深度学习TF—12.Auto-Encoders（AE）与Varational Auto Encoder（VAE）实战

一、Auto-Encoders（AE）实战

导入模块

# 图片的重建
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# 使用GPU，设置内存自动增长
from tensorflow.compat.v1 import ConfigProto
from tensorflow.compat.v1 import InteractiveSession

config = ConfigProto()
config.gpu_options.allow_growth = True
session = InteractiveSession(config=config)

import numpy as np
import random
import tensorflow as tf
import matplotlib.pyplot as plt
from PIL import Image
from tensorflow import keras
from tensorflow.keras import Sequential,layers

# 设置随机数种子
def seed_everying(SEED):
    os.environ['TF_DETERMINISTIC_OPS'] = '1'
    os.environ['PYTHONHASHSEED'] = str(SEED)
    random.seed(SEED)
    np.random.seed(SEED)
    tf.random.set_seed(SEED)
    
seed_everying(42)

assert tf.__version__.startswith('2.')

保存图片

# 定义Image保存函数,将多张Image保存到一张Image里面去
def save_images(imgs,name):
    news_im = Image.new('L',(280,280))

    index = 0
    # 将10张Image保存到一张Image里面去
    for i in range(0,280,28):
        for j in range(0,280,28):
            im = imgs[index]
            im = Image.fromarray(im,mode='L')
            news_im.paste(im,(i,j))
            index += 1

        news_im.save(name)

加载数据集并进行预处理

h_dim = 20  # 将原来的784维的数据降维到20维上
batchsz = 512
lr = 1e-3

# 加载数据集
(x_train, y_train), (x_test, y_test) = keras.datasets.fashion_mnist.load_data()
x_train, x_test = x_train.astype(np.float32) / 255., x_test.astype(np.float32) / 255.
# 无监督学习，我们不需要使用label，即y_train、y_test不使用
train_db = tf.data.Dataset.from_tensor_slices((x_train))
train_db = train_db.shuffle(batchsz * 5).batch(batchsz)
test_db = tf.data.Dataset.from_tensor_slices((x_test))
test_db = test_db.batch(batchsz)

print(x_train.shape, y_train.shape)
print(x_test.shape, y_test.shape)

搭建网络

# 建立模型
class AE(keras.Model):

    def __init__(self):
        super(AE, self).__init__()

        # Encoders-编码
        self.encoder = Sequential([
            layers.Dense(256, activation=tf.nn.relu),
            layers.Dense(128, activation=tf.nn.relu),
            layers.Dense(h_dim)
        ])

        # Decoders-解码
        self.decoder = Sequential([
            layers.Dense(128, activation=tf.nn.relu),
            layers.Dense(256, activation=tf.nn.relu),
            layers.Dense(784)
        ])

    def call(self, inputs, training=None):
        # [b,784] -> [b,10]
        h = self.encoder(inputs)
        # [b,10] -> [b,784]
        x_hat = self.decoder(h)

        return x_hat

全部代码：

# 图片的重建
import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# 使用GPU，设置内存自动增长
from tensorflow.compat.v1 import ConfigProto
from tensorflow.compat.v1 import InteractiveSession

config = ConfigProto()
config.gpu_options.allow_growth = True
session = InteractiveSession(config=config)

import numpy as np
import random
import tensorflow as tf
import matplotlib.pyplot as plt
from PIL import Image
from tensorflow import keras
from tensorflow.keras import Sequential, layers


# 设置随机数种子
# 随机数种子
def seed_everying(SEED):
    os.environ['TF_DETERMINISTIC_OPS'] = '1'
    os.environ['PYTHONHASHSEED'] = str(SEED)
    random.seed(SEED)
    np.random.seed(SEED)
    tf.random.set_seed(SEED)


seed_everying(42)

assert tf.__version__.startswith('2.')


# 定义Image保存函数,将多张Image保存到一张Image里面去
def save_images(imgs, name):
    news_im = Image.new('L', (280, 280))

    index = 0
    # 将10张Image保存到一张Image里面去
    for i in range(0, 280, 28):
        for j in range(0, 280, 28):
            im = imgs[index]
            im = Image.fromarray(im, mode='L')
            news_im.paste(im, (i, j))
            index += 1

        news_im.save(name)


h_dim = 20  # 将原来的784维的数据降维到20维上
batchsz = 512
lr = 1e-3

# 加载数据集
(x_train, y_train), (x_test, y_test) = keras.datasets.fashion_mnist.load_data()
x_train, x_test = x_train.astype(np.float32) / 255., x_test.astype(np.float32) / 255.
# 无监督学习，我们不需要使用label，即y_train、y_test不使用
train_db = tf.data.Dataset.from_tensor_slices((x_train))
train_db = train_db.shuffle(batchsz * 5).batch(batchsz)
test_db = tf.data.Dataset.from_tensor_slices((x_test))
test_db = test_db.batch(batchsz)

print(x_train.shape, y_train.shape)
print(x_test.shape, y_test.shape)


