失眠网 > 机器学习笔记——生成式对抗网络GAN

机器学习笔记——生成式对抗网络GAN

时间：2022-04-14 09:03:30

相关推荐

机器学习笔记——生成式对抗网络GAN

前言

今天下午我们一起学习有趣的生成式对抗网络GAN。

GAN介绍

生成对抗网络（Generative Adversarial Networks，GAN）最早由 Ian Goodfellow 在年提出，是目前深度学习领域最具潜力的研究成果之一。它的核心思想是：同时训练两个相互协作、同时又相互竞争的深度神经网络（一个称为生成器 Generator，另一个称为判别器 Discriminator）来处理无监督学习的相关问题。

通常，我们会用下面这个例子来说明 GAN 的原理：将警察视为判别器，制造假币的犯罪分子视为生成器。一开始，犯罪分子会首先向警察展示一张假币。警察识别出该假币，并向犯罪分子反馈哪些地方是假的。接着，根据警察的反馈，犯罪分子改进工艺，制作一张更逼真的假币给警方检查。这时警方再反馈，犯罪分子再改进工艺。不断重复这一过程，直到警察识别不出真假，那么模型就训练成功了。

GAN的变体非常多，我们就以深度卷积生成对抗网络（Deep Convolutional GAN，DCGAN）为例，自动生成 MNIST 手写体数字。

判别器

判别器的作用是判断一个模型生成的图像和真实图像比，有多逼真。它的基本结构就是如下图所示的卷积神经网络（Convolutional Neural Network，CNN）。对于 MNIST 数据集来说，模型输入是一个 28x28 像素的单通道图像。Sigmoid 函数的输出值在 0-1 之间，表示图像真实度的概率，其中 0 表示肯定是假的，1 表示肯定是真的。与典型的 CNN 结构相比，这里去掉了层之间的 max-pooling。这里每个 CNN 层都以 LeakyReLU 为激活函数。而且为了防止过拟合，层之间的 dropout 值均被设置在 0.4-0.7 之间，模型结构如下：

ReLU激活函数极为f(x)=alpha * x for x < 0, f(x) = x for x>=0。alpha是一个小的非零数。在我以前的学习中，LeakyReLU一般和BN结合使用来防止过拟合这里把他俩分开了我不是特别理解。

生成器

生成器的作用是合成假的图像，其基本机构如下图所示。图中，我们使用了卷积的倒数，即转置卷积（transposed convolution），从 100 维的噪声（满足 -1 至 1 之间的均匀分布）中生成了假图像。这里我们采用了模型前三层之间的上采样来合成更逼真的手写图像。在层与层之间，我们采用了批量归一化的方法来平稳化训练过程。以 ReLU 函数为每一层结构之后的激活函数。最后一层 Sigmoid 函数输出最后的假图像。第一层设置了 0.3-0.5 之间的 dropout 值来防止过拟合。

批量正则化：（这就是BN）

GAN应用

1.图像生成：http://make.girls.moe/。

2.向量空间运算

3.文本转图像

4.超分辨率

主干代码

完整代码我会放在下载里，好像是没法设置0积分我只能设置1积分，这里说声抱歉，这里附上主干代码，基本就是完整的，只少了两个展示效果。

import numpy as npfrom keras.datasets import mnistfrom keras.models import Sequentialfrom keras.layers import Dense, Activation, Flatten, Reshapefrom keras.layers import Conv2D, Conv2DTranspose, UpSampling2Dfrom keras.layers import LeakyReLU, Dropoutfrom keras.layers import BatchNormalizationfrom keras.optimizers import RMSpropimport matplotlib.pyplot as pltclass DCGAN(object):def __init__(self, img_rows=28, img_cols=28, channel=1):# 初始化图片的行列通道数self.img_rows = img_rowsself.img_cols = img_colsself.channel = channelself.D = None # discriminator 判别器self.G = None # generator 生成器self.AM = None # adversarial model 对抗模型self.DM = None # discriminator model 判别模型# 判别模型def discriminator(self):if self.D:return self.Dself.D = Sequential()# 定义通道数64depth = 64# dropout系数dropout = 0.4# 输入28*28*1input_shape = (self.img_rows, self.img_cols, self.channel)# 输出14*14*64self.D.add(Conv2D(depth*1, 5, strides=2, input_shape=input_shape, padding='same'))self.D.add(LeakyReLU(alpha=0.2))self.D.add(Dropout(dropout))# 输出7*7*128self.D.add(Conv2D(depth*2, 5, strides=2, padding='same'))self.D.add(LeakyReLU(alpha=0.2))self.D.add(Dropout(dropout))# 输出4*4*256self.D.add(Conv2D(depth*4, 5, strides=2, padding='same'))self.D.add(LeakyReLU(alpha=0.2))self.D.add(Dropout(dropout))# 输出4*4*512self.D.add(Conv2D(depth*8, 5, strides=1, padding='same'))self.D.add(LeakyReLU(alpha=0.2))self.D.add(Dropout(dropout))# 全连接层self.D.add(Flatten())self.D.add(Dense(1))self.D.add(Activation('sigmoid'))self.D.summary()return self.D# 生成模型def generator(self):if self.G:return self.Gself.G = Sequential()# dropout系数dropout = 0.4# 通道数256depth = 64*4# 初始平面大小设置dim = 7# 全连接层，100个的随机噪声数据，7*7*256个神经网络self.G.add(Dense(dim*dim*depth, input_dim=100))self.G.add(BatchNormalization(momentum=0.9))self.G.add(Activation('relu'))# 把1维的向量变成3维数据(7,7,256)self.G.add(Reshape((dim, dim, depth)))self.G.add(Dropout(dropout))# 用法和 MaxPooling2D 基本相反，比如：UpSampling2D(size=(2, 2))# 就相当于将输入图片的长宽各拉伸一倍，整个图片被放大了# 上采样，采样后得到数据格式(14,14,256)self.G.add(UpSampling2D()) # 转置卷积，得到数据格式(14,14,128) self.G.add(Conv2DTranspose(int(depth/2), 5, padding='same')) self.G.add(BatchNormalization(momentum=0.9))self.G.add(Activation('relu'))# 上采样，采样后得到数据格式(28,28,128)self.G.add(UpSampling2D()) # 