
HoNCAML does not assume any kind of technical knowledge, but at the same time it is designed to be extended by expert people. Therefore, its user base may range from:
-
Basic users: In terms of programming experience and/or machine learning knowledge. It would be possible for them to get results in an easy way.
-
Advanced users: It is possible to customize experiments in order to adapt to a specific use case that may be needed by an expert person.
Regarding each of the following concepts, HoNCAML supports specific sets of them; nevertheless, due to its nature, extend the library further should be not only feasible, but intuitive.
For now only data with tabular format is supported. However, HoNCAML provides special preprocessing methods if needed:
- Normalization
- One hot encoding of categorical features
At this moment, the following types of problems are supported:
- Regression
- Classification
Regarding available models, the following are supported:
- Sklearn models (ML)
- Pytorch models (DL)
To use HoNCAML, it is required to have Python >= 3.10.
To install HoNCAML, run: pip install honcaml
For a quick usage with example data and configuration, just run:
honcaml -e {example_directory}
This would create a directory containing sample data and configuration to see
how HoNCAML works in a straightforward manner. Just enter the specified
directory: cd {example_directory}
and run one of the pipelines located in
files directory. For example, a benchmark for a classification task:
honcaml -c files/classification_benchmark.yaml
To start a HoNCAML execution for a particular pipeline, first it is needed to generate the configuration file for it. It may be easy to start with a template, which is provided by the CLI itself.
In case a basic configuration file is enough, with the minimum required options, the following should be invoked:
honcaml -b {config_file} -t {pipeline_type}
On the other hand, there is the possibility of generating an advanced configuration file, with all the supported options:
honcaml -a {config_file} -t {pipeline_type}
In both cases, {config_file}
should be a path to the file containing the
configuration in yaml extension, and {pipeline_type}
one of the supported:
train, predict or benchmark.
When having a filled configuration file to run the pipeline, it is just a matter of executing it:
honcaml -c {config_file}
For example, the following basic configuration would train a default model for classification and store it.
```yaml
global:
problem_type: classification
steps:
data:
extract:
filepath: data/dataset.csv
target: class
transform:
model:
transform:
fit:
load:
filepath: default_model.sav
```
To run the HoNCAML GUI locally in a web browser tab, run the following command:
honcaml -g
It allows to execute HoNCAML by interactively selecting pipeline options, although it is possible to run a pipeline by uploading its configuration file as well.
All contributions are more than welcome! For further information, please refer to the contribution documentation.
If you find any bug, please check if there is any existing issues, and if not, open a new one with a clear description.
Should you have any inquiry regarding the library or its development, please contact the Applied Machine Learning team.