import numpy as np
import pandas as pd
import matplotlib.pyplot as plt


def load_data():
    df_orig_train = pd.read_csv('mnist.csv')
    df_digits = df_orig_train.drop('label', axis=1)

    return df_digits.to_numpy()


mnist = load_data()


def sigmoid(x):
    return 1.0 / (1.0 + np.exp(-x))  # Sigmoidfunktion


class RBM:

    def __init__(self, visible_size: int, hidden_size: int, learnrate: float = 0.1, epochs: int = 20):
        """__init__ Initializes a newly created Restricted Bolzmann Machine.

        Args:
            visible_size (int): amount of neurons inside the visible layer
            hidden_size (int): amount of neurons inside the hidden layer
            learnrate (float, optional): learnrate eta in [0;1]. Defaults to 0.1.
            epochs (int, optional): training epochs. Defaults to 20.
        """
        self.learnrate = learnrate
        self.visible_size = visible_size
        self.hidden_size = hidden_size
        self.epochs = epochs

        # initialize/reset learnable attributes
        self.weights = np.random.randn(self.visible_size, self.hidden_size)
        self.visible_bias = np.zeros(self.visible_size) * 0.1
        self.hidden_bias = np.zeros(self.hidden_size) * 0.1

    def activate(self, v0):
        return sigmoid(np.matmul(v0.T, self.weights) + self.hidden_bias)

    def reactivate(self, h0):
        return sigmoid(np.matmul(self.weights, h0.T) + self.visible_bias)

    def contrastive_divergence(self, v0, h0, v1, h1):
        # calculate gradients
        postive_gradient = np.outer(v0, h0)
        negative_gradient = np.outer(v1, h0)

        # Adjust weights by delta
        self.weights += self.learnrate * (postive_gradient - negative_gradient)

        # Adjust biases by delta
        self.visible_bias += self.learnrate * (v0 - v1)
        self.hidden_bias += self.learnrate * (h0 - h1)

    def train(self, v0):
        for _ in range(self.epochs):
            # activate hidden layer
            h0 = self.activate(v0)

            # reactivate visible layer
            v1 = self.reactivate(h0)

            # activate next hidden layer
            h1 = self.activate(v1)

            # Adjust weights
            self.contrastive_divergence(v0, h0, v1, h1)

    def run(self, v0):
        # activate hidden layer
        h0 = self.activate(v0)
        v1 = self.reactivate(h0)

        return h0, v1


rbm = RBM(28 ** 2, 100, 0.2, epochs=1)

for i in range(100):
    # normalize mnist data and train
    number = mnist[i] / 255
    rbm.train(number)

# plot results
rows, columns = (9, 9)

fig = plt.figure(figsize=(10, 7))
fig.canvas.manager.set_window_title("Reconstruction of MNIST Numbers using a Restricted Boltzmann Machine")

for i in range((rows * columns)):
    if i % 3 == 0:
        number = mnist[i] / 255
        (hidden, visible) = rbm.run(number)

        results = [hidden.reshape((10, 10)), visible.reshape((28, 28)), number.reshape((28, 28))]

        for j, item in enumerate(results):
            fig.add_subplot(rows, columns, i + j + 1)
            plt.imshow(item, cmap='gray')
            plt.axis('off')

plt.show()