注：因?yàn)橛布?#xff0c;這次的實(shí)驗(yàn)并沒(méi)有生成圖片，但是代碼應(yīng)該是沒(méi)有問(wèn)題的，可以參考學(xué)習(xí)一下。
本次基于DCGAN實(shí)現(xiàn)動(dòng)漫人物的生成。最終的效果可以參考大神**K同學(xué)啊**的博客。與上篇文章基于DCGAN生成手寫數(shù)字的步驟基本一致。

1.導(dǎo)入庫(kù)

import tensorflow as tf import numpy as np import glob,imageio,os,PIL,pathlib import matplotlib.pyplot as plt# 支持中文 plt.rcParams['font.sans-serif'] = ['SimHei'] # 用來(lái)正常顯示中文標(biāo)簽 plt.rcParams['axes.unicode_minus'] = False # 用來(lái)正常顯示負(fù)號(hào)

2.數(shù)據(jù)準(zhǔn)備

data_dir = "E:/tmp/.keras/datasets/car_face_photos" data_dir = pathlib.Path(data_dir)pic_paths = list(data_dir.glob('*')) pic_paths = [str(path) for path in pic_paths] img_count = len(list(pic_paths))#共21551張圖片 plt.figure(figsize=(10, 5)) plt.suptitle("數(shù)據(jù)示例", fontsize=15)for i in range(40):plt.subplot(5, 8, i + 1)plt.xticks([])plt.yticks([])# 顯示圖片images = plt.imread(pic_paths[i])plt.imshow(images)plt.show()

查看圖片：

數(shù)據(jù)預(yù)處理：
1.歸一化到[-1,1]之間
2.調(diào)整圖片大小為[64,64]
3.將數(shù)據(jù)按照batch_size劃分開，并打亂

#數(shù)據(jù)處理 def preprocess_image(image):image = tf.image.decode_jpeg(image,channels=3)image = tf.image.resize(image,[64,64])return (image - 127.5)/127.5 def load_and_preprocess_image(path):image = tf.io.read_file(path)return preprocess_image(image)path_ds = tf.data.Dataset.from_tensor_slices(pic_paths) image_ds = path_ds.map(load_and_preprocess_image,num_parallel_calls=tf.data.experimental.AUTOTUNE)buffer_size = 60000 batch_size = 256 dataset = image_ds.shuffle(buffer_size).batch(batch_size)

3.生成器與判別器的構(gòu)建

生成器采用tf.keras.layers.Conv2DTranspose（上采樣層）從噪聲數(shù)據(jù)中產(chǎn)生圖片。以一個(gè)使用該種子作為輸入的 Dense 層開始，然后多次上采樣直到達(dá)到所期望的 64x64x3 的圖片尺寸。
除了最后一層使用tanh作為激活函數(shù)外，其余的都采用LeakyReLU作為激活函數(shù)。

def Geberator_model():model = tf.keras.Sequential([])model.add(tf.keras.layers.Dense(4*4*1024,use_bias=False,input_shape=(100,)))model.add(tf.keras.layers.BatchNormalization())model.add(tf.keras.layers.LeakyReLU())model.add(tf.keras.layers.Reshape((4,4,1024)))assert model.output_shape == (None,4,4,1024)#1model.add(tf.keras.layers.Conv2DTranspose(512,(5,5),strides=(2,2),padding="same",use_bias=False))assert model.output_shape == (None,8,8,512)model.add(tf.keras.layers.BatchNormalization())model.add(tf.keras.layers.LeakyReLU())#2model.add(tf.keras.layers.Conv2DTranspose(256, (5, 5), strides=(2, 2), padding="same", use_bias=False))assert model.output_shape == (None, 16, 16, 256)model.add(tf.keras.layers.BatchNormalization())model.add(tf.keras.layers.LeakyReLU())#3model.add(tf.keras.layers.Conv2DTranspose(128, (5, 5), strides=(2, 2), padding='same', use_bias=False))assert model.output_shape == (None, 32, 32, 128)model.add(tf.keras.layers.BatchNormalization())model.add(tf.keras.layers.LeakyReLU())#4model.add(tf.keras.layers.Conv2DTranspose(3, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh'))assert model.output_shape == (None, 64, 64, 3)return model generator = Geberator_model() generator.summary() Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= dense (Dense) (None, 16384) 1638400 _________________________________________________________________ batch_normalization (BatchNo (None, 16384) 65536 _________________________________________________________________ leaky_re_lu (LeakyReLU) (None, 16384) 0 _________________________________________________________________ reshape (Reshape) (None, 4, 4, 1024) 0 _________________________________________________________________ conv2d_transpose (Conv2DTran (None, 8, 8, 512) 13107200 _________________________________________________________________ batch_normalization_1 (Batch (None, 8, 8, 512) 2048 _________________________________________________________________ leaky_re_lu_1 (LeakyReLU) (None, 8, 8, 512) 0 _________________________________________________________________ conv2d_transpose_1 (Conv2DTr (None, 16, 16, 256) 3276800 _________________________________________________________________ batch_normalization_2 (Batch (None, 16, 16, 256) 1024 _________________________________________________________________ leaky_re_lu_2 (LeakyReLU) (None, 16, 16, 256) 0 _________________________________________________________________ conv2d_transpose_2 (Conv2DTr (None, 32, 32, 128) 819200 _________________________________________________________________ batch_normalization_3 (Batch (None, 32, 32, 128) 512 _________________________________________________________________ leaky_re_lu_3 (LeakyReLU) (None, 32, 32, 128) 0 _________________________________________________________________ conv2d_transpose_3 (Conv2DTr (None, 64, 64, 3) 9600 ================================================================= Total params: 18,920,320 Trainable params: 18,885,760 Non-trainable params: 34,560 _________________________________________________________________

