我想为神经网络绘制一张动态图，以观察学习过程中权重的变化和神经元的激活。如何在 Python 中模拟该过程？

更准确地说，如果网络形状是:[1000, 300, 50]， 然后我想画一个三层神经网络，分别包含 1000、300 和 50 个神经元。 此外，我希望这张图片能反射(reflect)每个时期每一层神经元的饱和度。

我不知道该怎么做。有人能帮我解释一下吗？

最佳答案

我根据米洛的回答改编了一些部分

from matplotlib import pyplot
from math import cos, sin, atan


class Neuron():
    def __init__(self, x, y):
        self.x = x
        self.y = y

    def draw(self, neuron_radius):
        circle = pyplot.Circle((self.x, self.y), radius=neuron_radius, fill=False)
        pyplot.gca().add_patch(circle)


class Layer():
    def __init__(self, network, number_of_neurons, number_of_neurons_in_widest_layer):
        self.vertical_distance_between_layers = 6
        self.horizontal_distance_between_neurons = 2
        self.neuron_radius = 0.5
        self.number_of_neurons_in_widest_layer = number_of_neurons_in_widest_layer
        self.previous_layer = self.__get_previous_layer(network)
        self.y = self.__calculate_layer_y_position()
        self.neurons = self.__intialise_neurons(number_of_neurons)

    def __intialise_neurons(self, number_of_neurons):
        neurons = []
        x = self.__calculate_left_margin_so_layer_is_centered(number_of_neurons)
        for iteration in xrange(number_of_neurons):
            neuron = Neuron(x, self.y)
            neurons.append(neuron)
            x += self.horizontal_distance_between_neurons
        return neurons

    def __calculate_left_margin_so_layer_is_centered(self, number_of_neurons):
        return self.horizontal_distance_between_neurons * (self.number_of_neurons_in_widest_layer - number_of_neurons) / 2

    def __calculate_layer_y_position(self):
        if self.previous_layer:
            return self.previous_layer.y + self.vertical_distance_between_layers
        else:
            return 0

    def __get_previous_layer(self, network):
        if len(network.layers) > 0:
            return network.layers[-1]
        else:
            return None

    def __line_between_two_neurons(self, neuron1, neuron2):
        angle = atan((neuron2.x - neuron1.x) / float(neuron2.y - neuron1.y))
        x_adjustment = self.neuron_radius * sin(angle)
        y_adjustment = self.neuron_radius * cos(angle)
        line = pyplot.Line2D((neuron1.x - x_adjustment, neuron2.x + x_adjustment), (neuron1.y - y_adjustment, neuron2.y + y_adjustment))
        pyplot.gca().add_line(line)

    def draw(self, layerType=0):
        for neuron in self.neurons:
            neuron.draw( self.neuron_radius )
            if self.previous_layer:
                for previous_layer_neuron in self.previous_layer.neurons:
                    self.__line_between_two_neurons(neuron, previous_layer_neuron)
        # write Text
        x_text = self.number_of_neurons_in_widest_layer * self.horizontal_distance_between_neurons
        if layerType == 0:
            pyplot.text(x_text, self.y, 'Input Layer', fontsize = 12)
        elif layerType == -1:
            pyplot.text(x_text, self.y, 'Output Layer', fontsize = 12)
        else:
            pyplot.text(x_text, self.y, 'Hidden Layer '+str(layerType), fontsize = 12)

class NeuralNetwork():
    def __init__(self, number_of_neurons_in_widest_layer):
        self.number_of_neurons_in_widest_layer = number_of_neurons_in_widest_layer
        self.layers = []
        self.layertype = 0

    def add_layer(self, number_of_neurons ):
        layer = Layer(self, number_of_neurons, self.number_of_neurons_in_widest_layer)
        self.layers.append(layer)

    def draw(self):
        pyplot.figure()
        for i in range( len(self.layers) ):
            layer = self.layers[i]
            if i == len(self.layers)-1:
                i = -1
            layer.draw( i )
        pyplot.axis('scaled')
        pyplot.axis('off')
        pyplot.title( 'Neural Network architecture', fontsize=15 )
        pyplot.show()

class DrawNN():
    def __init__( self, neural_network ):
        self.neural_network = neural_network

    def draw( self ):
        widest_layer = max( self.neural_network )
        network = NeuralNetwork( widest_layer )
        for l in self.neural_network:
            network.add_layer(l)
        network.draw()

现在图层也被标记了，轴被删除并且构建图更容易。只需通过以下方式完成:

network = DrawNN( [2,8,8,1] )
network.draw()

这里构造了一个具有以下结构的网络:

• 输入层中有 2 个神经元
• 第一个隐藏层中有 8 个神经元
• 第二个隐藏层中有 8 个神经元
• 输出层中的 1 个神经元

关于python - 如何可视化神经网络，我们在Stack Overflow上找到一个类似的问题： https://stackoverflow.com/questions/29888233/