我们将实现一个 4
层的全连接网络,来完成二分类任务。网络输入节点数为 2
,隐藏 层的节点数设计为:25、50
和25
,输出层两个节点,分别表示属于类别 1
的概率和类别 2
的概率,如下图所示。这里并没有采用 Softmax
函数将网络输出概率值之和进行约束, 而是直接利用均方误差函数计算与 One-hot
编码的真实标签之间的误差,所有的网络激活 函数全部采用 Sigmoid
函数,这些设计都是为了能直接利用我们的梯度传播公式。
import numpy as np import matplotlib.pyplot as plt from sklearn import datasets from sklearn.model_selection import train_test_split
X, y = datasets.make_moons(n_samples=1000, noise=0.2, random_state=100) X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42) print(X.shape, y.shape) # (1000, 2) (1000,)
(1000, 2) (1000,)
def make_plot(X, y, plot_name): plt.figure(figsize=(12, 8)) plt.title(plot_name, fontsize=30) plt.scatter(X[y==0, 0], X[y==0, 1]) plt.scatter(X[y==1, 0], X[y==1, 1])
make_plot(X, y, "Classification Dataset Visualization ")
Layer
实现一个网络层,需要传入网络层的输入节点数、输出节点数、激 活函数类型等参数weights
和偏置张量 bias
在初始化时根据输入、输出节点数自动 生成并初始化class Layer: # 全链接网络层 def __init__(self, n_input, n_output, activation=None, weights=None, bias=None): """ :param int n_input: 输入节点数 :param int n_output: 输出节点数 :param str activation: 激活函数类型 :param weights: 权值张量,默认类内部生成 :param bias: 偏置,默认类内部生成 """ self.weights = weights if weights is not None else np.random.randn(n_input, n_output) * np.sqrt(1 / n_output) self.bias = bias if bias is not None else np.random.rand(n_output) * 0.1 self.activation = activation # 激活函数类型,如’sigmoid’ self.activation_output = None # 激活函数的输出值 o self.error = None # 用于计算当前层的 delta 变量的中间变量 self.delta = None # 记录当前层的 delta 变量,用于计算梯度 def activate(self, X): # 前向计算函数 r = np.dot(X, self.weights) + self.bias # X@W + b # 通过激活函数,得到全连接层的输出 o (activation_output) self.activation_output = self._apply_activation(r) return self.activation_output def _apply_activation(self, r): # 计算激活函数的输出 if self.activation is None: return r # 无激活函数,直接返回 elif self.activation == 'relu': return np.maximum(r, 0) elif self.activation == 'tanh': return np.tanh(r) elif self.activation == 'sigmoid': return 1 / (1 + np.exp(-r)) return r def apply_activation_derivative(self, r): # 计算激活函数的导数 # 无激活函数, 导数为 1 if self.activation is None: return np.ones_like(r) # ReLU 函数的导数 elif self.activation == 'relu': grad = np.array(r, copy=True) grad[r > 0] = 1. grad[r <= 0] = 0. return grad # tanh 函数的导数实现 elif self.activation == 'tanh': return 1 - r ** 2 # Sigmoid 函数的导数实现 elif self.activation == 'sigmoid': return r * (1 - r) return r
NeuralNetwork
类Layer
类对象,可以通过 add_layer
函数追加网络层,y_test.flatten().shape # (300,)
(300,)