判別器為基于CNN的圖片分類器

#判別器的構(gòu)建 def Discriminator_model():model = tf.keras.Sequential([tf.keras.layers.Conv2D(128,(5,5),strides=(2,2),padding="same",input_shape=[64,64,1]),tf.keras.layers.LeakyReLU(),tf.keras.layers.Dropout(0.3),tf.keras.layers.Conv2D(128,(5,5),strides=(2,2),padding="same"),tf.keras.layers.LeakyReLU(),tf.keras.layers.Dropout(0.3),tf.keras.layers.Conv2D(256, (5, 5), strides=(2, 2), padding="same"),tf.keras.layers.LeakyReLU(),tf.keras.layers.Dropout(0.3),tf.keras.layers.Conv2D(512, (5, 5), strides=(2, 2), padding="same"),tf.keras.layers.LeakyReLU(),tf.keras.layers.Dropout(0.3),tf.keras.layers.Flatten(),tf.keras.layers.Dense(1,activation='sigmoid')])return model discriminator = Discriminator_model() discriminator.summary() Model: "sequential_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d (Conv2D) (None, 32, 32, 128) 3328 _________________________________________________________________ leaky_re_lu_4 (LeakyReLU) (None, 32, 32, 128) 0 _________________________________________________________________ dropout (Dropout) (None, 32, 32, 128) 0 _________________________________________________________________ conv2d_1 (Conv2D) (None, 16, 16, 128) 409728 _________________________________________________________________ leaky_re_lu_5 (LeakyReLU) (None, 16, 16, 128) 0 _________________________________________________________________ dropout_1 (Dropout) (None, 16, 16, 128) 0 _________________________________________________________________ conv2d_2 (Conv2D) (None, 8, 8, 256) 819456 _________________________________________________________________ leaky_re_lu_6 (LeakyReLU) (None, 8, 8, 256) 0 _________________________________________________________________ dropout_2 (Dropout) (None, 8, 8, 256) 0 _________________________________________________________________ conv2d_3 (Conv2D) (None, 4, 4, 512) 3277312 _________________________________________________________________ leaky_re_lu_7 (LeakyReLU) (None, 4, 4, 512) 0 _________________________________________________________________ dropout_3 (Dropout) (None, 4, 4, 512) 0 _________________________________________________________________ flatten (Flatten) (None, 8192) 0 _________________________________________________________________ dense_1 (Dense) (None, 1) 8193 ================================================================= Total params: 4,518,017 Trainable params: 4,518,017 Non-trainable params: 0 _________________________________________________________________

4.loss值與優(yōu)化器

計(jì)算交叉熵

cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)

因?yàn)榉譃榕袆e器的生成器，因此loss值的計(jì)算方式也是不同的。
判別器的loss值：判斷真實(shí)圖片為1的loss與判斷生成圖片為0的loss之和。因?yàn)榕袆e器希望將真實(shí)圖片判別為1，將生成圖片判別為0.
生成器的loss值：判斷生成圖片為1的loss。因?yàn)樯善飨Ｍ傻膱D片是真實(shí)圖片，即判別為1.

