logistic_regression.py

# -*- coding: utf-8 -*-
# logistic regression on MNIST binary image classification.

# I have completed the major framework, you can fill the space 
# to make it work.
# By the way, achieving the logistic regression with Numpy 
# is a classical coding problem when you confront to 
# look for an intern in some AI labs. 
# So u may treat this as the rehearsal.
# Gooood luck for your first programming assignment. ;)

__author__ = "Zifeng Wang"
__email__ = "wangzf18@mails.tsinghua.edu.cn"
__date__ = "20190920"

import numpy as np
import pdb
import os

# import your own tools
from load_mnist import load_mnist, filter_dataset

np.random.seed(2019)


class logistic_regression:
    """To develop a classifier,
    u may build its two basic functions first:
    the train and predict.
    """

    def __init__(self):
        self.acc_func = lambda x, y: 100 * np.sum(x == y) / x.shape[0]

    def train(self, x_train, y_train):
        """Receive the input training data, then learn the model.
        Inputs:
        x_train: np.array, shape (num_samples, num_features)
        y_train: np.array, shape (num_samples, )

        Outputs：
        None
        """
        self.w = np.random.randn(784)
        self.learning_rate = 0.1

        # update the parameters
        for i in range(100):
            """
            Please Fill Your Code Here.
            """
            self.w = self.w + self.learning_rate * (y_train[i] - 1/(1+np.exp(-self.w.dot(x_train[i])))) * x_train[i]

    def predict(self, x_test):
        """Do prediction via the learned model.
        Inputs:
        x_test: np.array, shape (num_samples, num_features)

        Outputs:
        pred: np.array, shape (num_samples, )
        """

        """
        Please Fill Your Code Here.
        """
        pred = x_test.dot(self.w)
        return pred


if __name__ == '__main__':
    # load data
    x_train, y_train, x_test, y_test = load_mnist()
    x_train, y_train = filter_dataset(x_train, y_train, 1, 7)
    x_test, y_test = filter_dataset(x_test, y_test, 1, 7)

    # train ur classifier
    lr = logistic_regression()
    lr.train(x_train, y_train)
    y_test_pred = lr.predict(x_test)

    # evaluate the prediction
    y_test_pred[y_test_pred > 0.5] = 1.0
    y_test_pred[y_test_pred <= 0.5] = 0.0
    print(np.bincount(y_test_pred.astype(int)))

    test_acc = lr.acc_func(y_test, y_test_pred)
    print("Your model acquires test acc: {:.4f} %".format(test_acc))

    if test_acc >= 95:
        print("Congratulations! Your classifier works!")
    else:
        print("Check your code! Sth went wrong.")