Super basic multilayer perceptron neural network library built from scratch, with love, using only Python and numpy
'If you wish to make an apple pie from scratch, you must first invent the universe' -- Carl Sagan
I trained the network on the first 50,000 images of MNIST, and tested it on the last 10,000, and I was able to get ~98% accuracy.
from MNIST import labels, images
from Network import Network
training = images[:50_000]
labels = labels[:50_000]
probe_training = images[50_000:]
probe_labels = labels[50_000:]
nn = Network([784, 40, 40, 10], training=training, labels=labels, probe_training=probe_training, probe_labels=probe_labels)
nn.train(epochs=15, batch_size=15, learning_rate=0.1)epoch 1/15 - accuracy: 94.04% - loss: 2886.80
epoch 2/15 - accuracy: 95.78% - loss: 2083.07
epoch 3/15 - accuracy: 96.65% - loss: 1660.40
epoch 4/15 - accuracy: 97.13% - loss: 1586.17
epoch 5/15 - accuracy: 97.50% - loss: 1265.56
epoch 6/15 - accuracy: 97.86% - loss: 1110.86
epoch 7/15 - accuracy: 97.80% - loss: 1191.73
epoch 8/15 - accuracy: 98.21% - loss: 1064.28
epoch 9/15 - accuracy: 98.00% - loss: 1013.17
epoch 10/15 - accuracy: 98.55% - loss: 766.64
epoch 11/15 - accuracy: 97.52% - loss: 1273.52
epoch 12/15 - accuracy: 98.68% - loss: 728.28
epoch 13/15 - accuracy: 98.70% - loss: 729.98
epoch 14/15 - accuracy: 98.77% - loss: 684.22
epoch 15/15 - accuracy: 98.68% - loss: 747.66
nn.accuracy()98.68
from matplotlib import pyplot as plt
import numpy as np
for i in range(30, 35):
img = training[i]
plt.imshow(np.reshape(img, (28, 28)))
plt.show()
print("Digit is:", np.argmax(nn.classify(img)))
print("\n\n")
Digit is: 3
Digit is: 8
Digit is: 6
Digit is: 9
Digit is: 0