MXNet网络模型（四）GAN神经网络

概述

GAN 神经网络(2014)

• 《Generative Adversarial Network》
• https://arxiv.org/pdf/1406.2661.pdf

GAN网络也是一直在进步

原理

GAN训练可以分为5步

1. 样本图片x输入分类器
2. 随机种子z经由生成器生成模拟图片x*
3. 模拟图片x*输入分类器
4. 反馈给分类器
5. 反馈给生成器

MXNet代码

导入库

import time
import gzip
import numpy as np
import matplotlib.pyplot as plt![请添加图片描述](https://img-blog.csdnimg.cn/153a1da1e6d8438681cddca1fae72021.bmp)

import mxnet as mx

环境和批大小

batch_size = 10
device = mx.cpu()

读取训练数据

def load_dataset():
    transform = mx.gluon.data.vision.transforms.ToTensor()
    
    train_img = [ transform(img).asnumpy() for img, lbl in mx.gluon.data.vision.datasets.MNIST(root='mnist', train=True )]
    train_lbl = [ np.array(lbl)            for img, lbl in mx.gluon.data.vision.datasets.MNIST(root='mnist', train=True )]
    eval_img  = [ transform(img).asnumpy() for img, lbl in mx.gluon.data.vision.datasets.MNIST(root='mnist', train=False)]
    eval_lbl  = [ np.array(lbl)            for img, lbl in mx.gluon.data.vision.datasets.MNIST(root='mnist', train=False)]
    
    return train_img, train_lbl, eval_img, eval_lbl

train_img, train_lbl, eval_img, eval_lbl = load_dataset()

train_data = mx.gluon.data.DataLoader(
    mx.gluon.data.ArrayDataset(train_img, train_lbl),
    batch_size=batch_size,
    shuffle=True
)
eval_data = mx.gluon.data.DataLoader(
    mx.gluon.data.ArrayDataset(eval_img, eval_lbl),
    batch_size=batch_size,
    shuffle=False
)

预览训练数据

idxs = (25, 47, 74, 88, 92)
for i in range(5):
    plt.subplot(1, 5, i + 1)
    idx = idxs[i]
    plt.xticks([])
    plt.yticks([])
    img = train_img[idx][0].astype( np.float32 )
    plt.imshow(img, interpolation='none', cmap='Blues')
plt.show()

分类器

class Discriminator():
    def __init__(self):
        self.loss_fn = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
        self.metric = mx.metric.Loss()
        
        self.net = mx.gluon.nn.HybridSequential()
        # 第一层
        self.net.add(
            mx.gluon.nn.Flatten(),
            mx.gluon.nn.Dense(units=200),
            mx.gluon.nn.LeakyReLU(alpha=0.02),
            mx.gluon.nn.LayerNorm()
        )
        # 第二层
        self.net.add(
            mx.gluon.nn.Dense(units=1),
        )
            
        self.net.initialize( init=mx.init.Xavier(rnd_type='gaussian'), ctx=device )
        self.trainer = mx.gluon.Trainer(
            params=self.net.collect_params(),
            optimizer=mx.optimizer.Adam(learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-08, lazy_update=True)
        )
        self.net.summary(mx.ndarray.zeros(shape=(50, 1, 28, 28), dtype=np.float32, ctx=device))
    
discriminator = Discriminator()

--------------------------------------------------------------------------------
        Layer (type)                                Output Shape         Param #
================================================================================
               Input                             (50, 1, 28, 28)               0
           Flatten-1                                   (50, 784)               0
             Dense-2                                   (50, 200)          157000
         LeakyReLU-3                                   (50, 200)               0
         LayerNorm-4                                   (50, 200)             400
             Dense-5                                     (50, 1)             201
================================================================================
Parameters in forward computation graph, duplicate included
   Total params: 157601
   Trainable params: 157601
   Non-trainable params: 0
Shared params in forward computation graph: 0
Unique parameters in model: 157601
--------------------------------------------------------------------------------

生成器

class Generator():
    def __init__(self):
        #self.loss_fn = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
        self.metric = mx.metric.Loss()
        
        self.net = mx.gluon.nn.HybridSequential()
        # 第一层
        self.net.add(
            mx.gluon.nn.Dense(units=200),
            mx.gluon.nn.LeakyReLU(alpha=0.02),
            mx.gluon.nn.LayerNorm(),
        )
        # 第二层
        self.net.add(
            mx.gluon.nn.Dense(units=784),
            mx.gluon.nn.Activation(activation='sigmoid'),
            mx.gluon.nn.HybridLambda(lambda F, x: F.reshape(x, shape=(0, -1, 28, 28)))
        )
            
        self.net.initialize( init=mx.init.Xavier(rnd_type='gaussian'), ctx=device )
        self.trainer = mx.gluon.Trainer(
            params=self.net.collect_params(),
            optimizer=mx.optimizer.Adam(learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-08, lazy_update=True)
        )
        self.net.summary(mx.ndarray.zeros(shape=(50, 100), dtype=np.float32, ctx=device))

generator = Generator()