class NeuralNetwork: def __init__(self): self._layers = [] # 网络层对象列表 def add_layer(self, layer): self._layers.append(layer) def feed_forward(self, X): # 前向传播(求导) for layer in self._layers: X = layer.activate(X) return X def backpropagation(self, X, y, learning_rate): # 反向传播算法实现 # 向前计算,得到最终输出值 output = self.feed_forward(X) for i in reversed(range(len(self._layers))): # 反向循环 layer = self._layers[i] if layer == self._layers[-1]: # 如果是输出层 layer.error = y - output # 计算最后一层的 delta,参考输出层的梯度公式 layer.delta = layer.error * layer.apply_activation_derivative(output) else: # 如果是隐藏层 next_layer = self._layers[i + 1] layer.error = np.dot(next_layer.weights, next_layer.delta) layer.delta = layer.error*layer.apply_activation_derivative(layer.activation_output) # 循环更新权值 for i in range(len(self._layers)): layer = self._layers[i] # o_i 为上一网络层的输出 o_i = np.atleast_2d(X if i == 0 else self._layers[i - 1].activation_output) # 梯度下降算法,delta 是公式中的负数,故这里用加号 layer.weights += layer.delta * o_i.T * learning_rate def train(self, X_train, X_test, y_train, y_test, learning_rate, max_epochs): # 网络训练函数 # one-hot 编码 y_onehot = np.zeros((y_train.shape[0], 2)) y_onehot[np.arange(y_train.shape[0]), y_train] = 1 mses = [] for i in range(max_epochs): # 训练 100 个 epoch for j in range(len(X_train)): # 一次训练一个样本 self.backpropagation(X_train[j], y_onehot[j], learning_rate) if i % 10 == 0: # 打印出 MSE Loss mse = np.mean(np.square(y_onehot - self.feed_forward(X_train))) mses.append(mse) print('Epoch: #%s, MSE: %f, Accuracy: %.2f%%' % (i, float(mse), self.accuracy(self.predict(X_test), y_test.flatten()) * 100)) return mses def accuracy(self, y_predict, y_test): # 计算准确度 return np.sum(y_predict == y_test) / len(y_test) def predict(self, X_predict): y_predict = self.feed_forward(X_predict) # 此时的 y_predict 形状是 [600 * 2],第二个维度表示两个输出的概率 y_predict = np.argmax(y_predict, axis=1) return y_predict
nn = NeuralNetwork() # 实例化网络类 nn.add_layer(Layer(2, 25, 'sigmoid')) # 隐藏层 1, 2=>25 nn.add_layer(Layer(25, 50, 'sigmoid')) # 隐藏层 2, 25=>50 nn.add_layer(Layer(50, 25, 'sigmoid')) # 隐藏层 3, 50=>25 nn.add_layer(Layer(25, 2, 'sigmoid')) # 输出层, 25=>2
# nn.train(X_train, X_test, y_train, y_test, learning_rate=0.01, max_epochs=50)
def plot_decision_boundary(model, axis): x0, x1 = np.meshgrid( np.linspace(axis[0], axis[1], int((axis[1] - axis[0])*100)).reshape(1, -1), np.linspace(axis[2], axis[3], int((axis[3] - axis[2])*100)).reshape(-1, 1) ) X_new = np.c_[x0.ravel(), x1.ravel()] y_predic = model.predict(X_new) zz = y_predic.reshape(x0.shape) from matplotlib.colors import ListedColormap custom_cmap = ListedColormap(['#EF9A9A', '#FFF590', '#90CAF9']) plt.contourf(x0, x1, zz, linewidth=5, cmap=custom_cmap)
plt.figure(figsize=(12, 8)) plot_decision_boundary(nn, [-2, 2.5, -1, 2]) plt.scatter(X[y==0, 0], X[y==0, 1]) plt.scatter(X[y==1, 0], X[y==1, 1])
<matplotlib.collections.PathCollection at 0x29018d6dfd0>
y_predict = nn.predict(X_test)
y_predict[:10] # array([1, 1, 0, 1, 0, 0, 0, 1, 1, 1], dtype=int64)
array([1, 1, 0, 1, 0, 0, 0, 1, 1, 1], dtype=int64)
y_test[:10] # array([1, 1, 0, 1, 0, 0, 0, 1, 1, 1], dtype=int64)
array([1, 1, 0, 1, 0, 0, 0, 1, 1, 1], dtype=int64)
nn.accuracy(y_predict, y_test.flatten()) # 0.86
0.86