转置卷积，得到数据格式(28,28,64) self.G.add(Conv2DTranspose(int(depth/4), 5, padding='same'))self.G.add(BatchNormalization(momentum=0.9))self.G.add(Activation('relu'))# 转置卷积，得到数据格式(28,28,32) self.G.add(Conv2DTranspose(int(depth/8), 5, padding='same')) self.G.add(BatchNormalization(momentum=0.9))self.G.add(Activation('relu'))# 转置卷积，得到数据格式(28,28,1) self.G.add(Conv2DTranspose(1, 5, padding='same'))self.G.add(Activation('sigmoid'))self.G.summary()return self.G# 定义判别模型def discriminator_model(self):if self.DM:return self.DM# 定义优化器optimizer = RMSprop(lr=0.0002, decay=6e-8)# 构建模型self.DM = Sequential()self.DM.add(self.discriminator())pile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])return self.DM# 定义对抗模型def adversarial_model(self):if self.AM:return self.AM# 定义优化器optimizer = RMSprop(lr=0.0001, decay=3e-8)# 构建模型self.AM = Sequential()# 生成器self.AM.add(self.generator())# 判别器self.AM.add(self.discriminator())pile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])return self.AMclass MNIST_DCGAN(object):def __init__(self):# 图片的行数self.img_rows = 28# 图片的列数self.img_cols = 28# 图片的通道数self.channel = 1# 载入数据(x_train,y_train),(x_test,y_test) = mnist.load_data()# (60000,28,28)self.x_train = x_train/255.0# 改变数据格式(samples, rows, cols, channel)(60000,28,28,1)self.x_train = self.x_train.reshape(-1, self.img_rows, self.img_cols, 1).astype(np.float32)# 实例化DCGAN类self.DCGAN = DCGAN()# 定义判别器模型self.discriminator = self.DCGAN.discriminator_model()# 定义对抗模型self.adversarial = self.DCGAN.adversarial_model()# 定义生成器self.generator = self.DCGAN.generator()# 训练模型def train(self, train_steps=2000, batch_size=256, save_interval=0):noise_input = Noneif save_interval>0:# 生成16个100维的噪声数据noise_input = np.random.uniform(-1.0, 1.0, size=[16, 100])for i in range(train_steps):# 训练判别器，提升判别能力# 随机得到一个batch的图片数据images_train = self.x_train[np.random.randint(0, self.x_train.shape[0], size=batch_size), :, :, :]# 随机生成一个batch的噪声数据noise = np.random.uniform(-1.0, 1.0, size=[batch_size, 100])# 生成伪造的图片数据images_fake = self.generator.predict(noise)# 合并一个batch的真实图片和一个batch的伪造图片x = np.concatenate((images_train, images_fake))# 定义标签，真实数据的标签为1，伪造数据的标签为0y = np.ones([2*batch_size, 1])y[batch_size:, :] = 0# 把数据放到判别器中进行判断d_loss = self.discriminator.train_on_batch(x, y)# 训练对抗模型，提升生成器的造假能力# 标签都定义为1y = np.ones([batch_size, 1])# 生成一个batch的噪声数据noise = np.random.uniform(-1.0, 1.0, size=[batch_size, 100])# 训练对抗模型a_loss = self.adversarial.train_on_batch(noise, y)# 打印判别器的loss和准确率，以及对抗模型的loss和准确率log_mesg = "%d: [D loss: %f, acc: %f]" % (i, d_loss[0], d_loss[1])log_mesg = "%s [A loss: %f, acc: %f]" % (log_mesg, a_loss[0], a_loss[1])print(log_mesg)# 如果需要保存图片if save_interval>0:# 每save_interval次保存一次if (i+1)%save_interval==0:self.plot_images(save2file=True, samples=noise_input.shape[0], noise=noise_input, step=(i+1))# 保存图片def plot_images(self, save2file=False, fake=True, samples=16, noise=None, step=0):filename = 'mnist.png'if fake:if noise is None:noise = np.random.uniform(-1.0, 1.0, size=[samples, 100])else:filename = "mnist_%d.png" % step# 生成伪造的图片数据images = self.generator.predict(noise)else:# 获得真实图片数据i = np.random.randint(0, self.x_train.shape[0], samples)images = self.x_train[i, :, :, :]# 设置图片大小plt.figure(figsize=(10,10))# 生成16张图片for i in range(images.shape[0]):plt.subplot(4, 4, i+1)# 获取一个张图片数据image = images[i, :, :, :]# 变成2维的图片image = np.reshape(image, [self.img_rows, self.img_cols])# 显示灰度图片plt.imshow(image, cmap='gray')# 不显示坐标轴plt.axis('off')# 保存图片if save2file:plt.savefig(filename)plt.close('all')# 不保存的话就显示图片else:plt.show()# 实例化网络的类mnist_dcgan = MNIST_DCGAN()# 训练模型mnist_dcgan.train(train_steps=10000, batch_size=256, save_interval=500)