生成种子

def make_seed(batach, num, device=None):
    if device is None:
        device = mx.cpu()
    return mx.ndarray.normal(loc=0, scale=1, shape=(batach, num), ctx=device)

训练

for epoch in range(120):
    
    discriminator.metric.reset(); discriminator.metric.reset(); tic = time.time()
    for datas, _ in train_data:
        # 批大小
        actually_batch_size = datas.shape[0]
        # CPU 移到 GPU
        datas  = mx.gluon.utils.split_and_load( datas,  [device] )
        # 生成种子
        seeds = [make_seed(data.shape[0], 100, device) for data in datas]
        
        # 训练鉴别器
        for data, seed in zip(datas, seeds):
            length = data.shape[0]
            lbl_real = mx.ndarray.ones(shape=(length,1), ctx=device)
            lbl_fake = mx.ndarray.zeros(shape=(length,1), ctx=device)

            with mx.autograd.record():
                a = discriminator.loss_fn(discriminator.net(data), lbl_real)
                b = discriminator.loss_fn(discriminator.net(generator.net(seed).detach()), lbl_fake)
                d_loss = a + b
            d_loss.backward()
            discriminator.metric.update(_, preds=d_loss)
        discriminator.trainer.step(actually_batch_size)
        
        # 训练生成器
        for seed in seeds:
            length = seed.shape[0]
            lbl_real = mx.ndarray.ones(shape=(length,1), ctx=device)
            
            with mx.autograd.record():
                img = generator.net(seed)
                g_loss = discriminator.loss_fn(discriminator.net(img), lbl_real)
            g_loss.backward()
            generator.metric.update(_, preds=g_loss)
        generator.trainer.step(actually_batch_size)
    
    print("Epoch {:>2d}: cost:{:.1f}s d_loss:{:.3f} g_loss:{:.3f}".format(epoch, time.time()-tic, discriminator.metric.get()[1], generator.metric.get()[1]))
    
    for i in range(10):
        plt.subplot(1, 10, i + 1)
        plt.xticks([]); plt.yticks([])
        seed = make_seed(length, 100, device)
        img = generator.net(seed).asnumpy()
        plt.imshow(img[0][0], interpolation='none', cmap='Blues')
    plt.show()

Epoch 0: cost:43.4s d_loss:0.372 g_loss:4.094
Epoch 1: cost:40.2s d_loss:0.602 g_loss:3.242
Epoch 2: cost:40.5s d_loss:0.661 g_loss:2.887
Epoch 3: cost:40.2s d_loss:0.701 g_loss:2.688
Epoch 4: cost:40.3s d_loss:0.707 g_loss:2.563
Epoch 5: cost:39.2s d_loss:0.708 g_loss:2.480
Epoch 6: cost:52.1s d_loss:0.703 g_loss:2.422
Epoch 7: cost:44.2s d_loss:0.691 g_loss:2.384
Epoch 8: cost:53.6s d_loss:0.675 g_loss:2.359
Epoch 9: cost:38.2s d_loss:0.659 g_loss:2.344
Epoch 10: cost:38.2s d_loss:0.645 g_loss:2.333
Epoch 11: cost:38.2s d_loss:0.640 g_loss:2.325
Epoch 12: cost:38.1s d_loss:0.625 g_loss:2.319
Epoch 13: cost:38.3s d_loss:0.619 g_loss:2.313
Epoch 14: cost:38.3s d_loss:0.610 g_loss:2.308
Epoch 15: cost:38.2s d_loss:0.606 g_loss:2.304
Epoch 16: cost:38.2s d_loss:0.598 g_loss:2.301
Epoch 17: cost:38.3s d_loss:0.590 g_loss:2.298
Epoch 18: cost:38.2s d_loss:0.589 g_loss:2.296
Epoch 19: cost:38.5s d_loss:0.585 g_loss:2.294
Epoch 20: cost:38.3s d_loss:0.585 g_loss:2.292

训练20轮输出如图

损失值评价

最理想的分离器损失值为
entropy = -1 x ln(0.5) = 0.693

训练早期（前6轮），分类器损失值迅速从0.372上升到0.708。同时生成器损失值也在迅速下降。此时生成器轻微领先。

随后通过继续学习（7-26轮），分类器和生成器的损失值都在下降，双方不断进步。

之后开始（27轮），分类器损失继续下降，生成器损失上升，说明分类器已经抛开生成器。