Keras入门——（3）生成式对抗网络GAN

导入 matplotlib 模块：

import matplotlib

查看自己版本所支持的backends：

print(matplotlib.rcsetup.all_backends)

返回信息：

['GTK3Agg', 'GTK3Cairo', 'MacOSX', 'nbAgg', 'Qt4Agg', 'Qt4Cairo', 'Qt5Agg', 'Qt5Cairo', 'TkAgg', 'TkCairo', 'WebAgg', 'WX', 'WXAgg', 'WXCairo', 'agg', 'cairo', 'pdf', 'pgf', 'ps', 'svg', 'template']

查看当前工作的matplotlibrc文件是哪个：

print(matplotlib.matplotlib_fname())

返回信息：

D:\ProgramData\Anaconda2\lib\site-packages\matplotlib\mpl-data\matplotlibrc

打开 matplotlibrc 查看相应内容：

将 backend 修改为 TkAgg：

执行如下代码：

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

from keras.models import Model
from keras.layers import Dense, Activation, Input, Reshape
from keras.layers import Conv1D, Flatten, Dropout
from keras.optimizers import SGD, Adam

def sample_data(n_samples=10000, x_vals=np.arange(0, 5, .1), max_offset=100, mul_range=[1, 2]):
    vectors = []
    for i in range(n_samples):
        offset = np.random.random() * max_offset
        mul = mul_range[0] + np.random.random() * (mul_range[1] - mul_range[0])
        vectors.append(
            np.sin(offset + x_vals * mul) / 2 + .5
        )
    return np.array(vectors)

ax = pd.DataFrame(np.transpose(sample_data(5))).plot()
plt.show()

生成图像：

执行代码：

def get_generative(G_in, dense_dim=200, out_dim=50, lr=1e-3):
    x = Dense(dense_dim)(G_in)
    x = Activation('tanh')(x)
    G_out = Dense(out_dim, activation='tanh')(x)
    G = Model(G_in, G_out)
    opt = SGD(lr=lr)
    G.compile(loss='binary_crossentropy', optimizer=opt)
    return G, G_out

G_in = Input(shape=[10])
G, G_out = get_generative(G_in)
G.summary()

生成图像：

执行代码：

def get_discriminative(D_in, lr=1e-3, drate=.25, n_channels=50, conv_sz=5, leak=.2):
    x = Reshape((-1, 1))(D_in)
    x = Conv1D(n_channels, conv_sz, activation='relu')(x)
    x = Dropout(drate)(x)
    x = Flatten()(x)
    x = Dense(n_channels)(x)
    D_out = Dense(2, activation='sigmoid')(x)
    D = Model(D_in, D_out)
    dopt = Adam(lr=lr)
    D.compile(loss='binary_crossentropy', optimizer=dopt)
    return D, D_out

D_in = Input(shape=[50])
D, D_out = get_discriminative(D_in)
D.summary()

生成图像：

执行代码：

def set_trainability(model, trainable=False):
    model.trainable = trainable
    for layer in model.layers:
        layer.trainable = trainable

def make_gan(GAN_in, G, D):
    set_trainability(D, False)
    x = G(GAN_in)
    GAN_out = D(x)
    GAN = Model(GAN_in, GAN_out)
    GAN.compile(loss='binary_crossentropy', optimizer=G.optimizer)
    return GAN, GAN_out

GAN_in = Input([10])
GAN, GAN_out = make_gan(GAN_in, G, D)
GAN.summary()

生成图像：

执行代码：

def sample_data_and_gen(G, noise_dim=10, n_samples=10000):
    XT = sample_data(n_samples=n_samples)
    XN_noise = np.random.uniform(0, 1, size=[n_samples, noise_dim])
    XN = G.predict(XN_noise)
    X = np.concatenate((XT, XN))
    y = np.zeros((2*n_samples, 2))
    y[:n_samples, 1] = 1
    y[n_samples:, 0] = 1
    return X, y

def pretrain(G, D, noise_dim=10, n_samples=10000, batch_size=32):
    X, y = sample_data_and_gen(G, n_samples=n_samples, noise_dim=noise_dim)
    set_trainability(D, True)
    D.fit(X, y, epochs=1, batch_size=batch_size)

pretrain(G, D)