结果分析

其实到写这篇博客为止我还没有跑完，我的电脑太次了，我只能更根据他们跑过的结果给大家分析一下。首先，我们分析一下，训练过程，我们要进行两个不同的训练，分别是判别器的训练和对抗模型的训练。也就是先让我们的“警察”变得更强，然后再去训练“小偷（犯罪分子，因为更形象所以以后都用小偷）”。重复这个过程，最后我们的目的是让小偷变成“神偷”。

我们打印一部分结果：

Successfully downloaded train-images-idx3-ubyte.gz 9912422 bytes.Extracting mnist\train-images-idx3-ubyte.gzSuccessfully downloaded train-labels-idx1-ubyte.gz 28881 bytes.Extracting mnist\train-labels-idx1-ubyte.gzSuccessfully downloaded t10k-images-idx3-ubyte.gz 1648877 bytes.Extracting mnist\t10k-images-idx3-ubyte.gzSuccessfully downloaded t10k-labels-idx1-ubyte.gz 4542 bytes.Extracting mnist\t10k-labels-idx1-ubyte.gz_________________________________________________________________Layer (type) Output Shape Param # =================================================================conv2d_1 (Conv2D) (None, 14, 14, 64) 1664_________________________________________________________________leaky_re_lu_1 (LeakyReLU) (None, 14, 14, 64) 0 _________________________________________________________________dropout_1 (Dropout)(None, 14, 14, 64) 0 _________________________________________________________________conv2d_2 (Conv2D) (None, 7, 7, 128) 204928 _________________________________________________________________leaky_re_lu_2 (LeakyReLU) (None, 7, 7, 128) 0 _________________________________________________________________dropout_2 (Dropout)(None, 7, 7, 128) 0 _________________________________________________________________conv2d_3 (Conv2D) (None, 4, 4, 256) 819456 _________________________________________________________________leaky_re_lu_3 (LeakyReLU) (None, 4, 4, 256) 0 _________________________________________________________________dropout_3 (Dropout)(None, 4, 4, 256) 0 _________________________________________________________________conv2d_4 (Conv2D) (None, 4, 4, 512) 3277312 _________________________________________________________________leaky_re_lu_4 (LeakyReLU) (None, 4, 4, 512) 0 _________________________________________________________________dropout_4 (Dropout)(None, 4, 4, 512) 0 _________________________________________________________________flatten_1 (Flatten)(None, 8192) 0 _________________________________________________________________dense_1 (Dense) (None, 1) 8193_________________________________________________________________activation_1 (Activation) (None, 1) 0 =================================================================Total params: 4,311,553Trainable params: 4,311,553Non-trainable params: 0__________________________________________________________________________________________________________________________________Layer (type) Output Shape Param # =================================================================dense_2 (Dense) (None, 12544) 1266944 _________________________________________________________________batch_normalization_1 (Batch (None, 12544) 50176_________________________________________________________________activation_2 (Activation) (None, 12544) 0 _________________________________________________________________reshape_1 (Reshape)(None, 7, 7, 256) 0 _________________________________________________________________dropout_5 (Dropout)(None, 7, 7, 256) 0 _________________________________________________________________up_sampling2d_1 (UpSampling2 (None, 14, 14, 256) 0 _________________________________________________________________conv2d_transpose_1 (Conv2DTr (None, 14, 14, 128) 819328 _________________________________________________________________batch_normalization_2 (Batch (None, 14, 14, 128) 512 _________________________________________________________________activation_3 (Activation) (None, 14, 14, 128) 0 _________________________________________________________________up_sampling2d_2 (UpSampling2 (None, 28, 28, 128) 0 _________________________________________________________________conv2d_transpose_2 (Conv2DTr (None, 28, 28, 64) 204864 _________________________________________________________________batch_normalization_3 (Batch (None, 28, 28, 64) 256 _________________________________________________________________activation_4 (Activation) (None, 28, 28, 64) 0 _________________________________________________________________conv2d_transpose_3 (Conv2DTr (None, 28, 28, 32) 51232_________________________________________________________________batch_normalization_4 (Batch (None, 28, 28, 32) 128 _________________________________________________________________activation_5 (Activation) (None, 28, 28, 32) 0 _________________________________________________________________conv2d_transpose_4 (Conv2DTr (None, 28, 28, 1) 801 _________________________________________________________________activation_6 (Activation) (None, 28, 28, 1) 0 =================================================================Total params: 2,394,241Trainable params: 2,368,705Non-trainable params: 25,536_________________________________________________________________0: [D loss: 0.691503, acc: 0.509766] [A loss: 1.231536, acc: 0.000000]1: [D loss: 0.648653, acc: 0.548828] [A loss: 1.004777, acc: 0.000000]2: [D loss: 0.596343, acc: 0.515625] [A loss: 1.437418, acc: 0.000000]3: [D loss: 0.592514, acc: 0.843750] [A loss: 1.165042, acc: 0.000000]4: [D loss: 0.474580, acc: 0.982422] [A loss: 1.572894, acc: 0.000000]5: [D loss: 0.524854, acc: 0.693359] [A loss: 2.045968, acc: 0.000000]6: [D loss: 0.607588, acc: 0.810547] [A loss: 0.921388, acc: 0.015625]7: [D loss: 0.671562, acc: 0.500000] [A loss: 1.372429, acc: 0.000000]8: [D loss: 0.520482, acc: 0.906250] [A loss: 1.128514, acc: 0.000000]9: [D loss: 0.434005, acc: 0.986328] [A loss: 1.172277, acc: 0.000000]10: [D loss: 0.443416, acc: 0.644531] [A loss: 1.703157, acc: 0.000000]11: [D loss: 0.382728, acc: 0.996094] [A loss: 1.243012, acc: 