vb690 / bayesian_ANN

Small project for exploring the usage of bayesian artificial neural networks implemented with PyMC3.

Bayesian Artificial Neural Networks

Small project for exploring the usage of bayesian artificial neural networks implemented with PyMC3.

Motivation

The aim of this project is to test the usage of PyMC3 library for creating a series of Artificial Neural Netowrks within a bayesian framework. The project also tries to provide the skeleton of an API with some basic building blocks such as layers, activation functions and likelyhood models for more easily implement different types of architectures.

Features

1. PyMC3 implementations of various Artificial Neural Networks layers "à la Keras":

   • Dense Layer
   • Embedding Layer (currently not working)
   • Vanilla RNN Layer (currently not working)

2. Flexible pirors specification.

3. Set of pre-implemented Artificial Neural Networks with flexible likelyhood specification:

   • Multilayer Perceptron
   • Autoencoder with undercomplete and denoising options
   • Word Embedding (currently not working)

Usage

Model Specification

Models can be specified stacking layers on top of each other and making sure to have a final likelyhood model.

import numpy as np

import pymc3 as pm

from modules.layers import Dense

def gaussian_lk(shape_in, input_tensor, out_shape, observed,
                prior, beta=5, **priors_kwargs):
  with pm.Model() as lk_model:

      mu = Dense(
          shape_in=shape_in,
          units=out_shape,
          layer_name='mu',
          prior=prior,
          activation='linear',
          **priors_kwargs
      )(input_tensor)

      sd = pm.HalfCauchy(
          name='sigma',
          beta=beta
      )

      out = pm.Normal(
          'y',
          mu=mu,
          sd=sd,
          observed=observed
      )

  return lk_model

X = np.random.random(size=(10000, 100))
y = np.random.random(size=(10000))
shape_in = dense.shape[1]

with pm.Model() as model:
   dense=X
   for layer_n, units in enumerate((100, 50, 25):

       dense = Dense(
           shape_in=shape_in,
           units=units,
           layer_name=layer_n,
           prior=pm.Normal,
           activation='relu',
           mu=0,
           sigma=1
       )(dense)
       shape_in = units
   
    out = likelyhood_model(
       shape_in=shape_in,
       input_tensor=dense,
       out_shape=1,
       observed=y,
       prior=pm.Normal,
       mu=0,
       sigma=1
   )

Using pre-implemented models

A more conveninet shortcut is to use pre-implemented architectures like those specificed in the Features section.

from sklearn.datasets import load_digits
from sklearn.preprocessing import MinMaxScaler

from sklearn.model_selection import StratifiedShuffleSplit

import pymc3 as pm

from modules.neural_networks import BayesianMLP


X, y = load_digits(return_X_y=True)

for tr_i, ts_i in StratifiedShuffleSplit(n_splits=1).split(X, y):
    
    X_tr, X_ts = X[tr_i], X[ts_i]
    y_tr, y_ts = y[tr_i], y[ts_i]

    scaler = MinMaxScaler()
    scaler.fit(X_tr)

    X_tr = scaler.transform(X_tr)
    X_ts = scaler.transform(X_ts)
    
categorical_perceptron = BayesianMLP(
    X=X_tr, 
    y=y_tr, 
    shape_out=10, 
    likelyhood_model='categorical_lk',
    layers=(128, 64, 32, 16), 
    activation='tanh',
    prior=pm.Normal,
    mu=0,
    sigma=0.1,
    batch_size=150
)

Notebooks

1. Binary and multiclass classification.
2. Regression with Normal and Stundent T likelyhoods.
3. Autoencoder.
4. Sampling the embedding.
5. Language Model.