blog

import matplotlib.pyplot as plt
import numpy as np

# 自作クラス
from mlp import (
    ActivationLayer,
    FullyConnectedLayer,
    InputLayer,
    MeanSquaredError,
    MultiLayerPerceptron,
    Sigmoid,
)
from numpy.random import MT19937, Generator, SeedSequence

np.set_printoptions(precision=3)

data = np.array(
    [
        [0, 0, 0],
        [1, 0, 1],
        [0, 1, 1],
        [1, 1, 0],
    ]
)

x = data[:, :-1]
y = data[:, -1:]


def model(activation, rng):
    mlp = MultiLayerPerceptron(
        [
            InputLayer(2),
            FullyConnectedLayer(2, 2, rng),
            ActivationLayer(Sigmoid()),
            FullyConnectedLayer(2, 1, rng),
            ActivationLayer(Sigmoid()),
        ],
        MeanSquaredError(),
        epochs=1000,
        learning_rate=1,
    )
    return mlp


fig = plt.figure(figsize=(15, 160))

seeds = 200
cols = 4
rows = seeds // cols

for seed in range(seeds):

    mlp_sigm_mt1 = model(Sigmoid(), Generator(MT19937(seed)))
    mlp_sigm_mt1.fit(x, y)

    mlp_sigm_pcg1 = model(Sigmoid(), np.random.default_rng(seed))
    mlp_sigm_pcg1.fit(x, y)

    ax = fig.add_subplot(rows, cols, seed + 1)
    ax.set_title("XOR {}".format(seed))
    ax.set_xlabel("epoch")
    ax.set_ylabel("cost")
    ax.plot(
        mlp_sigm_mt1.history,
        marker="o",
        markersize=1,
        alpha=0.3,
        label=r"$\sigma, \; MT19937$",
    )
    ax.plot(
        mlp_sigm_pcg1.history,
        marker="o",
        markersize=1,
        alpha=0.3,
        label=r"$\sigma, \; PCG64$",
    )
    ax.set_ylim(-0.1, ax.get_ylim()[1])
    ax.legend()
    ax.grid()
    ax.minorticks_on()
plt.tight_layout()
plt.show()