0.000000]12: [D loss: 0.537050, acc: 0.505859] [A loss: 2.008118, acc: 0.000000]13: [D loss: 0.472494, acc: 0.931641] [A loss: 1.150631, acc: 0.000000]14: [D loss: 0.390675, acc: 0.759766] [A loss: 1.370212, acc: 0.000000]15: [D loss: 0.326801, acc: 0.978516] [A loss: 1.478633, acc: 0.000000]16: [D loss: 0.324425, acc: 0.916016] [A loss: 1.741925, acc: 0.000000]17: [D loss: 0.298051, acc: 0.990234] [A loss: 1.635591, acc: 0.000000]18: [D loss: 0.299104, acc: 0.947266] [A loss: 1.917285, acc: 0.000000]19: [D loss: 0.267050, acc: 0.998047] [A loss: 1.520450, acc: 0.000000]20: [D loss: 0.366637, acc: 0.742188] [A loss: 2.631558, acc: 0.000000]21: [D loss: 0.433646, acc: 0.892578] [A loss: 1.052351, acc: 0.027344]22: [D loss: 0.617413, acc: 0.515625] [A loss: 1.846913, acc: 0.000000]23: [D loss: 0.278645, acc: 0.990234] [A loss: 1.351792, acc: 0.000000]24: [D loss: 0.279705, acc: 0.947266] [A loss: 1.632707, acc: 0.000000]25: [D loss: 0.236295, acc: 0.980469] [A loss: 1.755901, acc: 0.000000]26: [D loss: 0.250655, acc: 0.949219] [A loss: 1.947745, acc: 0.000000]27: [D loss: 0.224082, acc: 0.972656] [A loss: 1.925455, acc: 0.000000]28: [D loss: 0.228702, acc: 0.957031] [A loss: 2.121093, acc: 0.000000]29: [D loss: 0.183285, acc: 0.990234] [A loss: 1.802444, acc: 0.003906]30: [D loss: 0.240770, acc: 0.912109] [A loss: 2.647219, acc: 0.000000]31: [D loss: 0.263249, acc: 0.923828] [A loss: 0.990316, acc: 0.210938]32: [D loss: 0.990905, acc: 0.507812] [A loss: 2.862743, acc: 0.000000]33: [D loss: 0.431388, acc: 0.785156] [A loss: 1.073982, acc: 0.078125]34: [D loss: 0.358037, acc: 0.789062] [A loss: 1.360353, acc: 0.011719]35: [D loss: 0.240701, acc: 0.937500] [A loss: 1.453409, acc: 0.019531]36: [D loss: 0.277398, acc: 0.886719] [A loss: 1.705918, acc: 0.000000]37: [D loss: 0.241611, acc: 0.916016] [A loss: 1.769321, acc: 0.000000]38: [D loss: 0.253498, acc: 0.910156] [A loss: 1.983329, acc: 0.000000]39: [D loss: 0.225967, acc: 0.937500] [A loss: 1.657594, acc: 0.007812]40: [D loss: 0.273817, acc: 0.863281] [A loss: 2.194973, acc: 0.000000]41: [D loss: 0.221860, acc: 0.937500] [A loss: 1.231366, acc: 0.117188]42: [D loss: 0.506532, acc: 0.693359] [A loss: 2.995471, acc: 0.000000]43: [D loss: 0.495057, acc: 0.800781] [A loss: 0.590665, acc: 0.667969]44: [D loss: 0.826440, acc: 0.544922] [A loss: 1.846615, acc: 0.000000]45: [D loss: 0.240079, acc: 0.955078] [A loss: 1.020400, acc: 0.167969]46: [D loss: 0.429716, acc: 0.730469] [A loss: 1.589278, acc: 0.003906]47: [D loss: 0.238475, acc: 0.947266] [A loss: 1.356752, acc: 0.031250]48: [D loss: 0.344801, acc: 0.822266] [A loss: 1.712172, acc: 0.000000]49: [D loss: 0.259987, acc: 0.925781] [A loss: 1.410086, acc: 0.031250]50: [D loss: 0.384457, acc: 0.775391] [A loss: 2.109657, acc: 0.000000]51: [D loss: 0.296866, acc: 0.933594] [A loss: 0.972942, acc: 0.214844]52: [D loss: 0.579416, acc: 0.615234] [A loss: 2.511665, acc: 0.000000]53: [D loss: 0.370514, acc: 0.871094] [A loss: 0.617159, acc: 0.640625]54: [D loss: 0.769498, acc: 0.521484] [A loss: 1.882442, acc: 0.000000]55: [D loss: 0.259458, acc: 0.962891] [A loss: 1.124723, acc: 0.058594]56: [D loss: 0.430123, acc: 0.708984] [A loss: 1.732257, acc: 0.000000]57: [D loss: 0.313927, acc: 0.929688] [A loss: 1.152099, acc: 0.058594]58: [D loss: 0.470944, acc: 0.654297] [A loss: 1.974094, acc: 0.000000]59: [D loss: 0.327888, acc: 0.931641] [A loss: 0.991130, acc: 0.140625]60: [D loss: 0.639045, acc: 0.548828] [A loss: 2.348823, acc: 0.000000]61: [D loss: 0.384606, acc: 0.886719] [A loss: 0.701185, acc: 0.500000]62: [D loss: 0.782717, acc: 0.509766] [A loss: 1.906247, acc: 0.000000]63: [D loss: 0.350297, acc: 0.949219] [A loss: 1.011624, acc: 0.117188]64: [D loss: 0.546892, acc: 0.558594] [A loss: 1.799624, acc: 0.000000]65: [D loss: 0.384208, acc: 0.894531] [A loss: 1.125789, acc: 0.023438]66: [D loss: 0.565075, acc: 0.568359] [A loss: 2.057138, acc: 0.000000]67: [D loss: 0.409379, acc: 0.916016] [A loss: 0.877637, acc: 0.234375]68: [D loss: 0.689376, acc: 0.509766] [A loss: 2.064061, acc: 0.000000]69: [D loss: 0.403014, acc: 0.923828] [A loss: 0.860366, acc: 0.246094]70: [D loss: 0.700391, acc: 0.505859] [A loss: 2.017132, acc: 0.000000]71: [D loss: 0.420937, acc: 0.904297] [A loss: 0.904990, acc: 0.167969]72: [D loss: 0.658957, acc: 0.517578] [A loss: 1.933370, acc: 0.000000]73: [D loss: 0.439174, acc: 0.884766] [A loss: 0.918563, acc: 0.199219]74: [D loss: 0.633506, acc: 0.525391] [A loss: 2.004735, acc: 0.000000]75: [D loss: 0.437182, acc: 0.896484] [A loss: 0.876891, acc: 0.238281]76: [D loss: 0.677424, acc: 0.503906] [A loss: 2.020369, acc: 0.000000]77: [D loss: 0.430446, acc: 0.906250] [A loss: 0.815220, acc: 0.281250]78: [D loss: 0.690905, acc: 0.503906] [A loss: 2.050816, acc: 0.000000]79: [D loss: 0.445981, acc: 0.908203] [A loss: 0.799167, acc: 0.296875]80: [D loss: 0.684389, acc: 0.509766] [A loss: 1.916862, acc: 0.000000]81: [D loss: 0.447047, acc: 0.902344] [A loss: 0.813182, acc: 0.312500]82: [D loss: 0.681568, acc: 0.507812] [A loss: 1.950821, acc: 0.000000]83: [D loss: 0.440485, acc: 0.912109] [A loss: 0.906062, acc: 0.175781]84: [D loss: 0.621085, acc: 0.519531] [A loss: 1.890153, acc: 0.000000]85: [D loss: 0.440503, acc: 0.888672] [A loss: 0.945012, acc: 0.167969]86: [D loss: 0.620250, acc: 0.525391] [A loss: 2.101747, acc: 0.000000]87: [D loss: 0.457391, acc: 0.902344] [A loss: 0.709829, acc: 0.480469]88: [D loss: 0.733050, acc: 0.503906] [A loss: 2.112121, acc: 0.000000]89: [D loss: 0.458123, acc: 0.888672] [A loss: 0.705255, acc: 0.464844]90: [D loss: 0.700676, acc: 0.507812] [A loss: 1.904836, acc: 0.000000]91: [D loss: 0.452364, acc: 0.906250] [A loss: 0.805733, acc: 