使用不完整的线性自动编码器执行PCA

使用不完整的线性自动编码器执行PCA

如果自动编码器仅使用线性激活，并且成本函数是均方误差MSE，则最后执行的是主成分分析（PCA）

以下代码构建了一个简单的线性自动编码器，来对3D数据集执行PCA，将其投影到2D：

from tensorflow import keras

encoder = keras.models.Sequential([
    keras.layers.Dense(2, input_shape=[3])
])
decoder = keras.models.Sequential([
    keras.layers.Dense(3, input_shape=[2])
])
autoencoder = keras.models.Sequential([encoder, decoder])
autoencoder.compile(loss='mse', optimizer=keras.optimizers.SGD(lr=.1))

• 将自动编码器分为两个子主件：编码器和解码器。两者都具有单一Dense层的常规Sequential模型，而自动编码器是包含编码器和解码器的Sequential模型（一个模型可以用作另一个模型中的层）
• 自动编码器的输出数量等于输入的数量
• 为了执行简单的PCA，不使用任何激活函数（即所有神经元都是线性的），成本函数为MSE

import tensorflow as tf

X_train = tf.random.normal([1000, 3], 0, 1)
print(X_train.shape)
history = autoencoder.fit(X_train, X_train, epochs=20)
# X_train=X_train[np.newaxis,...]
codings = encoder.predict(X_train)
print(codings.shape)

(1000, 3)
Epoch 1/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3010
Epoch 2/20
32/32 [==============================] - 0s 2ms/step - loss: 0.3008
Epoch 3/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3005
Epoch 4/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3007
Epoch 5/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3009
Epoch 6/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3004
Epoch 7/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3004
Epoch 8/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3011
Epoch 9/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3001
Epoch 10/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3007
Epoch 11/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3010
Epoch 12/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3004
Epoch 13/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3004
Epoch 14/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3003
Epoch 15/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3007
Epoch 16/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3004
Epoch 17/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3005
Epoch 18/20
32/32 [==============================] - 0s 1ms/step - loss: 0.2999
Epoch 19/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3003
Epoch 20/20
32/32 [==============================] - 0s 1ms/step - loss: 0.3005
(1000, 2)

import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D

fig = plt.figure()
ax = fig.add_subplot(121, projection='3d')
ax.plot(X_train[:, 0], X_train[:, 1], X_train[:, 2])
ax = fig.add_subplot(122)
ax.plot(codings[:, 0], codings[:, 1])

​
​