返回信息：

Epoch 1/1

   32/20000 [..............................] - ETA: 6:42 - loss: 0.7347
  288/20000 [..............................] - ETA: 47s - loss: 0.4808
  544/20000 [..............................] - ETA: 26s - loss: 0.3318
  800/20000 [>.............................] - ETA: 19s - loss: 0.2359
 1056/20000 [>.............................] - ETA: 15s - loss: 0.1805
 1312/20000 [>.............................] - ETA: 12s - loss: 0.1459
 1568/20000 [=>............................] - ETA: 11s - loss: 0.1223
 1824/20000 [=>............................] - ETA: 10s - loss: 0.1053
 2048/20000 [==>...........................] - ETA: 9s - loss: 0.0938
 2272/20000 [==>...........................] - ETA: 8s - loss: 0.0847
 2528/20000 [==>...........................] - ETA: 8s - loss: 0.0761
 2784/20000 [===>..........................] - ETA: 7s - loss: 0.0692
 3040/20000 [===>..........................] - ETA: 7s - loss: 0.0634
 3296/20000 [===>..........................] - ETA: 6s - loss: 0.0585
 3552/20000 [====>.........................] - ETA: 6s - loss: 0.0543
 3808/20000 [====>.........................] - ETA: 6s - loss: 0.0507
 4064/20000 [=====>........................] - ETA: 5s - loss: 0.0475
 4352/20000 [=====>........................] - ETA: 5s - loss: 0.0444
 4608/20000 [=====>........................] - ETA: 5s - loss: 0.0420
 4864/20000 [======>.......................] - ETA: 5s - loss: 0.0398
 5120/20000 [======>.......................] - ETA: 4s - loss: 0.0378
 5376/20000 [=======>......................] - ETA: 4s - loss: 0.0360
 5632/20000 [=======>......................] - ETA: 4s - loss: 0.0344
 5888/20000 [=======>......................] - ETA: 4s - loss: 0.0329
 6144/20000 [========>.....................] - ETA: 4s - loss: 0.0315
 6400/20000 [========>.....................] - ETA: 4s - loss: 0.0303
 6656/20000 [========>.....................] - ETA: 4s - loss: 0.0291
 6880/20000 [=========>....................] - ETA: 3s - loss: 0.0282
 7136/20000 [=========>....................] - ETA: 3s - loss: 0.0272
 7392/20000 [==========>...................] - ETA: 3s - loss: 0.0262
 7648/20000 [==========>...................] - ETA: 3s - loss: 0.0254
 7904/20000 [==========>...................] - ETA: 3s - loss: 0.0246
 8160/20000 [===========>..................] - ETA: 3s - loss: 0.0238
 8416/20000 [===========>..................] - ETA: 3s - loss: 0.0231
 8672/20000 [============>.................] - ETA: 3s - loss: 0.0224
 8928/20000 [============>.................] - ETA: 3s - loss: 0.0218
 9184/20000 [============>.................] - ETA: 2s - loss: 0.0212
 9440/20000 [=============>................] - ETA: 2s - loss: 0.0206
 9696/20000 [=============>................] - ETA: 2s - loss: 0.0200
 9952/20000 [=============>................] - ETA: 2s - loss: 0.0195
10208/20000 [==============>...............] - ETA: 2s - loss: 0.0190
10464/20000 [==============>...............] - ETA: 2s - loss: 0.0186
10720/20000 [===============>..............] - ETA: 2s - loss: 0.0181
10976/20000 [===============>..............] - ETA: 2s - loss: 0.0177
11232/20000 [===============>..............] - ETA: 2s - loss: 0.0173
11488/20000 [================>.............] - ETA: 2s - loss: 0.0169
11712/20000 [================>.............] - ETA: 2s - loss: 0.0166
11968/20000 [================>.............] - ETA: 2s - loss: 0.0163
12224/20000 [=================>............] - ETA: 2s - loss: 0.0159
12480/20000 [=================>............] - ETA: 1s - loss: 0.0156
12736/20000 [==================>...........] - ETA: 1s - loss: 0.0153
12992/20000 [==================>...........] - ETA: 1s - loss: 0.0150
13248/20000 [==================>...........] - ETA: 1s - loss: 0.0147
13504/20000 [===================>..........] - ETA: 1s - loss: 0.0144
13760/20000 [===================>..........] - ETA: 1s - loss: 0.0141
14016/20000 [====================>.........] - ETA: 1s - loss: 0.0139
14272/20000 [====================>.........] - ETA: 1s - loss: 0.0136
14528/20000 [====================>.........] - ETA: 1s - loss: 0.0134
14784/20000 [=====================>........] - ETA: 1s - loss: 0.0132
15040/20000 [=====================>........] - ETA: 1s - loss: 0.0129
15296/20000 [=====================>........] - ETA: 1s - loss: 0.0127
15552/20000 [======================>.......] - ETA: 1s - loss: 0.0125
15808/20000 [======================>.......] - ETA: 1s - loss: 0.0123
16064/20000 [=======================>......] - ETA: 0s - loss: 0.0121
16320/20000 [=======================>......] - ETA: 0s - loss: 0.0119
16576/20000 [=======================>......] - ETA: 0s - loss: 0.0118
16832/20000 [========================>.....] - ETA: 0s - loss: 0.0116
17088/20000 [========================>.....] - ETA: 0s - loss: 0.0114
17344/20000 [=========================>....] - ETA: 0s - loss: 0.0112
17600/20000 [=========================>....] - ETA: 0s - loss: 0.0111
17856/20000 [=========================>....] - ETA: 0s - loss: 0.0109
18144/20000 [==========================>...] - ETA: 0s - loss: 0.0107
18400/20000 [==========================>...] - ETA: 0s - loss: 0.0106
18656/20000 [==========================>...] - ETA: 0s - loss: 0.0104
18912/20000 [===========================>..] - ETA: 0s - loss: 0.0103
19168/20000 [===========================>..] - ETA: 0s - loss: 0.0102
19456/20000 [============================>.] - ETA: 0s - loss: 0.0100
19712/20000 [============================>.] - ETA: 0s - loss: 0.0099
19968/20000 [============================>.] - ETA: 0s - loss: 0.0098
20000/20000 [==============================] - 5s 236us/step - loss: 0.0097