0.316406]92: [D loss: 0.647448, acc: 0.507812] [A loss: 1.923860, acc: 0.000000]93: [D loss: 0.451022, acc: 0.916016] [A loss: 0.832103, acc: 0.253906]94: [D loss: 0.658411, acc: 0.515625] [A loss: 2.052168, acc: 0.000000]95: [D loss: 0.438787, acc: 0.929688] [A loss: 0.815914, acc: 0.308594]96: [D loss: 0.659617, acc: 0.507812] [A loss: 2.098220, acc: 0.000000]97: [D loss: 0.448281, acc: 0.902344] [A loss: 0.803287, acc: 0.320312]98: [D loss: 0.671589, acc: 0.509766] [A loss: 2.052079, acc: 0.000000]99: [D loss: 0.449434, acc: 0.898438] [A loss: 0.775447, acc: 0.367188]100: [D loss: 0.694258, acc: 0.507812] [A loss: 2.082814, acc: 0.000000]101: [D loss: 0.472326, acc: 0.886719] [A loss: 0.704720, acc: 0.503906]102: [D loss: 0.714512, acc: 0.505859] [A loss: 1.996335, acc: 0.000000]103: [D loss: 0.465975, acc: 0.882812] [A loss: 0.700260, acc: 0.519531]104: [D loss: 0.719334, acc: 0.500000] [A loss: 1.992613, acc: 0.000000]105: [D loss: 0.458248, acc: 0.902344] [A loss: 0.829929, acc: 0.277344]106: [D loss: 0.702113, acc: 0.507812] [A loss: 2.078796, acc: 0.000000]107: [D loss: 0.478760, acc: 0.888672] [A loss: 0.795521, acc: 0.320312]108: [D loss: 0.707462, acc: 0.501953] [A loss: 2.077059, acc: 0.000000]109: [D loss: 0.508267, acc: 0.833984] [A loss: 0.736822, acc: 0.480469]110: [D loss: 0.726499, acc: 0.501953] [A loss: 2.034654, acc: 0.000000]111: [D loss: 0.486403, acc: 0.876953] [A loss: 0.773986, acc: 0.375000]112: [D loss: 0.720257, acc: 0.500000] [A loss: 2.039338, acc: 0.000000]113: [D loss: 0.508873, acc: 0.833984] [A loss: 0.771436, acc: 0.398438]114: [D loss: 0.699827, acc: 0.505859] [A loss: 1.971889, acc: 0.000000]115: [D loss: 0.535446, acc: 0.796875] [A loss: 0.780918, acc: 0.335938]116: [D loss: 0.735963, acc: 0.501953] [A loss: 2.047856, acc: 0.000000]117: [D loss: 0.532679, acc: 0.806641] [A loss: 0.758681, acc: 0.402344]118: [D loss: 0.748309, acc: 0.507812] [A loss: 2.047386, acc: 0.000000]119: [D loss: 0.545521, acc: 0.783203] [A loss: 0.733404, acc: 0.429688]120: [D loss: 0.743108, acc: 0.501953] [A loss: 1.937115, acc: 0.000000]121: [D loss: 0.565786, acc: 0.765625] [A loss: 0.775433, acc: 0.390625]122: [D loss: 0.754094, acc: 0.503906] [A loss: 2.029332, acc: 0.000000]123: [D loss: 0.568998, acc: 0.763672] [A loss: 0.752774, acc: 0.398438]124: [D loss: 0.768901, acc: 0.505859] [A loss: 2.115191, acc: 0.000000]125: [D loss: 0.575241, acc: 0.751953] [A loss: 0.712867, acc: 0.468750]126: [D loss: 0.764882, acc: 0.503906] [A loss: 1.926544, acc: 0.000000]127: [D loss: 0.564186, acc: 0.775391] [A loss: 0.796656, acc: 0.316406]128: [D loss: 0.731247, acc: 0.501953] [A loss: 1.915877, acc: 0.000000]129: [D loss: 0.559361, acc: 0.785156] [A loss: 0.896263, acc: 0.175781]130: [D loss: 0.718427, acc: 0.517578] [A loss: 2.125160, acc: 0.000000]131: [D loss: 0.555089, acc: 0.763672] [A loss: 0.735170, acc: 0.449219]132: [D loss: 0.776140, acc: 0.498047] [A loss: 2.221849, acc: 0.000000]133: [D loss: 0.593860, acc: 0.707031] [A loss: 0.613928, acc: 0.679688]134: [D loss: 0.810136, acc: 0.496094] [A loss: 1.733192, acc: 0.000000]135: [D loss: 0.573212, acc: 0.767578] [A loss: 0.887116, acc: 0.167969]136: [D loss: 0.696365, acc: 0.503906] [A loss: 1.792508, acc: 0.000000]137: [D loss: 0.560364, acc: 0.789062] [A loss: 1.007125, acc: 0.093750]138: [D loss: 0.677562, acc: 0.525391] [A loss: 2.061942, acc: 0.000000]139: [D loss: 0.566491, acc: 0.773438] [A loss: 0.753326, acc: 0.425781]140: [D loss: 0.774658, acc: 0.503906] [A loss: 2.188344, acc: 0.000000]141: [D loss: 0.617926, acc: 0.673828] [A loss: 0.557792, acc: 0.816406]142: [D loss: 0.837046, acc: 0.494141] [A loss: 1.679428, acc: 0.000000]143: [D loss: 0.578102, acc: 0.775391] [A loss: 0.876964, acc: 0.187500]144: [D loss: 0.666602, acc: 0.517578] [A loss: 1.605749, acc: 0.000000]145: [D loss: 0.593821, acc: 0.705078] [A loss: 0.979484, acc: 0.074219]146: [D loss: 0.655855, acc: 0.527344] [A loss: 1.861180, acc: 0.000000]147: [D loss: 0.586890, acc: 0.761719] [A loss: 0.713094, acc: 0.457031]148: [D loss: 0.747398, acc: 0.509766] [A loss: 2.095814, acc: 0.000000]149: [D loss: 0.611762, acc: 0.689453] [A loss: 0.568410, acc: 0.789062]150: [D loss: 0.801940, acc: 0.500000] [A loss: 1.570221, acc: 0.000000]151: [D loss: 0.591073, acc: 0.750000] [A loss: 0.849759, acc: 0.222656]152: [D loss: 0.656044, acc: 0.517578] [A loss: 1.549002, acc: 0.000000]153: [D loss: 0.576636, acc: 0.728516] [A loss: 0.988255, acc: 0.113281]154: [D loss: 0.640300, acc: 0.537109] [A loss: 1.627752, acc: 0.000000]155: [D loss: 0.557738, acc: 0.806641] [A loss: 0.777479, acc: 0.328125]156: [D loss: 0.705224, acc: 0.500000] [A loss: 1.894259, acc: 0.000000]157: [D loss: 0.585216, acc: 0.716797] [A loss: 0.593588, acc: 0.726562]158: [D loss: 0.767763, acc: 0.496094] [A loss: 1.704471, acc: 0.000000]159: [D loss: 0.589467, acc: 0.722656] [A loss: 0.697022, acc: 0.464844]160: [D loss: 0.700470, acc: 0.500000] [A loss: 1.498897, acc: 0.000000]161: [D loss: 0.562630, acc: 0.796875] [A loss: 0.852995, acc: 0.207031]162: [D loss: 0.640407, acc: 0.515625] [A loss: 1.593685, acc: 0.000000]163: [D loss: 0.566859, acc: 0.751953] [A loss: 0.791467, acc: 0.324219]164: [D loss: 0.671670, acc: 0.513672] [A loss: 1.831024, acc: 0.000000]165: [D loss: 0.575037, acc: 0.750000] [A loss: 0.632219, acc: 0.687500]166: [D loss: 0.737036, acc: 0.500000] [A loss: 1.722258, acc: 0.000000]167: [D loss: 0.563757, acc: 0.777344] [A loss: 0.713480, acc: 0.460938]168: [D loss: 0.675642, acc: 0.505859] [A loss: 1.523493, acc: 0.000000]169: [D loss: 0.560679, acc: 0.787109] [A loss: 0.730296, acc: 0.421875]170: [D loss: 0.653577, acc: 0.523438] [A loss: 