低分辨率图像重建
任务总览数据加载与配置模型设置生成、判别、特征提取模块调用损失函数与训练测试今天我们来介绍利用对抗生成网络（GAN）对低分辨率图像进行重构的介绍。再开始今天的任务之前，给大家强调一下，我们需要使用1.x.x版本的tensorflow和tensorlayer，我是用的是3.6版本的python，3.4.1.15版本的opencv以及1.8.0版本的tensorflow和tensorlayer。另外还有其他的一些模块需要安装，直接按照错误提示安装即可。
任务总览
分辨率在图片中的直接反应就是图像的大小，分辨率越高，图像的初始大小越大。如果将不同分辨率的图像放缩到同样的大小，分辨率低的图像会更模糊。超分辨率重构就是将分辨率低的图片重构成清晰的高分辨率图像：
所需要用到的网络结构图为：
数据加载与配置
这个部分对应着生成网络和判别网络的input部分的初始化。
首先需要大家下载srgan任务，打开config文件，我们主要的参数都将在这个文件中进行修改：
from easydict import EasyDict as edictimport jsonconfig = edict()config.TRAIN = edict()## Adam# batch设置过大有可能会引发内存不足的报错config.TRAIN.batch_size = 4 # 可以适当调整config.TRAIN.lr_init = 1e-4config.TRAIN.beta1 = 0.9## 初始化生成器config.TRAIN.n_epoch_init = 100## 判别器学习 (SRGAN)config.TRAIN.n_epoch = 2000config.TRAIN.lr_decay = 0.1config.TRAIN.decay_every = int(config.TRAIN.n_epoch / 2)# 训练集路径指定config.TRAIN.hr_img_path = 'E:\srgan\srdata\srdata\DIV2K_train_HR'config.TRAIN.lr_img_path = 'E:\srgan\srdata\srdata\DIV2K_train_LR_bicubic\X4'config.VALID = edict()# 测试集路径制定config.VALID.hr_img_path = 'E:\srgan\srdata\srdata\DIV2K_valid_HR'config.VALID.lr_img_path = 'E:\srgan\srdata\srdata\DIV2K_valid_LR_bicubic\X4'def log_config(filename, cfg):with open(filename, 'w') as f:f.write("================================================\n")f.write(json.dumps(cfg, indent=4))f.write("\n================================================\n")
这里改好之后，我们需要对一些main.py文件里的函数进行一些设置，比如传递进batch_size，学习率，epoch等，同时要指定好生成的图像以及模型等文件的存储位置，之后把再图像读取进来：
import osimport timeimport pickle, randomimport numpy as npimport logging, scipyimport tensorflow as tfimport tensorlayer as tlfrom model import SRGAN_g, SRGAN_d, Vgg19_simple_apifrom utils import *from config import config, log_config## Adambatch_size = config.TRAIN.batch_size # 4lr_init = config.TRAIN.lr_init # 1e-4beta1 = config.TRAIN.beta1 # 0.9## 初始化生成器n_epoch_init = config.TRAIN.n_epoch_init # 100## 判别器学习(SRGAN)n_epoch = config.TRAIN.n_epoch # 2000lr_decay = config.TRAIN.lr_decay # 0.1decay_every = config.TRAIN.decay_every # 1000ni = int(np.sqrt(batch_size))def train():## 创建文件夹保存结果图像和训练模型save_dir_ginit = "samples/{}_ginit".format(tl.global_flag['mode'])save_dir_gan = "samples/{}_gan".format(tl.global_flag['mode'])tl.files.exists_or_mkdir(save_dir_ginit)tl.files.exists_or_mkdir(save_dir_gan)checkpoint_dir = "checkpoint" # checkpoint_resize_convtl.files.exists_or_mkdir(checkpoint_dir)# load_file_list可以把所有的文件都加载进来# path指定文件夹的路径# regx='.*.png'代表读取所有.png的文件train_hr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.hr_img_path, regx='.*.png', printable=False))[:800] # 如果出现memory error可以这样操作减少一次读取的数据量train_lr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.lr_img_path, regx='.*.png', printable=False)) # 不加切片读取也是可以的，但一定要注意传入的低分高分图像数量要匹配# 读取全部的内容花费时间较长valid_hr_img_list = sorted(tl.files.load_file_list(path=config.VALID.hr_img_path, regx='.*.png', printable=False))valid_lr_img_list = sorted(tl.files.load_file_list(path=config.VALID.lr_img_path, regx='.