# 建立模型
class AE(keras.Model):

    def __init__(self):
        super(AE, self).__init__()

        # Encoders-编码
        self.encoder = Sequential([
            layers.Dense(256, activation=tf.nn.relu),
            layers.Dense(128, activation=tf.nn.relu),
            layers.Dense(h_dim)
        ])

        # Decoders-解码
        self.decoder = Sequential([
            layers.Dense(128, activation=tf.nn.relu),
            layers.Dense(256, activation=tf.nn.relu),
            layers.Dense(784)
        ])

    def call(self, inputs, training=None):
        # [b,784] -> [b,10]
        h = self.encoder(inputs)
        # [b,10] -> [b,784]
        x_hat = self.decoder(h)

        return x_hat


model = AE()
model.build(input_shape=(None, 784))
# 查看网络结构
model.summary()

optimizer = tf.optimizers.Adam(lr=lr)
# training
for epoch in range(50):
    for step, x in enumerate(train_db):
        # [b,28,28] -> [b,784]
        x = tf.reshape(x, [-1, 784])

        with tf.GradientTape() as tape:
            x_rec_logits = model.call(x)

            rec_loss = tf.losses.binary_crossentropy(x, x_rec_logits, from_logits=True)
            rec_loss = tf.reduce_mean(rec_loss)

        grads = tape.gradient(rec_loss, model.trainable_variables)
        optimizer.apply_gradients(zip(grads, model.trainable_variables))

        # 打印training进度
        if step % 100 == 0:
            print(epoch, step, float(rec_loss))

        # evaluation
        x = next(iter(test_db))
        # training时，将x打平了
        x = tf.reshape(x, [-1, 784])        
        logits = model.predict(x)
        x_hat = tf.sigmoid(logits)
        # [b,784] => [b,28,28]
        x_hat = tf.reshape(x_hat, [-1, 28, 28])

        # [b,28,28] => [2b,28,28]
        # x_concat = tf.concat([x, x_hat], axis=0)
        x_concat = x_hat
        x_concat = x_concat.numpy() * 255.
        x_concat = x_concat.astype(np.uint8)
        save_images(x_concat, 'ae_images/rec_epoch_%d.png' % epoch)

(60000, 28, 28) (60000,)
(10000, 28, 28) (10000,)
Model: "ae"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
sequential (Sequential)      multiple                  235146    
_________________________________________________________________
sequential_1 (Sequential)    multiple                  235920    
=================================================================
Total params: 471,066
Trainable params: 471,066
Non-trainable params: 0
_________________________________________________________________
0 0 0.6929969787597656
0 100 0.3294386863708496
1 0 0.31267669796943665
1 100 0.30951932072639465
2 0 0.2991226315498352
2 100 0.30571430921554565
3 0 0.30124884843826294
3 100 0.29927968978881836
4 0 0.29633498191833496
4 100 0.29800713062286377
5 0 0.28601688146591187
...

每5epoch展示图片

---

二、Varational Auto Encoder（VAE）实战

导入模块

# 图片的重建
import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# 使用GPU，设置内存自动增长
from tensorflow.compat.v1 import ConfigProto
from tensorflow.compat.v1 import InteractiveSession

config = ConfigProto()
config.gpu_options.allow_growth = True
session = InteractiveSession(config=config)

import numpy as np
import random
import tensorflow as tf
import matplotlib.pyplot as plt
from PIL import Image
from tensorflow import keras
from tensorflow.keras import Sequential, layers


# 设置随机数种子
def seed_everying(SEED):
    os.environ['TF_DETERMINISTIC_OPS'] = '1'
    os.environ['PYTHONHASHSEED'] = str(SEED)
    random.seed(SEED)
    np.random.seed(SEED)
    tf.random.set_seed(SEED)


seed_everying(42)

assert tf.__version__.startswith('2.')