1.702063, acc: 0.000000]171: [D loss: 0.561906, acc: 0.798828] [A loss: 0.664845, acc: 0.578125]172: [D loss: 0.698300, acc: 0.509766] [A loss: 1.723585, acc: 0.000000]173: [D loss: 0.541326, acc: 0.826172] [A loss: 0.709710, acc: 0.476562]174: [D loss: 0.694211, acc: 0.505859] [A loss: 1.701803, acc: 0.000000]175: [D loss: 0.548754, acc: 0.806641] [A loss: 0.673307, acc: 0.574219]176: [D loss: 0.678137, acc: 0.517578] [A loss: 1.591114, acc: 0.000000]177: [D loss: 0.538395, acc: 0.826172] [A loss: 0.824897, acc: 0.250000]178: [D loss: 0.676258, acc: 0.519531] [A loss: 1.867238, acc: 0.000000]179: [D loss: 0.567595, acc: 0.751953] [A loss: 0.609377, acc: 0.710938]180: [D loss: 0.723833, acc: 0.503906] [A loss: 1.589976, acc: 0.000000]181: [D loss: 0.570847, acc: 0.791016] [A loss: 0.784557, acc: 0.312500]182: [D loss: 0.674772, acc: 0.521484] [A loss: 1.637760, acc: 0.000000]183: [D loss: 0.545820, acc: 0.833984] [A loss: 0.815327, acc: 0.253906]184: [D loss: 0.650002, acc: 0.527344] [A loss: 1.694071, acc: 0.000000]185: [D loss: 0.559047, acc: 0.783203] [A loss: 0.684570, acc: 0.550781]186: [D loss: 0.713607, acc: 0.503906] [A loss: 1.838257, acc: 0.000000]187: [D loss: 0.567683, acc: 0.730469] [A loss: 0.626130, acc: 0.675781]188: [D loss: 0.722959, acc: 0.503906] [A loss: 1.648201, acc: 0.000000]189: [D loss: 0.575623, acc: 0.777344] [A loss: 0.710549, acc: 0.503906]190: [D loss: 0.679402, acc: 0.527344] [A loss: 1.537004, acc: 0.000000]191: [D loss: 0.570763, acc: 0.783203] [A loss: 0.822068, acc: 0.277344]192: [D loss: 0.694234, acc: 0.525391] [A loss: 1.666622, acc: 0.000000]193: [D loss: 0.572525, acc: 0.763672] [A loss: 0.705894, acc: 0.503906]194: [D loss: 0.693591, acc: 0.523438] [A loss: 1.578833, acc: 0.000000]195: [D loss: 0.572520, acc: 0.771484] [A loss: 0.752325, acc: 0.421875]196: [D loss: 0.707219, acc: 0.517578] [A loss: 1.623304, acc: 0.000000]197: [D loss: 0.571517, acc: 0.771484] [A loss: 0.724463, acc: 0.480469]198: [D loss: 0.721095, acc: 0.511719] [A loss: 1.644762, acc: 0.000000]199: [D loss: 0.597435, acc: 0.677734] [A loss: 0.671912, acc: 0.585938]200: [D loss: 0.712119, acc: 0.517578] [A loss: 1.471917, acc: 0.000000]201: [D loss: 0.583933, acc: 0.757812] [A loss: 0.801758, acc: 0.289062]202: [D loss: 0.678271, acc: 0.531250] [A loss: 1.509630, acc: 0.000000]203: [D loss: 0.585138, acc: 0.750000] [A loss: 0.781951, acc: 0.312500]204: [D loss: 0.683118, acc: 0.519531] [A loss: 1.558321, acc: 0.000000]205: [D loss: 0.574745, acc: 0.753906] [A loss: 0.718502, acc: 0.449219]206: [D loss: 0.689886, acc: 0.509766] [A loss: 1.646207, acc: 0.000000]207: [D loss: 0.579706, acc: 0.759766] [A loss: 0.710435, acc: 0.503906]208: [D loss: 0.709791, acc: 0.513672] [A loss: 1.577824, acc: 0.000000]209: [D loss: 0.581872, acc: 0.740234] [A loss: 0.666501, acc: 0.589844]210: [D loss: 0.696718, acc: 0.515625] [A loss: 1.595843, acc: 0.000000]211: [D loss: 0.582528, acc: 0.744141] [A loss: 0.706088, acc: 0.492188]212: [D loss: 0.692366, acc: 0.519531] [A loss: 1.561489, acc: 0.000000]213: [D loss: 0.584884, acc: 0.742188] [A loss: 0.706848, acc: 0.484375]214: [D loss: 0.684679, acc: 0.525391] [A loss: 1.480523, acc: 0.000000]215: [D loss: 0.583968, acc: 0.757812] [A loss: 0.779733, acc: 0.339844]216: [D loss: 0.669083, acc: 0.527344] [A loss: 1.524182, acc: 0.000000]217: [D loss: 0.580016, acc: 0.783203] [A loss: 0.707451, acc: 0.496094]218: [D loss: 0.702822, acc: 0.519531] [A loss: 1.549842, acc: 0.000000]219: [D loss: 0.608712, acc: 0.720703] [A loss: 0.738430, acc: 0.429688]220: [D loss: 0.683573, acc: 0.517578] [A loss: 1.524086, acc: 0.000000]221: [D loss: 0.607246, acc: 0.693359] [A loss: 0.702375, acc: 0.515625]222: [D loss: 0.704177, acc: 0.525391] [A loss: 1.657308, acc: 0.000000]223: [D loss: 0.603331, acc: 0.703125] [A loss: 0.610635, acc: 0.710938]224: [D loss: 0.717118, acc: 0.501953] [A loss: 1.430291, acc: 0.000000]225: [D loss: 0.601408, acc: 0.744141] [A loss: 0.758834, acc: 0.402344]226: [D loss: 0.682414, acc: 0.527344] [A loss: 1.442136, acc: 0.000000]227: [D loss: 0.592158, acc: 0.750000] [A loss: 0.733325, acc: 0.425781]228: [D loss: 0.689754, acc: 0.527344] [A loss: 1.474277, acc: 0.000000]229: [D loss: 0.600397, acc: 0.712891] [A loss: 0.754566, acc: 0.414062]230: [D loss: 0.702799, acc: 0.521484] [A loss: 1.466574, acc: 0.000000]231: [D loss: 0.609971, acc: 0.718750] [A loss: 0.753828, acc: 0.437500]232: [D loss: 0.696546, acc: 0.527344] [A loss: 1.433056, acc: 0.000000]233: [D loss: 0.603584, acc: 0.734375] [A loss: 0.729402, acc: 0.472656]234: [D loss: 0.691293, acc: 0.533203] [A loss: 1.500463, acc: 0.000000]235: [D loss: 0.612721, acc: 0.679688] [A loss: 0.748835, acc: 0.410156]236: [D loss: 0.709877, acc: 0.515625] [A loss: 1.482194, acc: 0.000000]237: [D loss: 0.603751, acc: 0.728516] [A loss: 0.715858, acc: 0.503906]238: [D loss: 0.710946, acc: 0.515625] [A loss: 1.565704, acc: 0.000000]239: [D loss: 0.613677, acc: 0.685547] [A loss: 0.616414, acc: 0.675781]240: [D loss: 0.744525, acc: 0.505859] [A loss: 1.453091, acc: 0.003906]241: [D loss: 0.635414, acc: 0.648438] [A loss: 0.671547, acc: 0.585938]242: [D loss: 0.704229, acc: 0.505859] [A loss: 1.248188, acc: 0.000000]243: [D loss: 0.620378, acc: 0.679688] [A loss: 0.802517, acc: 0.269531]244: [D loss: 0.676368, acc: 0.554688] [A loss: 1.273573, acc: 0.007812]245: [D loss: 0.627741, acc: 0.697266] [A loss: 0.805003, acc: 0.277344]246: [D loss: 0.680954, acc: 0.535156] [A loss: 1.364167, acc: 0.000000]247: [D loss: 0.617311, acc: 0.708984] [A loss: 0.779782, acc: 0.324219]248: [D loss: 0.704194, acc: 0.525391] [A loss: 1.512358, acc: 0.000000]249: [D loss: 0.618787, acc: 0.705078] [A