类神经网路 Neural_Networks - Ex 1: Visualization of MLP weights on MNIST

http://scikit-learn.org/stable/auto_examples/neural_networks/plot_mnist_filters.html

此范例将使用MNIST dataset的训练资料集去训练MLPClassifier，资料集中每张图片都是28*28，对于第一层的每个神经元都会有28*28个特征，输出结果是将训练资料的每个像素点对于神经元的权重画成28*28的图，用来表示图片上每个像素点对于神经元的权重多寡。

(一)引入函式库与资料

1.matplotlib.pyplot:用来绘制影像
2.sklearn.datasets:引入内建的手写数字资料库
3.sklearn.neural_network:引入类神经网路的套件

import matplotlib.pyplot as plt
from sklearn.datasets import fetch_mldata
from sklearn.neural_network import MLPClassifier
mnist = fetch_mldata("MNIST original")

(二)将资料切割成训练集与测试集

# 将灰阶影像降尺度降到[0,1]
X, y = mnist.data / 255., mnist.target
X_train, X_test = X[:60000], X[60000:]
y_train, y_test = y[:60000], y[60000:]

(三)设定分类器参数与训练网路并画出权重矩阵

#hidden_layer_sizes=(50)此处使用1层隐藏层，只有50个神经元，max_iter=10叠代训练10次
mlp = MLPClassifier(hidden_layer_sizes=(50), max_iter=10, alpha=1e-4,
                    solver='sgd', verbose=10, tol=1e-4, random_state=1,
                    learning_rate_init=.1)
mlp.fit(X_train, y_train)
#画出16个神经元的权重图，黑色表示负的权重，越深色表示数值越大，白色表示正的权重，越浅色表示数值越大
fig, axes = plt.subplots(4, 4)
# use global min / max to ensure all weights are shown on the same scale
vmin, vmax = mlp.coefs_[0].min(), mlp.coefs_[0].max()
for coef, ax in zip(mlp.coefs_[0].T, axes.ravel()):
    ax.matshow(coef.reshape(28, 28), cmap=plt.cm.gray, vmin=.5 * vmin,
               vmax=.5 * vmax)
    ax.set_xticks(())
    ax.set_yticks(())
plt.show()

(四)完整程式码

print(__doc__)
import matplotlib.pyplot as plt
from sklearn.datasets import fetch_mldata
from sklearn.neural_network import MLPClassifier
mnist = fetch_mldata("MNIST original")
X, y = mnist.data / 255., mnist.target
X_train, X_test = X[:60000], X[60000:]
y_train, y_test = y[:60000], y[60000:]
mlp = MLPClassifier(hidden_layer_sizes=(50), max_iter=10, alpha=1e-4,
                    solver='sgd', verbose=10, tol=1e-4, random_state=1,
                    learning_rate_init=.1)
mlp.fit(X_train, y_train)
print("Training set score: %f" % mlp.score(X_train, y_train))
print("Test set score: %f" % mlp.score(X_test, y_test))
fig, axes = plt.subplots(4, 4)
vmin, vmax = mlp.coefs_[0].min(), mlp.coefs_[0].max()
for coef, ax in zip(mlp.coefs_[0].T, axes.ravel()):
    ax.matshow(coef.reshape(28, 28), cmap=plt.cm.gray, vmin=.5 * vmin,
               vmax=.5 * vmax)
    ax.set_xticks(())
    ax.set_yticks(())
plt.show()