引入模块：

from tqdm import tqdm_notebook as tqdm

执行代码：

def sample_noise(G, noise_dim=10, n_samples=10000):
    X = np.random.uniform(0, 1, size=[n_samples, noise_dim])
    y = np.zeros((n_samples, 2))
    y[:, 1] = 1
    return X, y

def train(GAN, G, D, epochs=200, n_samples=10000, noise_dim=10, batch_size=32, verbose=False, v_freq=50):
    d_loss = []
    g_loss = []
    e_range = range(epochs)
    if verbose:
        e_range = tqdm(e_range)
    for epoch in e_range:
        X, y = sample_data_and_gen(G, n_samples=n_samples, noise_dim=noise_dim)
        set_trainability(D, True)
        d_loss.append(D.train_on_batch(X, y))

        X, y = sample_noise(G, n_samples=n_samples, noise_dim=noise_dim)
        set_trainability(D, False)
        g_loss.append(GAN.train_on_batch(X, y))
        if verbose and (epoch + 1) % v_freq == 0:
            print("Epoch #{}: Generative Loss: {}, Discriminative Loss: {}".format(epoch + 1, g_loss[-1], d_loss[-1]))
    return d_loss, g_loss

d_loss, g_loss = train(GAN, G, D, verbose=True)

返回信息：

HBox(children=(IntProgress(value=0, max=200), HTML(value='')))
Epoch #50: Generative Loss: 5.842154026031494, Discriminative Loss: 0.4683375060558319
Epoch #100: Generative Loss: 3.4111320972442627, Discriminative Loss: 0.13123030960559845
Epoch #150: Generative Loss: 5.5205817222595215, Discriminative Loss: 0.03762095794081688
Epoch #200: Generative Loss: 4.994686603546143, Discriminative Loss: 0.045186348259449005

