BaseBayesianClassifier.py

import numpy as np
import pandas as pd
import numpy.linalg as LA
from scipy.linalg import cholesky, solve_triangular
from scipy.linalg.lapack import dtrtri


class BaseBayesianClassifier:
  def __init__(self):
    pass

  def _estimate_a_priori(self, y):
    a_priori = np.bincount(y.flatten().astype(int)) / y.size
    # Q3: para que sirve bincount?
    return np.log(a_priori)

  def _fit_params(self, X, y):
    # estimate all needed parameters for given model
    raise NotImplementedError()

  def _predict_log_conditional(self, x, class_idx):
    # predict the log(P(x|G=class_idx)), the log of the conditional probability of x given the class
    # this should depend on the model used
    raise NotImplementedError()

  def fit(self, X, y, a_priori=None):
    # if it's needed, estimate a priori probabilities
    self.log_a_priori = self._estimate_a_priori(y) if a_priori is None else np.log(a_priori)

    # now that everything else is in place, estimate all needed parameters for given model
    self._fit_params(X, y)
    # Q4: por que el _fit_params va al final? no se puede mover a, por ejemplo, antes de la priori?

  def predict(self, X):
    # this is actually an individual prediction encased in a for-loop
    m_obs = X.shape[1]
    y_hat = np.empty(m_obs, dtype=int)

    for i in range(m_obs):
      y_hat[i] = self._predict_one(X[:,i].reshape(-1,1))

    # return prediction as a row vector (matching y)
    return y_hat.reshape(1,-1)

  def _predict_one(self, x):
    # calculate all log posteriori probabilities (actually, +C)
    log_posteriori = [ log_a_priori_i + self._predict_log_conditional(x, idx) for idx, log_a_priori_i
                  in enumerate(self.log_a_priori) ]

    # return the class that has maximum a posteriori probability
    return np.argmax(log_posteriori)