def Discriminator_loss(real_out,fake_out):real_loss = cross_entropy(tf.ones_like(real_out),real_out)fake_loss = cross_entropy(tf.zeros_like(fake_out),fake_out)return real_loss+fake_loss def Generator_loss(fake_out):return cross_entropy(tf.ones_like(fake_out),fake_out)

優(yōu)化器也分為兩個(gè)：

generator_opt = tf.keras.optimizers.Adam(learning_rate=1e-4) discriminator_opt = tf.keras.optimizers.Adam(learning_rate=1e-4)

參數(shù)設(shè)置

epochs = 600 noise_dim = 100 num_exp_to_generate = 16seed = tf.random.normal([num_exp_to_generate,noise_dim])

5.批次訓(xùn)練

訓(xùn)練循環(huán)在生成器接收到一個(gè)隨機(jī)種子作為輸入時(shí)開始。該種子用于生產(chǎn)一張圖片。判別器隨后被用于區(qū)分真實(shí)圖片（選自訓(xùn)練集）和偽造圖片（由生成器生成）。針對(duì)這里的每一個(gè)模型都計(jì)算損失函數(shù)，并且計(jì)算梯度用于更新生成器與判別器。

def train_step(images):noise = tf.random.normal([batch_size,noise_dim])#生成一個(gè)batch_size*noise_dim的數(shù)據(jù)，相當(dāng)于生成了batch_size個(gè)長(zhǎng)度為100的隨機(jī)向量with tf.GradientTape() as gen_tape,tf.GradientTape() as dis_tape:#兩個(gè)Tape，一個(gè)代表生成器，一個(gè)代表判別器。real_out = discriminator(images,training = True)#利用判別器對(duì)真實(shí)的圖片進(jìn)行訓(xùn)練，得到一個(gè)modelgen_image = generator(noise,training = True)#利用生成器對(duì)噪聲數(shù)據(jù)生成圖片fake_out = discriminator(gen_image, training=True)#利用判別器對(duì)生成的圖片進(jìn)行訓(xùn)練gen_loss = Generator_loss(fake_out)#利用判別器對(duì)生成圖片的判斷計(jì)算生成器的loss值dis_loss = Discriminator_loss(real_out,fake_out)##利用判別器對(duì)生成圖片和真實(shí)圖片的判斷計(jì)算判別器的loss值gradient_gen = gen_tape.gradient(gen_loss,generator.trainable_variables)#根據(jù)生成器的loss值和網(wǎng)絡(luò)模型計(jì)算梯度gradient_dis = dis_tape.gradient(dis_loss, discriminator.trainable_variables)#根據(jù)判別器的loss值和網(wǎng)絡(luò)模型計(jì)算梯度Generator_opt.apply_gradients(zip(gradient_gen,generator.trainable_variables))#根據(jù)梯度對(duì)生成器進(jìn)行梯度更新Discriminator_opt.apply_gradients(zip(gradient_dis,discriminator.trainable_variables))#根據(jù)梯度對(duì)判別器進(jìn)行梯度更新

可視化圖片并保存到本地

def Generator_plot_image(gen_model,test_noise,epoch):pre_images = gen_model.predict(test_noise,training = False)fig = plt.figure(figsize=(4,4))for i in range(pre_images.shape[0]):plt.subplot(4,4,i+1)plt.imshow((pre_images[i,:,:,0]+1)/2)plt.axis('off')fig.savefig("E:/tmp/.keras/datasets/cartoon_photos_gen_DCGAN/%05d.png" % epoch)plt.close()

訓(xùn)練模型：

def train(dataset,epochs):for epoch in range(epochs):for image_batch in dataset:train_step(image_batch)print('.',end='')print()Generator_plot_image(generator,seed,epoch) train(dataset,epochs)

也可以用來(lái)生成其他的圖片，可以起到數(shù)據(jù)增強(qiáng)的效果。
努力加油a啊

總結(jié)

以上是生活随笔為你收集整理的深度学习之基于DCGAN实现动漫人物的生成的全部?jī)?nèi)容，希望文章能夠幫你解決所遇到的問(wèn)題。

如果覺(jué)得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。