执行代码：

ax = pd.DataFrame(
    {
        'Generative Loss': g_loss,
        'Discriminative Loss': d_loss,
    }
).plot(title='Training loss', logy=True)
ax.set_xlabel("Epochs")
ax.set_ylabel("Loss")
plt.show()

生成图像：

执行代码：

N_VIEWED_SAMPLES = 2
data_and_gen, _ = sample_data_and_gen(G, n_samples=N_VIEWED_SAMPLES)
pd.DataFrame(np.transpose(data_and_gen[N_VIEWED_SAMPLES:])).plot()
plt.show()

生成图像：

执行代码：

N_VIEWED_SAMPLES = 2
data_and_gen, _ = sample_data_and_gen(G, n_samples=N_VIEWED_SAMPLES)
pd.DataFrame(np.transpose(data_and_gen[N_VIEWED_SAMPLES:])).rolling(5).mean()[5:].plot()
plt.show()

生成图像：

完整代码如下：

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

from keras.models import Model
from keras.layers import Dense, Activation, Input, Reshape
from keras.layers import Conv1D, Flatten, Dropout
from keras.optimizers import SGD, Adam

from tqdm import tqdm_notebook as tqdm

#sec
def sample_data(n_samples=10000, x_vals=np.arange(0, 5, .1), max_offset=100, mul_range=[1, 2]):
    vectors = []
    for i in range(n_samples):
        offset = np.random.random() * max_offset
        mul = mul_range[0] + np.random.random() * (mul_range[1] - mul_range[0])
        vectors.append(
            np.sin(offset + x_vals * mul) / 2 + .5
        )
    return np.array(vectors)

ax = pd.DataFrame(np.transpose(sample_data(5))).plot()
plt.show()

#sec
def get_generative(G_in, dense_dim=200, out_dim=50, lr=1e-3):
    x = Dense(dense_dim)(G_in)
    x = Activation('tanh')(x)
    G_out = Dense(out_dim, activation='tanh')(x)
    G = Model(G_in, G_out)
    opt = SGD(lr=lr)
    G.compile(loss='binary_crossentropy', optimizer=opt)
    return G, G_out

G_in = Input(shape=[10])
G, G_out = get_generative(G_in)
G.summary()

#sec
def get_discriminative(D_in, lr=1e-3, drate=.25, n_channels=50, conv_sz=5, leak=.2):
    x = Reshape((-1, 1))(D_in)
    x = Conv1D(n_channels, conv_sz, activation='relu')(x)
    x = Dropout(drate)(x)
    x = Flatten()(x)
    x = Dense(n_channels)(x)
    D_out = Dense(2, activation='sigmoid')(x)
    D = Model(D_in, D_out)
    dopt = Adam(lr=lr)
    D.compile(loss='binary_crossentropy', optimizer=dopt)
    return D, D_out

D_in = Input(shape=[50])
D, D_out = get_discriminative(D_in)
D.summary()

#sec
def set_trainability(model, trainable=False):
    model.trainable = trainable
    for layer in model.layers:
        layer.trainable = trainable

def make_gan(GAN_in, G, D):
    set_trainability(D, False)
    x = G(GAN_in)
    GAN_out = D(x)
    GAN = Model(GAN_in, GAN_out)
    GAN.compile(loss='binary_crossentropy', optimizer=G.optimizer)
    return GAN, GAN_out

GAN_in = Input([10])
GAN, GAN_out = make_gan(GAN_in, G, D)
GAN.summary()

