赞
踩
基于Python,MATLAB设计,OpenCV,,CNN,机器学习,R-CNN,GCN,LSTM,SVM,BP目标检测、语义分割、Re-ID、医学图像分割、目标跟踪、人脸识别、数据增广、
人脸检测、显著性目标检测、自动驾驶、人群密度估计、3D目标检测、CNN、AutoML、图像分割、SLAM、实例分割、人体姿态估计、视频目标分割,PyTorch、人脸检测、车道线检测、去雾 、全景分割、
行人检测、文本检测、OCR、姿态估计、边缘检测、场景文本检测、视频实例分割、3D点云、模型压缩、人脸对齐、超分辨、去噪、强化学习、行为识别、OpenCV、场景文本识别、去雨、机器学习、风格迁移、
视频目标检测、去模糊、活体检测、人脸关键点检测、3D目标跟踪、视频修复、人脸表情识别、时序动作检测、图像检索、异常检测等毕设指导,毕设选题,毕业设计开题报告,
生成对抗网络(GAN)是当今计算机科学领域最有趣的想法之一。两个模型通过对抗过程同时训练。一个生成器模型(“艺术家”)学习创造看起来真实的图像,而判别器模型(“艺术评论家”)学习区分真假图像。
GAN 的应用十分广泛,它的应用包括图像合成、风格迁移、照片修复以及照片编辑,数据增强等等,这次我将讲解如何用生成对抗网络生成小姐姐。
先看看我生成的小姐姐一睹为快
import tensorflow as tf
gpus = tf.config.list_physical_devices("GPU")
if gpus:
tf.config.experimental.set_memory_growth(gpus[0], True) #设置GPU显存用量按需使用
tf.config.set_visible_devices([gpus[0]],"GPU")
# 打印显卡信息,确认GPU可用
print(gpus)
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
from tensorflow.keras import layers
from IPython import display
import numpy as np
import glob,imageio,os,PIL,time,pathlib
import matplotlib.pyplot as plt
# 支持中文
plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号
您将使用 MNIST 数据集来训练生成器和判别器。生成器将生成类似于 MNIST 数据集的手写数字。
data_dir = "D:/jupyter notebook/DL-100-days/datasets/020_cartoon_face"
data_dir = pathlib.Path(data_dir)
pictures_paths = list(data_dir.glob('*'))
pictures_paths = [str(path) for path in pictures_paths]
pictures_paths[:3]
['D:\\jupyter notebook\\DL-100-days\\datasets\\020_cartoon_face\\1.png',
'D:\\jupyter notebook\\DL-100-days\\datasets\\020_cartoon_face\\10.png',
'D:\\jupyter notebook\\DL-100-days\\datasets\\020_cartoon_face\\100.png']
image_count = len(list(pictures_paths))
print("图片总数为:",image_count)
图片总数为: 21551
plt.figure(figsize=(10,5))
plt.suptitle("数据示例",fontsize=15)
for i in range(40):
plt.subplot(5,8,i+1)
plt.xticks([])
plt.yticks([])
# 显示图片
images = plt.imread(pictures_paths[i])
plt.imshow(images)
# plt.show()
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)
AUTOTUNE = tf.data.experimental.AUTOTUNE
path_ds = tf.data.Dataset.from_tensor_slices(pictures_paths)
image_ds = path_ds.map(load_and_preprocess_image, num_parallel_calls=AUTOTUNE)
BUFFER_SIZE = 60000
BATCH_SIZE = 256
# 批量化和打乱数据
train_dataset = image_ds.shuffle(BUFFER_SIZE).batch(BATCH_SIZE)
生成器使用 tf.keras.layers.Conv2DTranspose
(上采样)层来从种子(随机噪声)中产生图片。以一个使用该种子作为输入的 Dense
层开始,然后多次上采样直到达到所期望的 28x28x1 的图片尺寸。注意除了输出层使用 tanh 之外,其他每层均使用 tf.keras.layers.LeakyReLU
作为激活函数。
def make_generator_model(): model = tf.keras.Sequential() model.add(layers.Dense(4*4*1024, use_bias=False, input_shape=(100,))) model.add(layers.BatchNormalization()) model.add(layers.LeakyReLU()) model.add(layers.Reshape((4, 4, 1024))) assert model.output_shape == (None, 4, 4, 1024) # 第一层 model.add(layers.Conv2DTranspose(512, (5, 5), strides=(2, 2), padding='same', use_bias=False)) assert model.output_shape == (None, 8, 8, 512) model.add(layers.BatchNormalization()) model.add(layers.LeakyReLU()) # 第二层 model.add(layers.Conv2DTranspose(256, (5, 5), strides=(2, 2), padding='same', use_bias=False)) assert model.output_shape == (None, 16, 16, 256) model.add(layers.BatchNormalization()) model.add(layers.LeakyReLU()) # 第三层 model.add(layers.Conv2DTranspose(128, (5, 5), strides=(2, 2), padding='same', use_bias=False)) assert model.output_shape == (None, 32, 32, 128) model.add(layers.BatchNormalization()) model.add(layers.LeakyReLU()) # 第四层 model.add(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 = make_generator_model() generator.summary()
