mirror of
https://github.com/guezoloic/neural-network.git
synced 2026-01-25 01:34:21 +00:00
refactor(network.py): annotate Neuron constructor (comments other functions)
This commit is contained in:
226
network.py
226
network.py
@@ -1,6 +1,7 @@
|
||||
import math
|
||||
import random
|
||||
|
||||
|
||||
def sigmoid(x: float) -> float:
|
||||
return 1 / (1 + math.exp(-x))
|
||||
|
||||
@@ -9,123 +10,154 @@ def sigmoid_deriv(x: float) -> float:
|
||||
y: float = sigmoid(x)
|
||||
return y * (1 - y)
|
||||
|
||||
# neuron class
|
||||
|
||||
class Neuron:
|
||||
"""
|
||||
z : linear combination of inputs and weights plus bias (pre-activation)
|
||||
y : output of the activation function (sigmoid(z))
|
||||
w : list of weights, one for each input
|
||||
"""
|
||||
def __init__(self, isize):
|
||||
# number of inputs to this neuron
|
||||
self.isize = isize
|
||||
# importance to each input
|
||||
self.weight = [random.uniform(-1, 1) for _ in range(self.isize)]
|
||||
# importance of the neuron
|
||||
self.bias = random.uniform(-1, 1)
|
||||
|
||||
def forward(self, x, activate=True):
|
||||
def __init__(self, input_size: int) -> None:
|
||||
"""
|
||||
x : list of input values to the neuron
|
||||
Set up the neuron's parameters (weights, and bias)
|
||||
|
||||
:param input_size: Number of incomming inputs (must be > 0)
|
||||
"""
|
||||
# computes the weighted sum of inputs and add the bias
|
||||
self.z = sum(w * xi for w, xi in zip(self.weight, x)) + self.bias
|
||||
# normalize the output between 0 and 1 if activate
|
||||
last_output = sigmoid(self.z) if activate else self.z
|
||||
# Store input dimensions for structural consistency
|
||||
self._input_size: int = input_size
|
||||
|
||||
return last_output
|
||||
|
||||
# adjust weight and bias of neuron
|
||||
def backward(self, x, dcost_dy, learning_rate):
|
||||
"""
|
||||
x : list of input values to the neuron
|
||||
dcost_dy : derivate of the cost function `(2 * (output - target))`
|
||||
learning_rate : learning factor (adjust the speed of weight/bias change during training)
|
||||
# Scale each input influence using random weights.
|
||||
self._weight: list[float] = [
|
||||
random.uniform(-1., 1.) for _ in range(input_size)
|
||||
]
|
||||
|
||||
weight -= learning_rate * dC/dy * dy/dz * dz/dw
|
||||
bias -= learning_rate * dC/dy * dy/dz * dz/db
|
||||
"""
|
||||
# dy/dz: derivate of the sigmoid activation
|
||||
dy_dz = sigmoid_deriv(self.z)
|
||||
# dz/dw = x
|
||||
dz_dw = x
|
||||
# Initialize a shift to the activation threshold with a random bias
|
||||
self._bias: float = random.uniform(-1., 1.)