保存图片

# 定义Image保存函数,将多张Image保存到一张Image里面去
def save_images(imgs,name):
    news_im = Image.new('L',(280,280))

    index = 0
    # 将10张Image保存到一张Image里面去
    for i in range(0,280,28):
        for j in range(0,280,28):
            im = imgs[index]
            im = Image.fromarray(im,mode='L')
            news_im.paste(im,(i,j))
            index += 1

        news_im.save(name)

加载数据并进行预处理

z_dim = 10
batchsz = 256
lr = 1e-3

# 加载数据集
(x_train, y_train), (x_test, y_test) = keras.datasets.fashion_mnist.load_data()
x_train, x_test = x_train.astype(np.float32) / 255., x_test.astype(np.float32) / 255.
# 无监督学习，我们不需要使用label，即y_train、y_test不使用
train_db = tf.data.Dataset.from_tensor_slices((x_train))
train_db = train_db.shuffle(batchsz * 5).batch(batchsz)
test_db = tf.data.Dataset.from_tensor_slices((x_test))
test_db = test_db.batch(batchsz)

print(x_train.shape, y_train.shape)
print(x_test.shape, y_test.shape)

损失函数定义

rec_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels = x,logits = x_hat)
rec_loss = tf.reduce_sum(rec_loss) / x.shape[0]

# compute KL divergence (mu, var) ~N(0,1)
# p~N(mu,∨var) q~N(0,1) 故u2=0,σ2=1
kl_div = -0.5 * (2*log_var + 1 - mu ** 2 - tf.exp(log_var))
kl_div = tf.reduce_mean(kl_div) / x.shape[0]

loss = rec_loss + 1. * kl_div

构建网络

class VAE(keras.Model):

    def __init__(self):
        super(VAE, self).__init__()

        # Encoder - 编码
        self.fc1 = layers.Dense(128)
        self.fc2 = layers.Dense(z_dim)  # get mean prediction
        self.fc3 = layers.Dense(z_dim)  # get var  prediction

        # Decoder
        self.fc4 = layers.Dense(128)
        self.fc5 = layers.Dense(784)

    # 定义编码函数
    def encoder(self, x):
        h = tf.nn.relu(self.fc1(x))
        # get mean
        mu = self.fc2(h)
        # get variance
        log_var = self.fc3(h)

        return mu, log_var

    # 定义解码函数
    def decoder(self, z):
        out = tf.nn.relu(self.fc4(z))
        out = self.fc5(out)
        return out

    # 定义z函数
    def reparameterize(self, mu, log_var):

        eps = tf.random.normal(log_var.shape)
        # 标准差
        std = tf.exp(log_var) ** 0.5
        # std = tf.exp(log_var * 0.5)
        z = mu + std * eps
        return z

    # 定义前向传播
    def call(self, inputs, training=None):
        # [b,784] => [b,z_dim],[b,z_dim]
        mu, log_var = self.encoder(inputs)
        # reparameterization trick
        z = self.reparameterize(mu, log_var)

        x_hat = self.decoder(z)
        # 除了前向传播不同以外，还有一个约束，使得mu，var趋向于正态分布
        return x_hat, mu, log_var

全部代码

# 图片的重建
import os

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

# 使用GPU，设置内存自动增长
from tensorflow.compat.v1 import ConfigProto
from tensorflow.compat.v1 import InteractiveSession

config = ConfigProto()
config.gpu_options.allow_growth = True
session = InteractiveSession(config=config)

import numpy as np
import random
import tensorflow as tf
import matplotlib.pyplot as plt
from PIL import Image
from tensorflow import keras
from tensorflow.keras import Sequential, layers


# 设置随机数种子
def seed_everying(SEED):
    os.environ['TF_DETERMINISTIC_OPS'] = '1'
    os.environ['PYTHONHASHSEED'] = str(SEED)
    random.seed(SEED)
    np.random.seed(SEED)
    tf.random.set_seed(SEED)


seed_everying(42)

assert tf.__version__.startswith('2.')

def save_images(imgs, name):
    new_im = Image.new('L', (280, 280))

    index = 0
    for i in range(0, 280, 28):
        for j in range(0, 280, 28):
            im = imgs[index]
            im = Image.fromarray(im, mode='L')
            new_im.paste(im, (i, j))
            index += 1

    new_im.save(name)



z_dim = 10
batchsz = 256
lr = 1e-3

# 加载数据集
(x_train, y_train), (x_test, y_test) = keras.datasets.fashion_mnist.load_data()
x_train, x_test = x_train.astype(np.float32) / 255., x_test.astype(np.float32) / 255.
# 无监督学习，我们不需要使用label，即y_train、y_test不使用
train_db = tf.data.Dataset.from_tensor_slices((x_train))
train_db = train_db.shuffle(batchsz * 5).batch(batchsz)
test_db = tf.data.Dataset.from_tensor_slices((x_test))
test_db = test_db.batch(batchsz)

print(x_train.shape, y_train.shape)
print(x_test.shape, y_test.shape)




class VAE(keras.Model):

    def __init__(self):
        super(VAE, self).__init__()

