ORCA-python is an experimental framework, completely built on Python (integrated with scikit-learn and sacred modules), that seeks to automatize the run of machine learning experiments through simple-to-understand configuration files.
ORCA-python has been initially created to test ordinal classification, but it can handle regular classification algorithms, as long as they are implemented in scikit-learn, or self-implemented following compatibility guidelines form scikit-learn.
In this README, we will explain how to use ORCA-python, and what you need to install in order to run it. A Jupyter notebook is also avaible in spanish.
ORCA-python has been developed and tested in GNU/Linux systems. It has been tested with Python 2.7.13 and Python 3.5.3.
Besides the need for the aforementioned Python interpreter, you will need to install the next Python modules in order to run an experiment (needs recent versions of scikit-learn >=0.20.0):
- numpy (tested with version 1.18.1)
- pandas (tested with version 1.0.1)
- sacred (tested with version 0.8.1)
- scikit-learn (tested with version 0.22.1)
- scipy (tested with version 1.4.1)
To install Python, you can use the package management system you like the most.
For the installation of the modules, you may follow this Python's Official Guide.
A requirements.txt file has been added to ease the installation of the different dependencies through pip.
To download ORCA-python you can simply clone this GitHub repository by using the following commands:
$ git clone https://github.com/ayrna/orca-python
All the contents of the repository can also be downloaded from the GitHub site by using the "Download ZIP" button.
Even if ORCA-Python is written in Python, some algorithms like REDSVM or SVOREX are implemented in C++ and C. Before using the framework is necesary to compile these algorithms using the make command in the repository root.
The algorithms will be compiled for the system or virtual environment default Python interpreter. If the framework is executed with another interpreter different from the one used for compiling, the algorithms wont work.
If executing the framework with a different Python interpreter is necesary, execute the make clean command in the repository root to clean the old compilation and use the make command again.
Inside the orca-python root execute pip install . Orca-python should be compiled before installing. If you want to compile it after installing add the --editable argument: pip install --editable .
We provide a pre-made experiment (dataset and configuration file) to test if everything has been correctly installed.
The way to run this test (and all experiments) is the following:
# Go to framework main folder
$ python config.py with orca_python/configurations/full_functionality_test.json -l ERROR
This tutorial uses three small datasets (balance-scale, contact-lenses and tae) contained in "datasets" folder. The datasets are already partitioned with a 30-holdout experimental design (train and test pairs for each partition).
All experiments are run through configuration files, which are written in JSON format, and consist of two well differentiated sections:
general-conf: indicates basic information to run the experiment, such as the location to datasets, the names of the different datasets to run, etc.configurations: tells the framework what classification algorithms to apply over all the datasets, with the collection of hyper-parameters to tune.
Each one of this sections will be inside a dictionary, having the said section names as keys.
For a better understanding of the way this files works, it's better to follow an example, that can be found in: configurations/full_functionality_test.json.
"general_conf": {
"basedir": "ordinal-datasets/ordinal-regression/",
"datasets": ["tae", "balance-scale", "contact-lenses"],
"hyperparam_cv_folds": 3,
"jobs": 10,
"input_preprocessing": "std",
"output_folder": "my_runs/",
"metrics": ["ccr", "mae", "amae", "mze"],
"cv_metric": "mae"
}
note that all the keys (variable names) must be strings, while all pair: value elements are separated by commas.
basedir: folder containing all dataset subfolders, it doesn't allow more than one folder at a time. It can be indicated using a full path, or a relative one to the framework folder.datasets: name of datasets that will be experimented with. A subfolder with the same name must exist insidebasedir.hyperparam_cv_folds: number of folds used while cross-validating.jobs: number of jobs used for GridSearchCV during cross-validation.input_preprocessing: type of preprocessing to apply to the data,stdfor standardization andnormfor normalization. Assigning an empty srtring will omit the preprocessing process.output_folder: name of the folder where all experiment results will be stored.metrics: name of the accuracy metrics to measure the train and test performance of the classifier.cv_metric: error measure used for GridSearchCV to find the best set of hyper-parameters.
Most of this variables do have default values (specified in config.py), but "basedir" and "datasets" must always be written for the experiment to be run. Take into account, that all variable names in "general-conf" cannot be modified, otherwise the experiment will fail.
this dictionary will contain, at the same time, one dictionary for each configuration to try over the datasets during the experiment. This is, a classifier with some specific hyper-parameters to tune. (Keep in mind, that if two or more configurations share the same name, the later ones will be omitted)
"configurations": {
"SVM": {
"classifier": "sklearn.svm.SVC",
"parameters": {
"C": [0.001, 0.1, 1, 10, 100],
"gamma": [0.1, 1, 10]
}
},
"SVMOP": {
"classifier": "orca_python.classifiers.OrdinalDecomposition",
"parameters": {
"dtype": "ordered_partitions",
"decision_method": "frank_hall",
"base_classifier": "sklearn.svm.SVC",
"parameters": {
"C": [0.01, 0.1, 1, 10],
"gamma": [0.01, 0.1, 1, 10],
"probability": ["True"]
}
}
},
"LR": {
"classifier": "orca_python.classifiers.OrdinalDecomposition",
"parameters": {
"dtype": ["ordered_partitions", "one_vs_next"],
"decision_method": "exponential_loss",
"base_classifier": "sklearn.linear_model.LogisticRegression",
"parameters": {
"solver": ["liblinear"],
"C": [0.01, 0.1, 1, 10],
"penalty": ["l1","l2"]
}
}
},
"REDSVM": {
"classifier": "orca_python.classifiers.REDSVM",
"parameters": {
"t": 2,
"c": [0.1, 1, 10],
"g": [0.1, 1, 10],
"r": 0,
"m": 100,
"e": 0.001,
"h": 1
}
},
"SVOREX": {
"classifier": "orca_python.classifiers.SVOREX",
"parameters": {
"kernel_type": 0,
"c": [0.1, 1, 10],
"k": [0.1, 1, 10],
"t": 0.001
}
}
}
Each configuration has a name (whatever you want), and consists of:
classifier: tells the framework which classifier to use. Can be specified in two different ways:- A relative path to the classifier in sklearn module.
- The name of a built-in class in Classifiers folder (found in the main folder of the project).
parameters: hyper-parameters to tune, having each one of them a list of values to cross-validate (not really necessary, can be just one value).
In ensemble methods, as OrdinalDecomposition, you must nest another classifier (the base classifier, which doesn't have a configuration name), with it's respective parameters to tune.
As viewed in Installation Testing, running an experiment is as simple as executing Config.py with the python interpreter, and tell what configuration file to use for this experiment, resulting in the next command:
$ python config.py with experiment_file.json
Running an experiment this way has two problems though, one of them being an excessive verbosity from Sacred, while the other consists of the non-reproducibility of the results of the experiment, due to the lack of a fixed seed.
Both problems can be easily fixed. The seed can be specified after "with" in the command:
$ python config.py with experiment_file.json seed=12345
while we can silence Sacred just by adding "-l ERROR" at the end of the line (not necessarily at the end).
$ python config.py with experiment_file.json seed=12345 -l ERROR