|
||||
|
||||
assert len(dz_dw) >= self.isize, "too many value for input size"
|
||||
def __repr__(self) -> str:
|
||||
jmp: int = int(math.sqrt(self._input_size))
|
||||
text: list[str] = []
|
||||
|
||||
# dz/db = 1
|
||||
dz_db = 1
|
||||
for i in range(0, self._input_size, jmp):
|
||||
line: str = str.join("", str(self._weight[i: (i + jmp)]))
|
||||
text.append(line)
|
||||
|
||||
for i in range(self.isize):
|
||||
# update each weight `weight -= learning_rate * dC/dy * dy/dz * x_i`
|
||||
self.weight[i] -= learning_rate * dcost_dy * dy_dz * dz_dw[i]
|
||||
return f"weight:\n{str.join("\n", text)}\nbias: {self._bias}"
|
||||
|
||||
# update bias: bias -= learning_rate * dC/dy * dy/dz * dz/db
|
||||
self.bias -= learning_rate * dcost_dy * dy_dz * dz_db
|
||||
# # neuron class
|
||||
# class Neuron:
|
||||
# """
|
||||
# z : linear combination of inputs and weights plus bias (pre-activation)
|
||||
# y : output of the activation function (sigmoid(z))
|
||||
# w : list of weights, one for each input
|
||||
# """
|
||||
# def __init__(self, input_size):
|
||||
# # number of inputs to this neuron
|
||||
# self._input_size = input_size
|
||||
# # importance to each input
|
||||
# self._weight = [random.uniform(-1, 1) for _ in range(self._input_size)]
|
||||
# # importance of the neuron
|
||||
# self._bias = random.uniform(-1, 1)
|
||||
|
||||
# return gradient vector len(input) dimension
|
||||
return [dcost_dy * dy_dz * w for w in self.weight]
|
||||
# def forward(self, x, activate=True):
|
||||
# """
|
||||
# x : list of input values to the neuron
|
||||
# """
|
||||
# # computes the weighted sum of inputs and add the bias
|
||||
# self._z = sum(w * xi for w, xi in zip(self.weight, x)) + self.bias
|
||||
# # normalize the output between 0 and 1 if activate
|
||||
# last_output = sigmoid(self.z) if activate else self.z
|
||||
|
||||
# return last_output
|
||||
|
||||
class Layer:
|
||||
def __init__(self, input_size, output_size):
|
||||
"""
|
||||
input_size : size of each neuron input
|
||||
output_size : size of neurons
|
||||
"""
|
||||
self.size = output_size
|
||||
# list of neurons
|
||||
self.neurons = [Neuron(input_size) for _ in range(output_size)]
|
||||
# # adjust weight and bias of neuron
|
||||
# def backward(self, x, dcost_dy, learning_rate):
|
||||
# """
|
||||
# x : list of input values to the neuron
|
||||
# dcost_dy : derivate of the cost function `(2 * (output - target))`
|
||||
# learning_rate : learning factor (adjust the speed of weight/bias change during training)
|
||||
|
||||
def forward(self, inputs, activate=True):
|
||||
self.inputs = inputs
|
||||
# give the same inputs to each neuron in the layer
|
||||
return [neuron.forward(inputs, activate) for neuron in self.neurons]
|
||||
# weight -= learning_rate * dC/dy * dy/dz * dz/dw
|
||||
# bias -= learning_rate * dC/dy * dy/dz * dz/db
|
||||
# """
|
||||
# # dy/dz: derivate of the sigmoid activation
|
||||
# dy_dz = sigmoid_deriv(self.z)
|
||||
# # dz/dw = x
|
||||
# dz_dw = x
|
||||
|
||||
# adjust weight and bias of the layer (all neurons)
|
||||
def backward(self, dcost_dy_list, learning_rate=0.1):
|
||||
# init layer gradient vector len(input) dimention
|
||||
input_gradients = [0.0] * len(self.inputs)
|
||||
# assert len(dz_dw) >= self.isize, "too many value for input size"
|
||||
|
||||
for i, neuron in enumerate(self.neurons):
|
||||
dcost_dy = dcost_dy_list[i]
|
||||
grad_to_input = neuron.backward(self.inputs, dcost_dy, learning_rate)
|
||||
# # dz/db = 1
|
||||
# dz_db = 1
|
||||
|
||||
# accumulate the input gradients from all neurons
|
||||
for j in range(len(grad_to_input)):
|
||||
input_gradients[j] += grad_to_input[j]
|
||||
# for i in range(self.isize):
|
||||
# # update each weight `weight -= learning_rate * dC/dy * dy/dz * x_i`
|
||||
# self.weight[i] -= learning_rate * dcost_dy * dy_dz * dz_dw[i]
|
||||
|
||||
# return layer gradient
|
||||
return input_gradients
|
||||
# # update bias: bias -= learning_rate * dC/dy * dy/dz * dz/db
|
||||