        # Encoder - 编码
        self.fc1 = layers.Dense(128)
        self.fc2 = layers.Dense(z_dim)  # get mean prediction
        self.fc3 = layers.Dense(z_dim)  # get var  prediction

        # Decoder
        self.fc4 = layers.Dense(128)
        self.fc5 = layers.Dense(784)

    # 定义编码函数
    def encoder(self, x):
        h = tf.nn.relu(self.fc1(x))
        # get mean
        mu = self.fc2(h)
        # get variance
        log_var = self.fc3(h)

        return mu, log_var

    # 定义解码函数
    def decoder(self, z):
        out = tf.nn.relu(self.fc4(z))
        out = self.fc5(out)
        return out

    # 定义z函数
    def reparameterize(self, mu, log_var):

        eps = tf.random.normal(log_var.shape)
        # 标准差
        std = tf.exp(log_var) ** 0.5
        # std = tf.exp(log_var * 0.5)
        z = mu + std * eps
        return z

    # 定义前向传播
    def call(self, inputs, training=None):
        # [b,784] => [b,z_dim],[b,z_dim]
        mu, log_var = self.encoder(inputs)
        # reparameterization trick
        z = self.reparameterize(mu, log_var)

        x_hat = self.decoder(z)
        # 除了前向传播不同以外，还有一个约束，使得mu，var趋向于正态分布
        return x_hat, mu, log_var


model = VAE()
model.build(input_shape=(4, 784))               # 这一个与往常构造的时候有区别
optimizer = tf.optimizers.Adam(lr=lr)

for epoch in range(50):
    for step, x in enumerate(train_db):
        # [b,28,28] => [b,784]
        x = tf.reshape(x, [-1, 784])

        with tf.GradientTape() as tape:
            x_hat, mu, log_var = model(x)

            rec_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels = x,logits = x_hat)
            rec_loss = tf.reduce_sum(rec_loss) / x.shape[0]

            # compute KL divergence (mu, var) ~N(0,1)
            # p~N(mu,∨var) q~N(0,1) 故u2=0,σ2=1
            kl_div = -0.5 * (2*log_var + 1 - mu ** 2 - tf.exp(log_var))
            kl_div = tf.reduce_mean(kl_div) / x.shape[0]

            loss = rec_loss + 1. * kl_div

        grads = tape.gradient(loss, model.trainable_variables)
        optimizer.apply_gradients(zip(grads, model.trainable_variables))

        if step % 100 == 0:
            print(epoch, step, 'kl div:', float(kl_div), 'rec loss:', float(loss))


        # evaluation-测试生成效果
        z = tf.random.normal((batchsz,z_dim))
        logits = model.decoder(z)
        x_hat = tf.sigmoid(logits)
        x_hat = tf.reshape(x_hat,[-1,28,28]).numpy()* 255.
        x_hat = x_hat.astype(np.uint8)

        save_images(x_hat,'vae_image/sampled_epoch%d.png'%epoch)

        # 重建图片
        x = next(iter(test_db))
        x = tf.reshape(x, [-1, 784])
        x_hat_logits, _, _ = model(x)
        x_hat = tf.sigmoid(x_hat_logits)
        x_hat = tf.reshape(x_hat, [-1, 28, 28]).numpy() * 255.
        x_hat = x_hat.astype(np.uint8)
        save_images(x_hat,'vae_image_1/sampled_epoch%d.png'%epoch)

(60000, 28, 28) (60000,)
(10000, 28, 28) (10000,)
0 0 kl div: 0.001522199483588338 rec loss: 545.1880493164062
0 100 kl div: 0.12963518500328064 rec loss: 276.12176513671875
0 200 kl div: 0.15164682269096375 rec loss: 247.87600708007812
1 0 kl div: 0.1401350051164627 rec loss: 252.12911987304688
1 100 kl div: 0.14920486509799957 rec loss: 241.8990478515625
1 200 kl div: 0.16304132342338562 rec loss: 233.77679443359375
2 0 kl div: 0.16015133261680603 rec loss: 238.06817626953125
2 100 kl div: 0.15632715821266174 rec loss: 240.16061401367188
2 200 kl div: 0.157673180103302 rec loss: 233.2672882080078
3 0 kl div: 0.1759733110666275 rec loss: 230.8567352294922
3 100 kl div: 0.16381677985191345 rec loss: 230.72335815429688
3 200 kl div: 0.15521030128002167 rec loss: 233.7858123779297
4 0 kl div: 0.16824732720851898 rec loss: 227.0357208251953
4 100 kl div: 0.16316646337509155 rec loss: 235.9604949951172
4 200 kl div: 0.1655232012271881 rec loss: 229.02232360839844
...

每5epoch打印一次：