*.png', printable=False))## 如果计算机内存够大，可以加在全部内容# n_threads可以当成多线程，这里意思是每8张一组一并处理train_hr_imgs = tl.vis.read_images(train_hr_img_list, path=config.TRAIN.hr_img_path, n_threads=8)## 设置生成器、判别器和特征提取模块的输入内容# 制作生成器和判别器的输入数据t_image = tf.placeholder('float32', [batch_size, 96, 96, 3], name='t_image_input_to_SRGAN_generator')# 判别器接收的原始高分辨图像t_target_image = tf.placeholder('float32', [batch_size, 384, 384, 3], name='t_target_image')# vgg特征提取模块初始化设置t_target_image_224 = tf.image.resize_images(t_target_image, size=[224, 224], method=0, # 剪切成对应的大小align_corners=False) t_predict_image_224 = tf.image.resize_images(net_g.outputs, size=[224, 224], method=0, align_corners=False)
这样一来，我们就完成了数据的加载和小部分参数的配置。接下来我们就需要在main.py文件中继续调整生成模块、判别模块、特征提取、损失函数设置和测试模块。
模型设置
以上我们已经完成了读取文件夹内的图像内容的任务，接下来就需要用生成器和判别器分别处理各自的输入内容了。源码中生成器和判别器的具体操作是在model.py文件中执行的，main.py只是负责调用这个模块。因此我们先讲解model中的内容。首先说生成器：
生成器所需要用到的卷积和残差模块，以及对应结果加和处理都需要在这里进行设置：
import tensorflow as tfimport tensorlayer as tlfrom tensorlayer.layers import *import timeimport os# 生成网络def SRGAN_g(t_image, is_train=False, reuse=False):w_init = tf.random_normal_initializer(stddev=0.02)b_init = None # tf.constant_initializer(value=0.0)g_init = tf.random_normal_initializer(1., 0.02)with tf.variable_scope("SRGAN_g", reuse=reuse) as vs:# 输入层，内容+名字n = InputLayer(t_image, name='in')# 进行卷积（初始化）n = Conv2d(n, 64, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', W_init=w_init, name='n64s1/c')temp = n# 设置16个残差模块for i in range(16):nn = Conv2d(n, 64, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='n64s1/c1/%s' % i)nn = BatchNormLayer(nn, act=tf.nn.relu, is_train=is_train, gamma_init=g_init, name='n64s1/b1/%s' % i)nn = Conv2d(nn, 64, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='n64s1/c2/%s' % i)nn = BatchNormLayer(nn, is_train=is_train, gamma_init=g_init, name='n64s1/b2/%s' % i)nn = ElementwiseLayer([n, nn], tf.add, name='b_residual_add/%s' % i)n = nn# 残差信息整合# 对应网络示意图中的skip connection步骤n = Conv2d(n, 64, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='n64s1/c/m')n = BatchNormLayer(n, is_train=is_train, gamma_init=g_init, name='n64s1/b/m')# 把最开始的结果（temp）加到当前的结果当中n = ElementwiseLayer([n, temp], tf.add, name='add3')# 重构出图n = Conv2d(n, 256, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='n256s1/1')n = SubpixelConv2d(n, scale=2, n_out_channel=None, act=tf.nn.relu, name='pixelshufflerx2/1')n = Conv2d(n, 256, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='n256s1/2')n = SubpixelConv2d(n, scale=2, n_out_channel=None, act=tf.nn.relu, name='pixelshufflerx2/2')n = Conv2d(n, 3, (1, 1), (1, 1), act=tf.nn.tanh, padding='SAME', W_init=w_init, name='out')return n
对于判别器，也要在model中进行设置：
def SRGAN_d(input_images, is_train=True, reuse=False): # reuse指定为True意味着输入的图像是从原始数据集中取到的，# False意味着图像是生成器生成的# 参数的初始化指定w_init = tf.random_normal_initializer(stddev=0.02)b_init = None # tf.constant_initializer(value=0.0)gamma_init = tf.random_normal_initializer(1., 0.02)df_dim = 64lrelu = lambda x: tl.act.lrelu(x, 0.2)# 基础的判别网络with tf.variable_scope("SRGAN_d", reuse=reuse):tl.layers.set_name_reuse(reuse)net_in = InputLayer(input_images, name='input/images')net_h0 = Conv2d(net_in, df_dim, (4, 4), (2, 2), act=lrelu, padding='SAME', W_init=w_init, name='h0/c')net_h1 = Conv2d(net_h0, df_dim * 2, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='h1/c')net_h1 = BatchNormLayer(net_h1, act=lrelu, is_train=is_train, gamma_init=gamma_init, name='h1/bn')net_h2 = Conv2d(net_h1, df_dim * 4, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='h2/c')net_h2 = BatchNormLayer(net_h2, act=lrelu, is_train=is_train, gamma_init=gamma_init, name='h2/bn')net_h3 = Conv2d(net_h2, df_dim * 8, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='h3/c')net_h3 = BatchNormLayer(net_h3, act=lrelu, is_train=is_train, gamma_init=gamma_init, name='h3/bn')net_h4 = Conv2d(net_h3, df_dim * 16, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='h4/c')net_h4 = BatchNormLayer(net_h4, act=lrelu, is_train=is_train, gamma_init=gamma_init, name='h4/bn')net_h5 = Conv2d(net_h4, df_dim * 32, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='h5/c')net_h5 = BatchNormLayer(net_h5, act=lrelu, is_train=is_train, gamma_init=gamma_init, name='h5/bn')net_h6 = Conv2d(net_h5, df_dim * 16, (1, 1), (1, 1), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='h6/c')net_h6 = BatchNormLayer(net_h6, act=lrelu, is_train=is_train, gamma_init=gamma_init, name='h6/bn')net_h7 = Conv2d(net_h6, df_dim * 8, (1, 1), (1, 1), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='h7/c')net_h7 = BatchNormLayer(net_h7, is_train=is_train, gamma_init=gamma_init, name='h7/bn')net = Conv2d(net_h7, df_dim * 2, (1, 1), (1, 1), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='res/c')net = BatchNormLayer(net, act=lrelu, is_train=is_train, gamma_init=gamma_init, name='res/bn')net = Conv2d(net, df_dim * 2, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='res/c2')net = BatchNormLayer(net, act=lrelu, is_train=is_train, gamma_init=gamma_init, name='res/bn2')net = Conv2d(net, df_dim * 8, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, b_init=b_init, name='res/c3')net = BatchNormLayer(net, is_train=is_train, gamma_init=gamma_init, name='res/bn3')net_h8 = ElementwiseLayer([net_h7, net], combine_fn=tf.add, name='res/add')net_h8.outputs = tl.act.lrelu(net_h8.outputs, 0.2)net_ho = FlattenLayer(net_h8, name='ho/flatten') # 池化net_ho = DenseLayer(net_ho, n_units=1, act=tf.identity, W_init=w_init, name='ho/dense')logits = net_ho.outputsnet_ho.outputs = tf.nn.sigmoid(net_ho.outputs)return net_ho, logits
如果上述内容中有不懂的参数，可以查询文档。
还有，我们需要把特征提取模块（VGG）加进来，这个模块具体的作用会在损失函数里具体介绍，我们这里只需要知道vgg会帮我们提取生成图像和原始高清图像做特征比对，我们把它也写到model里：
def Vgg19_simple_api(rgb, reuse):# 减均值VGG_MEAN = [103.939, 116.779, 123.68]with tf.variable_scope("VGG19", reuse=reuse) as vs:start_time = time.time()print("build model started")rgb_scaled = rgb * 255.0# Convert RGB to BGRred, green, blue = tf.split(rgb_scaled, 3, 3)assert red.get_shape().as_list()[1:] == [224, 224, 1]assert green.get_shape().as_list()[1:] == [224, 224, 1]assert blue.get_shape().as_list()[1:] == [224, 224, 1]bgr = tf.concat([blue - VGG_MEAN[0],green - VGG_MEAN[1],red - VGG_MEAN[2],], axis=3)assert bgr.get_shape().as_list()[1:] == [224, 224, 3] #""" input layer """net_in = InputLayer(bgr, name='input')# 每卷积一轮特征图的大小缩小为原来的1/4（长宽各缩小一半）""" conv1 """network = Conv2d(net_in, n_filter=64, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv1_1')network = Conv2d(network, n_filter=64, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv1_2')network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool1')""" conv2 """network = Conv2d(network, n_filter=128, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv2_1')network = Conv2d(network, n_filter=128, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv2_2')network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool2')""" conv3 """network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv3_1')network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv3_2')network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv3_3')network = Conv2d(network, n_filter=256, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv3_4')network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool3')""" conv4 """network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv4_1')network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv4_2')network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv4_3')network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv4_4')network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool4') # (batch_size, 14, 14, 512)conv = network""" conv5 """network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv5_1')network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv5_2')network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv5_3')network = Conv2d(network, n_filter=512, filter_size=(3, 3), strides=(1, 1), act=tf.nn.relu, padding='SAME', name='conv5_4')network = MaxPool2d(network, filter_size=(2, 2), strides=(2, 2), padding='SAME', name='pool5') # (batch_size, 7, 7, 512)""" fc 6~8 """network = FlattenLayer(network, name='flatten')network = DenseLayer(network, n_units=4096, act=tf.nn.relu, name='fc6')network = DenseLayer(network, n_units=4096, act=tf.nn.relu, name='fc7')network = DenseLayer(network, n_units=1000, act=tf.identity, name='fc8')print("build model finished: %fs" % (time.time() - start_time))return network, conv
这样一来，训练所需要的模型我们就设置好了，跑代码的时候就可以直接调用了。
生成、判别、特征提取模块调用
GAN网络是一个由生成器和判别器互相耦合网络，生成器的生成并不直接受训练集的约束，因此生成器输入的内容与训练集输入不同。在上一讲中，由于任务简单，我们直接用随机噪声作为生成器的输入，但今天，为了适应较为复杂的超分辨率任务，srgan网络设计了用低分辨率图像作为生成器的初始化输入。下面我们写代码调用模块完成生成、判别和特征提取（继续补到main.py后面）：
# 数据传入网络net_g = SRGAN_g(t_image, is_train=True, reuse=False)net_d, logits_real = SRGAN_d(t_target_image, is_train=True, reuse=False) # 让网络判为真# reuse设置为真可以帮我们自动重新构建网络_, logits_fake = SRGAN_d(net_g.outputs, is_train=True, reuse=True) # 让网络判为假## 可以用以下方式查看网络的各种参数、层数等# net_g.print_params(False)# net_g.print_layers()# net_d.print_params(False)# net_d.print_layers()# reuse和之前一样，表示图像是生成结果（False）或数据集中取出的高清图片（True）net_vgg, vgg_target_emb = Vgg19_simple_api((t_target_image_224 + 1) / 2, reuse=False)_, vgg_predict_emb = Vgg19_simple_api((t_predict_image_224 + 1) / 2, reuse=True)
损失函数与训练