#sec
def sample_data_and_gen(G, noise_dim=10, n_samples=10000):
    XT = sample_data(n_samples=n_samples)
    XN_noise = np.random.uniform(0, 1, size=[n_samples, noise_dim])
    XN = G.predict(XN_noise)
    X = np.concatenate((XT, XN))
    y = np.zeros((2*n_samples, 2))
    y[:n_samples, 1] = 1
    y[n_samples:, 0] = 1
    return X, y

def pretrain(G, D, noise_dim=10, n_samples=10000, batch_size=32):
    X, y = sample_data_and_gen(G, n_samples=n_samples, noise_dim=noise_dim)
    set_trainability(D, True)
    D.fit(X, y, epochs=1, batch_size=batch_size)

pretrain(G, D)

#sec
def sample_noise(G, noise_dim=10, n_samples=10000):
    X = np.random.uniform(0, 1, size=[n_samples, noise_dim])
    y = np.zeros((n_samples, 2))
    y[:, 1] = 1
    return X, y

def train(GAN, G, D, epochs=200, n_samples=10000, noise_dim=10, batch_size=32, verbose=False, v_freq=50):
    d_loss = []
    g_loss = []
    e_range = range(epochs)
    if verbose:
        e_range = tqdm(e_range)
    for epoch in e_range:
        X, y = sample_data_and_gen(G, n_samples=n_samples, noise_dim=noise_dim)
        set_trainability(D, True)
        d_loss.append(D.train_on_batch(X, y))

        X, y = sample_noise(G, n_samples=n_samples, noise_dim=noise_dim)
        set_trainability(D, False)
        g_loss.append(GAN.train_on_batch(X, y))
        if verbose and (epoch + 1) % v_freq == 0:
            print("Epoch #{}: Generative Loss: {}, Discriminative Loss: {}".format(epoch + 1, g_loss[-1], d_loss[-1]))
    return d_loss, g_loss

d_loss, g_loss = train(GAN, G, D, verbose=True)

#sec
ax = pd.DataFrame(
    {
        'Generative Loss': g_loss,
        'Discriminative Loss': d_loss,
    }
).plot(title='Training loss', logy=True)
ax.set_xlabel("Epochs")
ax.set_ylabel("Loss")
plt.show()

#sec
N_VIEWED_SAMPLES = 2
data_and_gen, _ = sample_data_and_gen(G, n_samples=N_VIEWED_SAMPLES)
pd.DataFrame(np.transpose(data_and_gen[N_VIEWED_SAMPLES:])).plot()
plt.show()

#sec
N_VIEWED_SAMPLES = 2
data_and_gen, _ = sample_data_and_gen(G, n_samples=N_VIEWED_SAMPLES)
pd.DataFrame(np.transpose(data_and_gen[N_VIEWED_SAMPLES:])).rolling(5).mean()[5:].plot()
plt.show()

参考：

https://blog.csdn.net/tanmx219/article/details/88074600

https://blog.csdn.net/xqf1528399071/article/details/53385593

http://www.rricard.me/machine/learning/generative/adversarial/networks/keras/tensorflow/2017/04/05/gans-part2.html#Imports

