Purestochastic

Purestochastic lets you deal with the uncertainty associated with the prediction of your machine learning model. It makes you able to use all the tools of the Tensorflow library and adds functionnality to be able to construct stochastic model easily. You can use a low-level interface by constructing model on your own way or a high-level interface by converting standard model into stochastic model.

⚠️ At present, the library is only constructed for the regression problem. At future, it will may be adapted to the classification problem.

Table of Contents

• Documentation : This section contains some ways to get few docs.
• Examples : This section will describe to you how you can use purestochastic.
• Installation : This section will explain to you how can you install this library
• Organisation : This section will explain you the structure of the github repository.

Documentation

The documentation is available at here at readthedocs. If an information is not available in the documentation, you can go directly in the source code.

Examples

The library enables to convert standard deterministic keras.Model into stochastic model with high-level methods. Suppose you have constructed a standard model from keras API.

inputs = Input(shape=(1,))
x = Dense(100, activation="relu")(inputs)
outputs = Dense(1)(x)
model = Model(inputs=inputs, outputs=outputs)

The library let you convert this model to a Deep Ensemble model with only one line of code.

from purestochastic.model.deep_ensemble import toDeepEnsemble
deep_ensemble = toDeepEnsemble(model, nb_models=5)
deep_ensemble.summary()
    _________________________________________________________________
    Layer (type)                        Output Shape        Param #   
    =================================================================
    input (InputLayer)                  [(None, 1)]         0                                                            
    ensemble_hidden_layer (Dense2Dto3D)  (None, 5, 100)     1000                                                                                           
    ensemble_output (Dense3Dto3D)        (None, 5, 1)       505                                                                                                                        
    =================================================================
    Total params: 1,505
    Trainable params: 1,505
    Non-trainable params: 0
    _________________________________________________________________

The model has been converted and it is a now a Deep Ensemble. Thus, it is possible to quantify the uncertainty. For more examples, you can go the documentation in the Examples section.

Installation

Firstly you need to clone this repo on your computer with this git command :

git clone https://github.com/purecontrol/purestochastic.git

Then it's recommended to install dependencies using pip3 tool :

pip3 install -r requirements.txt

Finally, you can install purestochastic with pip3 tool :

pip3 install .

Then, you can use it in your code with this import :

import purestochastic as ps

Organisation

The repository is organized as follows :

• \docs documentation files
• \examples simple examples on how to use the library
• \purestochastic purestochastic package (each method or class is documented)
  • \common module with new losses and metrics
    • losses.py stochastic losses
    • metrics.py stochastic metrics
  • model tools to create stochastic models
    • activations.py custom activations
    • layers.py custom layers
    • base_uncertainty_models.py custom class StochasticModel and GaussianRegression
    • deep_ensemble.py tools for the Deep Ensemble model
    • swag.py tools for the SWAG and the MultiSWAG model
    • orthonormal_certificates.py tools for Orthonormal Certificates model
• requirements.txt all the package you need to install so that you can use the project on your own