损失函数是可以对生成结果产生关键影响的部分，如何设计好便是关键。这里我们选择用以下三个方面作为损失函数的考量因素：
MSEloss均方误差损失，用生成图像与训练集中对应的高分辨率图像进行逐个像素点比较，计算对应的损失；VGG特征提取模块，将生成模块与原高清图像进行相同的卷积提取特征，然后将提取到的特征进行对比，计算对应损失；GANloss，这是经典的损失，用于衡量生成结果经过判别器所产生的损失。
我们先知道需要用到的这些损失即可，它们的作用会在注释中给大家介绍。下面的代码仍是main.py中的内容：
# 判别器loss，传递进数据时给数据打上标号# tf.ones_like(logits_real)生成了与logits_real等大的全1矩阵代表是真实取出的内容d_loss1 = tl.cost.sigmoid_cross_entropy(logits_real, tf.ones_like(logits_real), name='d1')# tf.ones_like(logits_fake)生成了与logits_fake等大的全0矩阵代表是生成器生成的内容d_loss2 = tl.cost.sigmoid_cross_entropy(logits_fake, tf.zeros_like(logits_fake), name='d2')d_loss = d_loss1 + d_loss2# 生成网络希望logits_fake和高分辨率原图更像# logits_fake是判别器对生成图像的特征判别结果，我们将用它与和它等大的全1矩阵进行比较，得到损失值g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(logits_fake, tf.ones_like(logits_fake), name='g')# 生成图像与真实高清图做逐帧比较mse_loss = tl.cost.mean_squared_error(net_g.outputs, t_target_image, is_mean=True)# 进行vgg特征图的比较vgg_loss = 2e-6 * tl.cost.mean_squared_error(vgg_predict_emb.outputs, vgg_target_emb.outputs, is_mean=True)g_loss = mse_loss + vgg_loss + g_gan_lossg_vars = tl.layers.get_variables_with_name('SRGAN_g', True, True)d_vars = tl.layers.get_variables_with_name('SRGAN_d', True, True)# 初始化学习率with tf.variable_scope('learning_rate'):lr_v = tf.Variable(lr_init, trainable=False)# 预训练g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(mse_loss, var_list=g_vars)## SRGANg_optim = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(g_loss, var_list=g_vars)d_optim = tf.train.AdamOptimizer(lr_v, beta1=beta1).minimize(d_loss, var_list=d_vars)##================重建模型====================sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False))tl.layers.initialize_global_variables(sess)# 查看checkpoint中是否有训练好的模型，有则加载，没有模型从0开始训练if tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_{}.npz'.format(tl.global_flag['mode']), network=net_g) is False:tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_{}_init.npz'.format(tl.global_flag['mode']), network=net_g)tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/d_{}.npz'.format(tl.global_flag['mode']), network=net_d)##===============加载特征提取模型===============vgg19_npy_path = "vgg19.npy"if not os.path.isfile(vgg19_npy_path):print("Please download vgg19.npz")exit()npz = np.load(vgg19_npy_path, encoding='latin1').item()params = []for val in sorted(npz.items()): # 提取vgg模块中的参数W = np.asarray(val[1][0])b = np.asarray(val[1][1])print(" Loading %s: %s, %s" % (val[0], W.shape, b.shape))params.extend([W, b])tl.files.assign_params(sess, params, net_vgg)print ('ok')## 开始训练## 在训练过程中使用训练集的第一个batch_size进行快速测试sample_imgs = train_hr_imgs[0:batch_size]sample_imgs_384 = tl.prepro.threading_data(sample_imgs, fn=crop_sub_imgs_fn, is_random=False)print('sample HR sub-image:', sample_imgs_384.shape, sample_imgs_384.min(), sample_imgs_384.max())sample_imgs_96 = tl.prepro.threading_data(sample_imgs_384, fn=downsample_fn)print('sample LR