# self.bias -= learning_rate * dcost_dy * dy_dz * dz_db
|
||||
|
||||
class NeuralNetwork:
|
||||
def __init__(self, layer_size):
|
||||
self.layers = [Layer(layer_size[i], layer_size[i+1]) for i in range(len(layer_size) - 1)]
|
||||
# # return gradient vector len(weight) dimension
|
||||
# return [dcost_dy * dy_dz * w for w in self.weight]
|
||||
|
||||
def forward(self, inputs):
|
||||
output = inputs
|
||||
for i, layer in enumerate(self.layers):
|
||||
activate = (i != len(self.layers) - 1) # deactivate sigmoid latest neuron
|
||||
output = layer.forward(output, activate=activate)
|
||||
return output
|
||||
# def __repr__(self):
|
||||
# pass
|
||||
|
||||
def backward(self, inputs, targets, learning_rate=0.1):
|
||||
"""
|
||||
target must be a list with the same length that the final layer
|
||||
input
|
||||
"""
|
||||
output = self.forward(inputs)
|
||||
# class Layer:
|
||||
# def __init__(self, input_size, output_size):
|
||||
# """
|
||||
# input_size : size of each neuron input
|
||||
# output_size : size of neurons
|
||||
# """
|
||||
# self.size = output_size
|
||||
# # list of neurons
|
||||
# self.neurons = [Neuron(input_size) for _ in range(output_size)]
|
||||
|
||||
# computes the initial gradient of the cost function for each neuron
|
||||
# by using Mean Squared Error's derivate: dC/dy = 2 * (output - target)
|
||||
dcost_dy_list = [2 * (o - t) for o, t in zip(output, targets)]
|
||||
# def forward(self, inputs, activate=True):
|
||||
# self.inputs = inputs
|
||||
# # give the same inputs to each neuron in the layer
|
||||
# return [neuron.forward(inputs, activate) for neuron in self.neurons]
|
||||
|
||||
grad = dcost_dy_list
|
||||
for layer in reversed(self.layers):
|
||||
# backpropagate the gradient of the layer to update weights and biases
|
||||
grad = layer.backward(grad, learning_rate)
|
||||
|
||||
# return final gradient
|
||||
return grad
|
||||
# # adjust weight and bias of the layer (all neurons)
|
||||
# def backward(self, dcost_dy_list, learning_rate=0.1):
|
||||
# # init layer gradient vector len(input) dimention
|
||||
# input_gradients = [0.0] * len(self.inputs)
|
||||
|
||||
if __name__ == "__main__":
|
||||
print("you might want to run main.py instead of network.py")
|
||||
# for i, neuron in enumerate(self.neurons):
|
||||
# dcost_dy = dcost_dy_list[i]
|
||||
# grad_to_input = neuron.backward(self.inputs, dcost_dy, learning_rate)
|
||||
|
||||
# # accumulate the input gradients from all neurons
|
||||
# for j in range(len(grad_to_input)):
|
||||
# input_gradients[j] += grad_to_input[j]
|
||||
|
||||
# # return layer gradient
|
||||
# return input_gradients
|
||||
|
||||
# class NeuralNetwork:
|
||||
# def __init__(self, layer_size):
|
||||
# self.layers = [Layer(layer_size[i], layer_size[i+1]) for i in range(len(layer_size) - 1)]
|
||||
|
||||
# def forward(self, inputs):
|
||||
# output = inputs
|
||||
# for i, layer in enumerate(self.layers):
|
||||
# activate = (i != len(self.layers) - 1) # deactivate sigmoid latest neuron
|
||||
# output = layer.forward(output, activate=activate)
|
||||
# return output
|
||||
|
||||
# def backward(self, inputs, targets, learning_rate=0.1):
|
||||
# """
|
||||
# target must be a list with the same length that the final layer
|
||||
# input
|
||||
# """
|
||||
# output = self.forward(inputs)
|
||||
|
||||
# # computes the initial gradient of the cost function for each neuron
|
||||
# # by using Mean Squared Error's derivate: dC/dy = 2 * (output - target)
|
||||
# dcost_dy_list = [2 * (o - t) for o, t in zip(output, targets)]
|
||||
|
||||
# grad = dcost_dy_list
|
||||
# for layer in reversed(self.layers):
|
||||
# # backpropagate the gradient of the layer to update weights and biases
|
||||
# grad = layer.backward(grad, learning_rate)
|
||||
|
||||
# # return final gradient
|
||||
# return grad
|
||||
|
||||
# if __name__ == "__main__":
|
||||
# print("you might want to run main.py instead of network.py")
|
||||
|
||||
Reference in New Issue
Block a user