loss: 0.656829, acc: 0.621094]250: [D loss: 0.749846, acc: 0.511719] [A loss: 1.544242, acc: 0.000000]251: [D loss: 0.619732, acc: 0.691406] [A loss: 0.703778, acc: 0.523438]252: [D loss: 0.743708, acc: 0.511719] [A loss: 1.425920, acc: 0.000000]253: [D loss: 0.626959, acc: 0.660156] [A loss: 0.713868, acc: 0.492188]254: [D loss: 0.715720, acc: 0.513672] [A loss: 1.382949, acc: 0.000000]255: [D loss: 0.630420, acc: 0.667969] [A loss: 0.766442, acc: 0.335938]256: [D loss: 0.706938, acc: 0.531250] [A loss: 1.348288, acc: 0.003906]257: [D loss: 0.633038, acc: 0.687500] [A loss: 0.775958, acc: 0.332031]258: [D loss: 0.696131, acc: 0.521484] [A loss: 1.309652, acc: 0.003906]259: [D loss: 0.622492, acc: 0.669922] [A loss: 0.793313, acc: 0.285156]260: [D loss: 0.685976, acc: 0.527344] [A loss: 1.376491, acc: 0.000000]261: [D loss: 0.638133, acc: 0.673828] [A loss: 0.790920, acc: 0.355469]262: [D loss: 0.679706, acc: 0.544922] [A loss: 1.427657, acc: 0.000000]263: [D loss: 0.644071, acc: 0.628906] [A loss: 0.745437, acc: 0.433594]264: [D loss: 0.732221, acc: 0.503906] [A loss: 1.472740, acc: 0.000000]265: [D loss: 0.658985, acc: 0.611328] [A loss: 0.653982, acc: 0.589844]266: [D loss: 0.738794, acc: 0.507812] [A loss: 1.425443, acc: 0.000000]267: [D loss: 0.661236, acc: 0.578125] [A loss: 0.620909, acc: 0.738281]268: [D loss: 0.729037, acc: 0.503906] [A loss: 1.274895, acc: 0.003906]269: [D loss: 0.646045, acc: 0.638672] [A loss: 0.715880, acc: 0.460938]270: [D loss: 0.718505, acc: 0.513672] [A loss: 1.197851, acc: 0.003906]271: [D loss: 0.641561, acc: 0.656250] [A loss: 0.762347, acc: 0.355469]272: [D loss: 0.680380, acc: 0.533203] [A loss: 1.152511, acc: 0.007812]273: [D loss: 0.655355, acc: 0.642578] [A loss: 0.818601, acc: 0.261719]274: [D loss: 0.679781, acc: 0.564453] [A loss: 1.128714, acc: 0.011719]275: [D loss: 0.655671, acc: 0.628906] [A loss: 0.856017, acc: 0.191406]276: [D loss: 0.690292, acc: 0.511719] [A loss: 1.310163, acc: 0.000000]277: [D loss: 0.647914, acc: 0.660156] [A loss: 0.750118, acc: 0.390625]278: [D loss: 0.704151, acc: 0.513672] [A loss: 1.374055, acc: 0.000000]279: [D loss: 0.654893, acc: 0.621094] [A loss: 0.650132, acc: 0.625000]280: [D loss: 0.729345, acc: 0.498047] [A loss: 1.403211, acc: 0.000000]281: [D loss: 0.657251, acc: 0.605469] [A loss: 0.647218, acc: 0.632812]282: [D loss: 0.732492, acc: 0.503906] [A loss: 1.268551, acc: 0.003906]283: [D loss: 0.662744, acc: 0.626953] [A loss: 0.689310, acc: 0.558594]284: [D loss: 0.700425, acc: 0.511719] [A loss: 1.162058, acc: 0.003906]285: [D loss: 0.651830, acc: 0.623047] [A loss: 0.780479, acc: 0.312500]286: [D loss: 0.691948, acc: 0.525391] [A loss: 1.133446, acc: 0.011719]287: [D loss: 0.650677, acc: 0.644531] [A loss: 0.793982, acc: 0.304688]288: [D loss: 0.698428, acc: 0.515625] [A loss: 1.247315, acc: 0.000000]289: [D loss: 0.657979, acc: 0.648438] [A loss: 0.722180, acc: 0.449219]290: [D loss: 0.701157, acc: 0.517578] [A loss: 1.253019, acc: 0.003906]291: [D loss: 0.649832, acc: 0.644531] [A loss: 0.701979, acc: 0.503906]292: [D loss: 0.712979, acc: 0.519531] [A loss: 1.271342, acc: 0.000000]293: [D loss: 0.660288, acc: 0.621094] [A loss: 0.689836, acc: 0.550781]294: [D loss: 0.726884, acc: 0.505859] [A loss: 1.93, acc: 0.003906]295: [D loss: 0.663560, acc: 0.617188] [A loss: 0.701607, acc: 0.488281]296: [D loss: 0.707670, acc: 0.507812] [A loss: 1.129499, acc: 0.003906]297: [D loss: 0.663256, acc: 0.617188] [A loss: 0.739957, acc: 0.371094]298: [D loss: 0.689587, acc: 0.523438] [A loss: 1.191701, acc: 0.000000]299: [D loss: 0.660023, acc: 0.623047] [A loss: 0.746944, acc: 0.371094]300: [D loss: 0.684651, acc: 0.531250] [A loss: 1.150684, acc: 0.000000]301: [D loss: 0.651196, acc: 0.642578] [A loss: 0.808909, acc: 0.246094]302: [D loss: 0.678586, acc: 0.552734] [A loss: 1.157491, acc: 0.003906]303: [D loss: 0.652035, acc: 0.646484] [A loss: 0.783414, acc: 0.289062]304: [D loss: 0.671595, acc: 0.564453] [A loss: 1.269552, acc: 0.000000]305: [D loss: 0.648778, acc: 0.662109] [A loss: 0.721312, acc: 0.453125]306: [D loss: 0.716662, acc: 0.517578] [A loss: 1.334316, acc: 0.000000]307: [D loss: 0.653186, acc: 0.613281] [A loss: 0.611153, acc: 0.742188]308: [D loss: 0.724283, acc: 0.505859] [A loss: 1.280901, acc: 0.000000]309: [D loss: 0.657101, acc: 0.617188] [A loss: 0.631597, acc: 0.683594]310: [D loss: 0.730646, acc: 0.500000] [A loss: 1.158158, acc: 0.000000]311: [D loss: 0.657883, acc: 0.625000] [A loss: 0.732748, acc: 0.410156]312: [D loss: 0.703416, acc: 0.523438] [A loss: 1.026374, acc: 0.015625]313: [D loss: 0.669057, acc: 0.599609] [A loss: 0.774714, acc: 0.257812]314: [D loss: 0.681841, acc: 0.539062] [A loss: 1.001734, acc: 0.019531]315: [D loss: 0.658778, acc: 0.632812] [A loss: 0.791964, acc: 0.246094]316: [D loss: 0.682347, acc: 0.537109] [A loss: 1.017623, acc: 0.011719]317: [D loss: 0.662230, acc: 0.595703] [A loss: 0.856697, acc: 0.121094]318: [D loss: 0.676194, acc: 0.546875] [A loss: 1.032536, acc: 0.019531]319: [D loss: 0.653828, acc: 0.617188] [A loss: 0.825725, acc: 0.207031]320: [D loss: 0.668046, acc: 0.556641] [A loss: 1.124810, acc: 0.015625]。。。。。。。。。。。。。。。。。。。。。..。