Keras入门——（3）生成式对抗网络GAN的更多相关文章

生成式对抗网络GAN 的研究进展与展望
生成式对抗网络GAN的研究进展与展望.pdf 摘要: 生成式对抗网络GAN (Generative adversarial networks) 目前已经成为人工智能学界一个热门的研究方向. GAN的基 ...
【CV论文阅读】生成式对抗网络GAN
生成式对抗网络GAN 1. 基本GAN 在论文<Generative Adversarial Nets>提出的GAN是最原始的框架,可以看成极大极小博弈的过程,因此称为“对抗网络”.一般 ...
生成式对抗网络(GAN)实战——书法字体生成练习赛
https://www.tinymind.cn/competitions/ai 生成式对抗网络(GAN)是近年来大热的深度学习模型. 目前GAN最常使用的场景就是图像生成,作为一种优秀的生成式模型,G ...
【神经网络与深度学习】生成式对抗网络GAN研究进展（五）——Deep Convolutional Generative Adversarial Nerworks，DCGAN
[前言] 本文首先介绍生成式模型,然后着重梳理生成式模型(Generative Models)中生成对抗网络(Generative Adversarial Network)的研究与发展.作者 ...
不要怂，就是GAN (生成式对抗网络) （一）
前面我们用 TensorFlow 写了简单的 cifar10 分类的代码,得到还不错的结果,下面我们来研究一下生成式对抗网络 GAN,并且用 TensorFlow 代码实现. 自从 Ian Goodf ...
不要怂，就是GAN (生成式对抗网络) （一）： GAN 简介
前面我们用 TensorFlow 写了简单的 cifar10 分类的代码,得到还不错的结果,下面我们来研究一下生成式对抗网络 GAN,并且用 TensorFlow 代码实现. 自从 Ian Goodf ...
生成式对抗网络（GAN）学习笔记
图像识别和自然语言处理是目前应用极为广泛的AI技术,这些技术不管是速度还是准确度都已经达到了相当的高度,具体应用例如智能手机的人脸解锁.内置的语音助手.这些技术的实现和发展都离不开神经网络,可是传统的 ...
AI 生成式对抗网络（GAN）
生成式对抗网络(Generative Adversarial Network,简称GAN),主要由两部分构成:生成模型G和判别模型D.训练GAN就是两种模型的对抗过程. 生成模型:利用任意噪音(ran ...
GAN生成式对抗网络（四）——SRGAN超高分辨率图片重构
论文pdf 地址:https://arxiv.org/pdf/1609.04802v1.pdf 我的实际效果清晰度距离我的期待有距离. 颜色上面存在差距. 解决想法增加一个颜色判别器.将颜色值反馈 ...

随机推荐

Java日志介绍(1)-java.util.logging.Logger
java.util.logging.Logger是JDK自带的日志工具,其简单实现了日志的功能,不是很完善,所以在实际应用中使用的比较少.本文直接用代码演示其使用方法,文中所使用到的软件版本:Java ...
Python - int()
参考 https://docs.python.org/3/library/functions.html?highlight=int#int If x is not a number or if bas ...
ANSYS单元应用简介1
目录 1. LINK单元 2. BEAM单元 3. PLANE单元 4. SHELL单元 5. SOLIDE单元 6. COMBIN单元 ANSYS中的单元针对不同的应用对象,有着不同的选择方法,下面 ...
02-16Android学习进度报告十六
今天主要学习了GridView(网格视图)的基本使用和一些基本概念. 下面是GridView中的一些属性: android:columnWidth:设置列的宽度 android:gravity:组件对 ...
Python3中reduce和lambda的用法
reduce() 函数将一个数据集合(iterable[, initializer])可以看出是包含了初始化数据的,且初始化数据位列第1位,即集合中的第1个元素)中的所有数据进行下列操作:先对集合中的 ...
http协议学习 —— post请求方法提交application/x-www-form-urlencoded类型的数据格式
先推荐一篇很不错的文章:https://imququ.com/post/four-ways-to-post-data-in-http.html 说一下,如果是自己编写底层,那么要注意了,不能只有提交数 ...
「JSOI2011」柠檬
「JSOI2011」柠檬传送门斜率优化题. 在优化前,还有一个值得一提的优化: 对于最后的最优分割方案,每一段的两个端点一定是同颜色的,并且作为这一段的 \(s_0\) 证明:如果不作为这一段的 ...
Faster-RCNN Pytorch实现的minibatch包装
实际上faster-rcnn对于输入的图片是有resize操作的,在resize的图片基础上提取feature map,而后generate一定数量的RoI. 我想首先去掉这个resize的操作,对每 ...
一文解读IBMS (转)
IBMS系统是面向建筑设备生命周期的管理:面向楼宇的管理者和使用者:是传统BAS系统数据及功能的聚合. 更高效的分析和管理.更复杂的应用:跨平台.跨品牌的整合各子系统数据提供一套采集楼宇设备实时数据的 ...
CF 1073 E. Segment Sum
https://codeforces.com/problemset/problem/1073/E 题意:[l,r]中,出现0—9数字的种类数不超过k的数的和 dp[i][j][0/1] 表示 dfs到 ...

Keras入门——（3）生成式对抗网络GAN

Keras入门——（3）生成式对抗网络GAN的更多相关文章

随机推荐

热门专题