vae_keras_cluster.py

#! -*- coding: utf-8 -*-

import numpy as np
from keras.layers import *
from keras.models import Model
from keras import backend as K
import imageio,os
from keras.datasets import mnist
# from keras.datasets import fashion_mnist as mnist


batch_size = 100
latent_dim = 20
epochs = 50
num_classes = 10
img_dim = 28
filters = 16
intermediate_dim = 256


# 加载MNIST数据集
(x_train, y_train_), (x_test, y_test_) = mnist.load_data()
x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
x_train = x_train.reshape((-1, img_dim, img_dim, 1))
x_test = x_test.reshape((-1, img_dim, img_dim, 1))



# 搭建模型
x = Input(shape=(img_dim, img_dim, 1))
h = x

for i in range(2):
    filters *= 2
    h = Conv2D(filters=filters,
               kernel_size=3,
               strides=2,
               padding='same')(h)
    h = LeakyReLU(0.2)(h)
    h = Conv2D(filters=filters,
               kernel_size=3,
               strides=1,
               padding='same')(h)
    h = LeakyReLU(0.2)(h)


h_shape = K.int_shape(h)[1:]
h = Flatten()(h)
z_mean = Dense(latent_dim)(h) # p(z|x)的均值
z_log_var = Dense(latent_dim)(h) # p(z|x)的方差

encoder = Model(x, z_mean) # 通常认为z_mean就是所需的隐变量编码


z = Input(shape=(latent_dim,))
h = z
h = Dense(np.prod(h_shape))(h)
h = Reshape(h_shape)(h)

for i in range(2):
    h = Conv2DTranspose(filters=filters,
                        kernel_size=3,
                        strides=1,
                        padding='same')(h)
    h = LeakyReLU(0.2)(h)
    h = Conv2DTranspose(filters=filters,
                        kernel_size=3,
                        strides=2,
                        padding='same')(h)
    h = LeakyReLU(0.2)(h)
    filters //= 2

x_recon = Conv2DTranspose(filters=1,
                          kernel_size=3,
                          activation='sigmoid',
                          padding='same')(h)


decoder = Model(z, x_recon) # 解码器
generator = decoder


z = Input(shape=(latent_dim,))
y = Dense(intermediate_dim, activation='relu')(z)
y = Dense(num_classes, activation='softmax')(y)

classfier = Model(z, y) # 隐变量分类器


# 重参数技巧
def sampling(args):
    z_mean, z_log_var = args
    epsilon = K.random_normal(shape=(K.shape(z_mean)[0], latent_dim))
    return z_mean + K.exp(z_log_var / 2) * epsilon

# 重参数层，相当于给输入加入噪声
z = Lambda(sampling, output_shape=(latent_dim,))([z_mean, z_log_var])
x_recon = decoder(z)
y = classfier(z)


class Gaussian(Layer):
    """这是个简单的层，定义q(z|y)中的均值参数，每个类别配一个均值。
    然后输出“z - 均值”，为后面计算loss准备。
    """
    def __init__(self, num_classes, **kwargs):
        self.num_classes = num_classes
        super(Gaussian, self).__init__(**kwargs)
    def build(self, input_shape):
        latent_dim = input_shape[-1]
        self.mean = self.add_weight(name='mean',
                                    shape=(self.num_classes, latent_dim),
                                    initializer='zeros')
    def call(self, inputs):
        z = inputs # z.shape=(batch_size, latent_dim)
        z = K.expand_dims(z, 1)
        return z - K.expand_dims(self.mean, 0)
    def compute_output_shape(self, input_shape):
        return (None, self.num_classes, input_shape[-1])

gaussian = Gaussian(num_classes)
z_prior_mean = gaussian(z)


# 建立模型
vae = Model(x, [x_recon, z_prior_mean, y])

# 下面一大通都是为了定义loss
z_mean = K.expand_dims(z_mean, 1)
z_log_var = K.expand_dims(z_log_var, 1)

lamb = 2.5 # 这是重构误差的权重，它的相反数就是重构方差，越大意味着方差越小。
xent_loss = 0.5 * K.mean((x - x_recon)**2, 0)
kl_loss = - 0.5 * (z_log_var - K.square(z_prior_mean))
kl_loss = K.mean(K.batch_dot(K.expand_dims(y, 1), kl_loss), 0)
cat_loss = K.mean(y * K.log(y + K.epsilon()), 0)
vae_loss = lamb * K.sum(xent_loss) + K.sum(kl_loss) + K.sum(cat_loss)


vae.add_loss(vae_loss)
vae.compile(optimizer='adam')
vae.summary()


vae.fit(x_train,
        shuffle=True,
        epochs=epochs,
        batch_size=batch_size,
        validation_data=(x_test, None))


means = K.eval(gaussian.mean)
x_train_encoded = encoder.predict(x_train)
y_train_pred = classfier.predict(x_train_encoded).argmax(axis=1)
x_test_encoded = encoder.predict(x_test)
y_test_pred = classfier.predict(x_test_encoded).argmax(axis=1)


def cluster_sample(path, category=0):
    """观察被模型聚为同一类的样本
    """
    n = 8
    figure = np.zeros((img_dim * n, img_dim * n))
    idxs = np.where(y_train_pred == category)[0]
    for i in range(n):
        for j in range(n):
            digit = x_train[np.random.choice(idxs)]
            digit = digit.reshape((img_dim, img_dim))
            figure[i * img_dim: (i + 1) * img_dim,
            j * img_dim: (j + 1) * img_dim] = digit
    imageio.imwrite(path, figure * 255)


def random_sample(path, category=0, std=1):
    """按照聚类结果进行条件随机生成
    """
    n = 8
    figure = np.zeros((img_dim * n, img_dim * n))
    for i in range(n):
        for j in range(n):
            noise_shape = (1, latent_dim)
            z_sample = np.array(np.random.randn(*noise_shape)) * std + means[category]
            x_recon = generator.predict(z_sample)
            digit = x_recon[0].reshape((img_dim, img_dim))
            figure[i * img_dim: (i + 1) * img_dim,
            j * img_dim: (j + 1) * img_dim] = digit
    imageio.imwrite(path, figure * 255)


if not os.path.exists('samples'):
    os.mkdir('samples')

for i in range(10):
    cluster_sample(u'samples/聚类类别_%s.png' % i, i)
    random_sample(u'samples/类别采样_%s.png' % i, i)


right = 0.
for i in range(10):
    _ = np.bincount(y_train_[y_train_pred == i])
    right += _.max()

print 'train acc: %s' % (right / len(y_train_))


right = 0.
for i in range(10):
    _ = np.bincount(y_test_[y_test_pred == i])
    right += _.max()

print 'test acc: %s' % (right / len(y_test_))