9919: [D loss: 0.702338, acc: 0.517578] [A loss: 0.747043, acc: 0.378906]9920: [D loss: 0.703451, acc: 0.494141] [A loss: 0.835588, acc: 0.164062]9921: [D loss: 0.693855, acc: 0.517578] [A loss: 0.730508, acc: 0.410156]9922: [D loss: 0.715099, acc: 0.472656] [A loss: 0.826207, acc: 0.195312]9923: [D loss: 0.706640, acc: 0.474609] [A loss: 0.751718, acc: 0.335938]9924: [D loss: 0.691947, acc: 0.550781] [A loss: 0.813644, acc: 0.226562]9925: [D loss: 0.697594, acc: 0.513672] [A loss: 0.768750, acc: 0.269531]9926: [D loss: 0.689501, acc: 0.552734] [A loss: 0.768900, acc: 0.335938]9927: [D loss: 0.696636, acc: 0.531250] [A loss: 0.772304, acc: 0.316406]9928: [D loss: 0.697596, acc: 0.523438] [A loss: 0.806830, acc: 0.261719]9929: [D loss: 0.697859, acc: 0.505859] [A loss: 0.727642, acc: 0.433594]9930: [D loss: 0.697926, acc: 0.523438] [A loss: 0.879466, acc: 0.101562]9931: [D loss: 0.697527, acc: 0.531250] [A loss: 0.701799, acc: 0.488281]9932: [D loss: 0.710694, acc: 0.490234] [A loss: 0.855742, acc: 0.128906]9933: [D loss: 0.705487, acc: 0.521484] [A loss: 0.727056, acc: 0.417969]9934: [D loss: 0.700702, acc: 0.496094] [A loss: 0.823942, acc: 0.179688]9935: [D loss: 0.704966, acc: 0.498047] [A loss: 0.767886, acc: 0.281250]9936: [D loss: 0.701806, acc: 0.509766] [A loss: 0.798015, acc: 0.238281]9937: [D loss: 0.715728, acc: 0.478516] [A loss: 0.776415, acc: 0.292969]9938: [D loss: 0.697116, acc: 0.511719] [A loss: 0.829303, acc: 0.195312]9939: [D loss: 0.699822, acc: 0.509766] [A loss: 0.740921, acc: 0.375000]9940: [D loss: 0.719507, acc: 0.478516] [A loss: 0.859272, acc: 0.160156]9941: [D loss: 0.695651, acc: 0.525391] [A loss: 0.704798, acc: 0.503906]9942: [D loss: 0.707328, acc: 0.503906] [A loss: 0.808904, acc: 0.218750]9943: [D loss: 0.688593, acc: 0.548828] [A loss: 0.744065, acc: 0.394531]9944: [D loss: 0.699065, acc: 0.523438] [A loss: 0.802332, acc: 0.257812]9945: [D loss: 0.698068, acc: 0.525391] [A loss: 0.745907, acc: 0.347656]9946: [D loss: 0.701033, acc: 0.498047] [A loss: 0.813438, acc: 0.199219]9947: [D loss: 0.691200, acc: 0.541016] [A loss: 0.756370, acc: 0.343750]9948: [D loss: 0.712499, acc: 0.476562] [A loss: 0.830884, acc: 0.136719]9949: [D loss: 0.696183, acc: 0.509766] [A loss: 0.704277, acc: 0.488281]9950: [D loss: 0.696827, acc: 0.523438] [A loss: 0.825646, acc: 0.183594]9951: [D loss: 0.692385, acc: 0.529297] [A loss: 0.743277, acc: 0.378906]9952: [D loss: 0.704989, acc: 0.513672] [A loss: 0.853841, acc: 0.144531]9953: [D loss: 0.703750, acc: 0.476562] [A loss: 0.710911, acc: 0.484375]9954: [D loss: 0.714116, acc: 0.507812] [A loss: 0.949648, acc: 0.078125]9955: [D loss: 0.694516, acc: 0.525391] [A loss: 0.714422, acc: 0.453125]9956: [D loss: 0.710579, acc: 0.500000] [A loss: 0.842483, acc: 0.171875]9957: [D loss: 0.693028, acc: 0.515625] [A loss: 0.750638, acc: 0.355469]9958: [D loss: 0.707715, acc: 0.515625] [A loss: 0.801792, acc: 0.230469]9959: [D loss: 0.706937, acc: 0.507812] [A loss: 0.780088, acc: 0.292969]9960: [D loss: 0.704332, acc: 0.519531] [A loss: 0.721530, acc: 0.414062]9961: [D loss: 0.708584, acc: 0.492188] [A loss: 0.830755, acc: 0.191406]9962: [D loss: 0.699920, acc: 0.500000] [A loss: 0.730089, acc: 0.402344]9963: [D loss: 0.709511, acc: 0.482422] [A loss: 0.819024, acc: 0.214844]9964: [D loss: 0.698606, acc: 0.513672] [A loss: 0.775723, acc: 0.296875]9965: [D loss: 0.707085, acc: 0.474609] [A loss: 0.83, acc: 0.226562]9966: [D loss: 0.708760, acc: 0.500000] [A loss: 0.773791, acc: 0.296875]9967: [D loss: 0.714988, acc: 0.501953] [A loss: 0.876916, acc: 0.128906]9968: [D loss: 0.699741, acc: 0.527344] [A loss: 0.761598, acc: 0.339844]9969: [D loss: 0.709108, acc: 0.505859] [A loss: 0.784633, acc: 0.292969]9970: [D loss: 0.704452, acc: 0.476562] [A loss: 0.811802, acc: 0.210938]9971: [D loss: 0.706212, acc: 0.490234] [A loss: 0.733744, acc: 0.398438]9972: [D loss: 0.709742, acc: 0.486328] [A loss: 0.834597, acc: 0.183594]9973: [D loss: 0.698140, acc: 0.531250] [A loss: 0.685229, acc: 0.550781]9974: [D loss: 0.711616, acc: 0.492188] [A loss: 0.845335, acc: 0.140625]9975: [D loss: 0.693955, acc: 0.531250] [A loss: 0.699906, acc: 0.519531]9976: [D loss: 0.705078, acc: 0.515625] [A loss: 0.821778, acc: 0.203125]9977: [D loss: 0.697455, acc: 0.513672] [A loss: 0.714708, acc: 0.457031]9978: [D loss: 0.705634, acc: 0.500000] [A loss: 0.791556, acc: 0.261719]9979: [D loss: 0.702151, acc: 0.509766] [A loss: 0.729714, acc: 0.398438]9980: [D loss: 0.710843, acc: 0.492188] [A loss: 0.803456, acc: 0.218750]9981: [D loss: 0.698539, acc: 0.500000] [A loss: 0.743020, acc: 0.351562]9982: [D loss: 0.702784, acc: 0.507812] [A loss: 0.829923, acc: 0.164062]9983: [D loss: 0.710229, acc: 0.480469] [A loss: 0.726384, acc: 0.417969]9984: [D loss: 0.711040, acc: 0.503906] [A loss: 0.807429, acc: 0.207031]9985: [D loss: 0.696956, acc: 0.533203] [A loss: 0.755409, acc: 0.363281]9986: [D loss: 0.700728, acc: 0.488281] [A loss: 0.772643, acc: 0.304688]9987: [D loss: 0.704541, acc: 0.492188] [A loss: 0.800799, acc: 0.261719]9988: [D loss: 0.700189, acc: 0.521484] [A loss: 0.770434, acc: 0.320312]9989: [D loss: 0.697253, acc: 0.509766] [A loss: 0.797975, acc: 0.242188]9990: [D loss: 0.706917, acc: 0.492188] [A loss: 0.883894, acc: 0.128906]9991: [D loss: 0.695292, acc: 0.505859] [A loss: 0.720615, acc: 0.433594]9992: [D loss: 0.698072, acc: 0.535156] [A loss: 0.877171, acc: 0.132812]9993: [D loss: 0.691101, acc: 0.515625] [A loss: 0.722928, acc: 0.472656]9994: [D loss: 0.703333, acc: 0.533203] [A loss: 0.944150, acc: 0.082031]9995: [D loss: 0.699371, acc: 0.521484] [A loss: 0.715438, acc: 0.464844]9996: [D loss: 0.710147, acc: 0.505859] [A loss: 0.822873, acc: 0.191406]9997: [D loss: 0.708276, acc: 0.480469] [A loss: 0.714018, acc: 0.453125]9998: [D loss: 0.701492, acc: 0.507812] [A loss: 0.802196, acc: 0.257812]9999: [D loss: 0.695118, acc: 0.513672] [A loss: 0.777079, acc: 0.292969]Elapsed: 2.9090983357694413 hr

我们的判别器的工作原理是这样的：我们先用生成器生成一堆图片便签为0（假）在用真实图片便签为1（真），然后拿给判别器训练。而对抗模型则是用生成器的产物便签为1，让判别器去判别。数据中D就代表判别器，A代表对抗模型我们可以看到一开始判别器，准确率很低，因为他不是一个好警察，然后上升到90左右说明他已经是一个好警察了，但是最后准确率不断下降，这并不是因为他的水平下降，而是我们的生成器愈发完美，成了“神偷”达到了我们的目的。在看对抗模型，也可以看出我们的警察不能识别出小偷造的假币了。所以GAN是很成功的。