sub-image:', sample_imgs_96.shape, sample_imgs_96.min(), sample_imgs_96.max())tl.vis.save_images(sample_imgs_96, [ni, ni], save_dir_ginit + '/_train_sample_96.png')tl.vis.save_images(sample_imgs_384, [ni, ni], save_dir_ginit + '/_train_sample_384.png')tl.vis.save_images(sample_imgs_96, [ni, ni], save_dir_gan + '/_train_sample_96.png')tl.vis.save_images(sample_imgs_384, [ni, ni], save_dir_gan + '/_train_sample_384.png')## 固定学习率学习sess.run(tf.assign(lr_v, lr_init))print(" ** fixed learning rate: %f (for init G)" % lr_init)for epoch in range(0, n_epoch_init + 1):epoch_time = time.time()total_mse_loss, n_iter = 0, 0for idx in range(0, len(train_hr_imgs), batch_size):step_time = time.time()b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)## 更新生成器errM, _ = sess.run([mse_loss, g_optim_init], {t_image: b_imgs_96, t_target_image: b_imgs_384})print("Epoch [%2d/%2d] %4d time: %4.4fs, mse: %.8f " % (epoch, n_epoch_init, n_iter, time.time() - step_time, errM))total_mse_loss += errMn_iter += 1log = "[*] Epoch: [%2d/%2d] time: %4.4fs, mse: %.8f" % (epoch, n_epoch_init, time.time() - epoch_time, total_mse_loss / n_iter)print(log)## 快速测试if (epoch != 0) and (epoch % 10 == 0):out = sess.run(net_g_test.outputs, {t_image: sample_imgs_96}) print("[*] save images")tl.vis.save_images(out, [ni, ni], save_dir_ginit + '/train_%d.png' % epoch)## 每十个epoch保存一次模型if (epoch != 0) and (epoch % 10 == 0):tl.files.save_npz(net_g.all_params, name=checkpoint_dir + '/g_{}_init.npz'.format(tl.global_flag['mode']), sess=sess)###========================= train GAN (SRGAN) =========================###for epoch in range(0, n_epoch + 1):## 更新学习率if epoch != 0 and (epoch % decay_every == 0):new_lr_decay = lr_decay**(epoch // decay_every)sess.run(tf.assign(lr_v, lr_init * new_lr_decay))log = " ** new learning rate: %f (for GAN)" % (lr_init * new_lr_decay)print(log)elif epoch == 0:sess.run(tf.assign(lr_v, lr_init))log = " ** init lr: %f decay_every_init: %d, lr_decay: %f (for GAN)" % (lr_init, decay_every, lr_decay)print(log)epoch_time = time.time()total_d_loss, total_g_loss, n_iter = 0, 0, 0for idx in range(0, len(train_hr_imgs), batch_size):step_time = time.time()b_imgs_384 = tl.prepro.threading_data(train_hr_imgs[idx:idx + batch_size], fn=crop_sub_imgs_fn, is_random=True)b_imgs_96 = tl.prepro.threading_data(b_imgs_384, fn=downsample_fn)## 更新判别器errD, _ = sess.run([d_loss, d_optim], {t_image: b_imgs_96, t_target_image: b_imgs_384})## 更新生成器errG, errM, errV, errA, _ = sess.run([g_loss, mse_loss, vgg_loss, g_gan_loss, g_optim], {t_image: b_imgs_96, t_target_image: b_imgs_384})print("Epoch [%2d/%2d] %4d time: %4.4fs, d_loss: %.8f g_loss: %.8f (mse: %.6f vgg: %.6f adv: %.6f)" %(epoch, n_epoch, n_iter, time.time() - step_time, errD, errG, errM, errV, errA))total_d_loss += errDtotal_g_loss += errGn_iter += 1log = "[*] Epoch: [%2d/%2d] time: %4.4fs, d_loss: %.8f g_loss: %.8f" % (epoch, n_epoch, time.time() - epoch_time, total_d_loss / n_iter,total_g_loss / n_iter)print(log)## quick evaluation on train setif (epoch != 0) and (epoch % 10 == 0):out = sess.run(net_g_test.outputs, {t_image: sample_imgs_96}) #; print('gen sub-image:', out.shape, out.min(), out.max())print("[*] save images")tl.vis.save_images(out, [ni, ni], save_dir_gan + '/train_%d.png' % epoch)## 保存模型if (epoch != 0) and (epoch % 10 == 0):tl.files.save_npz(net_g.all_params, name=checkpoint_dir + '/g_{}.npz'.format(tl.global_flag['mode']), sess=sess)tl.files.save_npz(net_d.all_params, name=checkpoint_dir + '/d_{}.npz'.format(tl.global_flag['mode']), sess=sess)def evaluate():## create folders to save result imagessave_dir = "samples/{}".format(tl.global_flag['mode'])tl.files.exists_or_mkdir(save_dir)checkpoint_dir = "checkpoint"valid_hr_img_list = sorted(tl.files.load_file_list(path=config.VALID.hr_img_path, regx='.*.png', printable=False))valid_lr_img_list = sorted(tl.files.load_file_list(path=config.VALID.lr_img_path, regx='.*.png', printable=False))valid_lr_imgs = tl.vis.read_images(valid_lr_img_list, path=config.VALID.lr_img_path, n_threads=8)valid_hr_imgs = tl.vis.read_images(valid_hr_img_list, path=config.VALID.hr_img_path, n_threads=8)
测试
测试部分主要是检查生成器跑出的模型（保存在checkpoint里的g_srgan.npz文件）是否实用，而判别器的学习结果我们并没有进行保存。想要进行测试，就需要传入低分辨率图像，让图像经过模型生成高分辨率图像，在与真实的高分辨率图像进行对比。先来看看代码，这部分代码依然是main.py的后续部分：
imid = 9 # 输入待超分的低分辨率图像序号valid_lr_img = valid_lr_imgs[imid]valid_hr_img = valid_hr_imgs[imid]valid_lr_img = (valid_lr_img / 127.5) - 1size = valid_lr_img.shapet_image = tf.placeholder('float32', [1, None, None, 3], name='input_image')net_g = SRGAN_g(t_image, is_train=False, reuse=False)###========================== RESTORE G =============================###sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False))tl.layers.initialize_global_variables(sess)tl.files.load_and_assign_npz(sess=sess, name=checkpoint_dir + '/g_srgan.npz', network=net_g)###======================= EVALUATION =============================###start_time = time.time()out = sess.run(net_g.outputs, {t_image: [valid_lr_img]})print("took: %4.4fs" % (time.time() - start_time))print("LR size: %s / generated HR size: %s" % (size, out.shape))print("[*] save images")tl.vis.save_image(out[0], save_dir + '/valid_gen.png')tl.vis.save_image(valid_lr_img, save_dir + '/valid_lr.png')tl.vis.save_image(valid_hr_img, save_dir + '/valid_hr.png')out_bicu = scipy.misc.imresize(valid_lr_img, [size[0] * 4, size[1] * 4], interp='bicubic', mode=None)tl.vis.save_image(out_bicu, save_dir + '/valid_bicubic.png')if __name__ == '__main__':import argparseparser = argparse.ArgumentParser()parser.add_argument('--mode', type=str, default='evaluate', help='srgan, evaluate')# 这个参数可以选择srgan或evaluateargs = parser.parse_args()tl.global_flag['mode'] = args.modeif tl.global_flag['mode'] == 'srgan':# 先训练后输出train()evaluate()elif tl.global_flag['mode'] == 'evaluate':evaluate()else:raise Exception("Unknow --mode")
这样我们就可以借助训练出的模型对图像进行超分了，运行的结果会给我们保存到evaluate文件夹中，我们可以来对比一下：
从左至右依次为原低分辨率图像、超分图像、原高分辨率图像。如果大家觉得结果对比不是很明显的话，可以自己运行一下试试看啦~
下面给大家贴一些训练时的部分pycharm终端内容：
（这后面还有一个2000epoch的训练我忘记截了，真的很难难跑…都看到这里了确定不点个赞嘛）
今天的内容就到这里了，代码很多，但是我们需要修改的地方并不太多，希望对大家有所帮助。
