import numpy as np
def sigmoid(x):
"""Compute sigmoid function for each element in x."""
return 1.0 / (1.0 + np.exp(-x))
def relu(x):
""" Compute RELU """
result = x.copy()
result[result < 0] = 0.
return result
def softmax(a):
"""
Compute softmax over the last dimension
:param a:
:return:
"""
exp_a = np.exp(a - np.max(a))
return exp_a / np.sum(exp_a, axis=-1)
class NeuralNetworkClassifier:
def __init__(self, n_input, n_hidden, n_output):
"""
:param n_input:
:param n_hidden:
:param n_output:
"""
self.n_input, self.n_hidden, self.n_output = n_input, n_hidden, n_output
self.hidden_weights = np.zeros((self.n_hidden, self.n_input))
self.output_weights = np.zeros((self.n_output, self.n_hidden))
def get_weights_shapes(self):
"""
:return: A list of 2 tuples for each layer of the network
"""
return [(self.n_hidden, self.n_input), (self.n_output, self.n_hidden)]
def set_weights(self, hidden_weights, output_weights):
self.hidden_weights[:] = hidden_weights
self.output_weights[:] = output_weights
def score(self, x, y):
"""
:param x: batch_size x n_input size
:param y: batch_size size
:return:
"""
hidden_activation = sigmoid(np.dot(self.hidden_weights, x.T)) # -> n_hidden x batch_size
output_activation = np.dot(self.output_weights, hidden_activation) # -> n_output x batch_size
output_labels = np.argmax(output_activation, axis=0) # -> batch_size
assert y.shape == output_labels.shape, "The shapes of y and output labels are %s, %s" % (y.shape, output_labels.shape)
n_correct = np.count_nonzero(y == output_labels)
n_total = len(y)
score = n_correct / n_total
return score
def main():
from sklearn.datasets import load_digits, fetch_mldata
SMALL_MNIST = False
if SMALL_MNIST:
mnist_digits = load_digits()
n_input = np.prod(mnist_digits.images.shape[1:])
n_images = len(mnist_digits.images) # 1797
data_images = mnist_digits.images.reshape(n_images, -1) / 16. # -> 1797 x 64
data_targets = mnist_digits.target
# im_size_x, im_size_y = 8, 8
else:
mnist_digits = fetch_mldata('MNIST original')
n_input = np.prod(mnist_digits.data.shape[1:])
data_images = mnist_digits.data / 255. # -> 70000 x 284
data_targets = mnist_digits.target
# im_size_x, im_size_y = 28, 28
n_hidden, n_output = 5, 10
nn = NeuralNetworkClassifier(n_input, n_hidden, n_output)
weight_shapes = nn.get_weights_shapes()
weights = []
for weight_shape in weight_shapes:
weights.append(np.random.randn(*weight_shape))
nn.set_weights(*weights)
score = nn.score(data_images, data_targets)
print("Score is: ", score)
if __name__ == '__main__':
main()