Model: "sequential_4" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= dense_4 (Dense) (None, 16384) 1638400 _________________________________________________________________ batch_normalization_8 (Batch (None, 16384) 65536 _________________________________________________________________ leaky_re_lu_16 (LeakyReLU) (None, 16384) 0 _________________________________________________________________ reshape_2 (Reshape) (None, 4, 4, 1024) 0 _________________________________________________________________ conv2d_transpose_8 (Conv2DTr (None, 8, 8, 512) 13107200 _________________________________________________________________ batch_normalization_9 (Batch (None, 8, 8, 512) 2048 _________________________________________________________________ leaky_re_lu_17 (LeakyReLU) (None, 8, 8, 512) 0 _________________________________________________________________ conv2d_transpose_9 (Conv2DTr (None, 16, 16, 256) 3276800 _________________________________________________________________ batch_normalization_10 (Batc (None, 16, 16, 256) 1024 _________________________________________________________________ leaky_re_lu_18 (LeakyReLU) (None, 16, 16, 256) 0 _________________________________________________________________ conv2d_transpose_10 (Conv2DT (None, 32, 32, 128) 819200 _________________________________________________________________ batch_normalization_11 (Batc (None, 32, 32, 128) 512 _________________________________________________________________ leaky_re_lu_19 (LeakyReLU) (None, 32, 32, 128) 0 _________________________________________________________________ conv2d_transpose_11 (Conv2DT (None, 64, 64, 3) 9600 ================================================================= Total params: 18,920,320 Trainable params: 18,885,760 Non-trainable params: 34,560 _________________________________________________________________
判别器是一个基于 CNN 的图片分类器。
def make_discriminator_model(): model = tf.keras.Sequential([ layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same', input_shape=[64, 64, 3]), layers.LeakyReLU(), layers.Dropout(0.3), layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same'), layers.LeakyReLU(), layers.Dropout(0.3), layers.Conv2D(256, (5, 5), strides=(2, 2), padding='same'), layers.LeakyReLU(), layers.Dropout(0.3), layers.Conv2D(512, (5, 5), strides=(2, 2), padding='same'), layers.LeakyReLU(), layers.Dropout(0.3), layers.Flatten(), layers.Dense(1) ]) return model discriminator = make_discriminator_model() discriminator.summary()
Model: "sequential_5" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= conv2d_8 (Conv2D) (None, 32, 32, 128) 9728 _________________________________________________________________ leaky_re_lu_20 (LeakyReLU) (None, 32, 32, 128) 0 _________________________________________________________________ dropout_8 (Dropout) (None, 32, 32, 128) 0 _________________________________________________________________ conv2d_9 (Conv2D) (None, 16, 16, 128) 409728 _________________________________________________________________ leaky_re_lu_21 (LeakyReLU) (None, 16, 16, 128) 0 _________________________________________________________________ dropout_9 (Dropout) (None, 16, 16, 128) 0 _________________________________________________________________ conv2d_10 (Conv2D) (None, 8, 8, 256) 819456 _________________________________________________________________ leaky_re_lu_22 (LeakyReLU) (None, 8, 8, 256) 0 _________________________________________________________________ dropout_10 (Dropout) (None, 8, 8, 256) 0 _________________________________________________________________ conv2d_11 (Conv2D) (None, 4, 4, 512) 3277312 _________________________________________________________________ leaky_re_lu_23 (LeakyReLU) (None, 4, 4, 512) 0 _________________________________________________________________ dropout_11 (Dropout) (None, 4, 4, 512) 0 _________________________________________________________________ flatten_2 (Flatten) (None, 8192) 0 _________________________________________________________________ dense_5 (Dense) (None, 1) 8193 ================================================================= Total params: 4,524,417 Trainable params: 4,524,417 Non-trainable params: 0 _________________________________________________________________
为两个模型定义损失函数和优化器。
# 该方法返回计算交叉熵损失的辅助函数
cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)
该方法量化判断真伪图片的能力。它将判别器对真实图片的预测值与值全为 1 的数组进行对比,将判别器对伪造(生成的)图片的预测值与值全为 0 的数组进行对比。
def discriminator_loss(real_output, fake_output):
real_loss = cross_entropy(tf.ones_like(real_output), real_output)
fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)
total_loss = real_loss + fake_loss
return total_loss
生成器损失量化其欺骗判别器的能力。直观来讲,如果生成器表现良好,判别器将会把伪造图片判断为真实图片(或 1)。这里我们将把判别器在生成图片上的判断结果与一个值全为 1 的数组进行对比。
def generator_loss(fake_output):
return cross_entropy(tf.ones_like(fake_output), fake_output)
由于我们需要分别训练两个网络,判别器和生成器的优化器是不同的。
generator_optimizer = tf.keras.optimizers.Adam(1e-4)
discriminator_optimizer = tf.keras.optimizers.Adam(1e-4)
tf.train.Checkpoint
只保存模型的参数,不保存模型的计算过程,因此一般用于在具有模型源代码的时候恢复之前训练好的模型参数。
# 定义模型保存路径
checkpoint_dir = './model/model_20/training_checkpoints'
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(generator_optimizer=generator_optimizer,
discriminator_optimizer=discriminator_optimizer,
generator=generator,
discriminator=discriminator)
EPOCHS = 600
noise_dim = 100
num_examples_to_generate = 16
# 我们将重复使用该种子(在 GIF 中更容易可视化进度)
seed = tf.random.normal([num_examples_to_generate, noise_dim])
训练循环在生成器接收到一个随机种子作为输入时开始。该种子用于生产一张图片。判别器随后被用于区分真实图片(选自训练集)和伪造图片(由生成器生成)。针对这里的每一个模型都计算损失函数,并且计算梯度用于更新生成器与判别器。
# 注意 `tf.function` 的使用 # 该注解使函数被“编译” @tf.function def train_step(images): # 生成噪音 noise = tf.random.normal([BATCH_SIZE, noise_dim]) with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape: generated_images = generator(noise, training=True) real_output = discriminator(images, training=True) fake_output = discriminator(generated_images, training=True) # 计算loss gen_loss = generator_loss(fake_output) disc_loss = discriminator_loss(real_output, fake_output) #计算梯度 gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables) gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables) #更新模型 generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables)) discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))
def train(dataset, epochs): for epoch in range(epochs): start = time.time() for image_batch in dataset: train_step(image_batch) # 实时更新生成的图片 display.clear_output(wait=True) generate_and_save_images(generator, epoch + 1, seed) # 每 15 个 epoch 保存一次模型 if (epoch + 1) % 100 == 0: checkpoint.save(file_prefix = checkpoint_prefix) print ('Time for epoch {} is {} sec'.format(epoch + 1, time.time()-start)) # 最后一个 epoch 结束后生成图片 display.clear_output(wait=True) generate_and_save_images(generator, epochs, seed)
生成与保存图片
def generate_and_save_images(model, epoch, test_input):
# 注意 training` 设定为 False
# 因此,所有层都在推理模式下运行(batchnorm)。
predictions = model(test_input, training=False)
fig = plt.figure(figsize=(5,5))
for i in range(predictions.shape[0]):
plt.subplot(4, 4, i+1)
plt.imshow(predictions[i] * 0.5 + 0.5) # 注意需要还原标准化的图片
plt.axis('off')
plt.savefig('./images/images_20/image_at_epoch_{:04d}.png'.format(epoch+600))
plt.show()
调用上面定义的 train()
方法来同时训练生成器和判别器。在训练之初,生成的图片看起来像是随机噪声。随着训练过程的进行,生成的数字将越来越真实。在大概 50 个 epoch 之后,这些图片看起来像是 MNIST 数字。
返回目录下最近一次checkpoint的文件名。例如如果save目录下有 model.ckpt-1.index
到 model.ckpt-10.index
的10个保存文件, tf.train.latest_checkpoint('./save')
即返回 ./save/model.ckpt-10
。
checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir))
<tensorflow.python.training.tracking.util.CheckpointLoadStatus at 0x20498267128>
%%time
:将会给出cell的代码运行一次所花费的时间。
%%time
train(train_dataset, EPOCHS)
Time for epoch 201 is 17.364601135253906 sec
import imageio,pathlib
def compose_gif():
# 图片地址
data_dir = "./images/images_20"
data_dir = pathlib.Path(data_dir)
paths = list(data_dir.glob('*'))
gif_images = []
for path in paths:
gif_images.append(imageio.imread(path))
imageio.mimsave("MINST_DCGAN_20.gif",gif_images,fps=8)
compose_gif()
print("GIF动图生成完成!")
GIF动图生成完成!
基于Python,MATLAB设计,OpenCV,,CNN,机器学习,R-CNN,GCN,LSTM,SVM,BP目标检测、语义分割、Re-ID、医学图像分割、目标跟踪、人脸识别、数据增广、
人脸检测、显著性目标检测、自动驾驶、人群密度估计、3D目标检测、CNN、AutoML、图像分割、SLAM、实例分割、人体姿态估计、视频目标分割,PyTorch、人脸检测、车道线检测、去雾 、全景分割、
行人检测、文本检测、OCR、姿态估计、边缘检测、场景文本检测、视频实例分割、3D点云、模型压缩、人脸对齐、超分辨、去噪、强化学习、行为识别、OpenCV、场景文本识别、去雨、机器学习、风格迁移、
视频目标检测、去模糊、活体检测、人脸关键点检测、3D目标跟踪、视频修复、人脸表情识别、时序动作检测、图像检索、异常检测等毕设指导,毕设选题,毕业设计开题报告,
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。