图像超分辨重构的原理，输入一张像素点少，像素较低的图像， 输出一张像素点多，像素较高的图像
而在作者的文章中，作者使用downsample_up, 使用imresize(img, []) 将图像的像素从原理的384，384降低到96， 96, 从而构造出高水平的图像和低水平的图像
作者使用了三个部分构成网络，
第一部分是生成网络，用于进行图片的生成，使用了16层的残差网络，最后的输出结果为tf.nn.tanh(),即为-1, 1, 因为图像进行了-1,1的预处理
第二部分是判别网络， 用于进行图片的判别操作，对于判别网络而言，是希望将生成的图片判别为假，将真的图片判别为真
第三部分是VGG19来提取生成图片和真实图片的conv5层卷积层的输出结果，用于生成局部部位的损失值mse
损失值说明：
d_loss:
d_loss_1: tl.cost.sigmoid_cross_entropy(logits_real, tf.ones_like(logits_real)) # 真实图像的判别结果的损失值
d_loss_2: tl.cost.sigmoid_cross_entrpopy(logits_fake, tf.zeros_like(logits_real)) # 生成图像的判别结果的损失值
g_loss:
g_gan_loss: 1e-3 * tl.cost.sigmoid_cross_entropy(logits_fake, tf.ones_like(logits_real)) # 损失值表示为 -log(D(g(lr))) # 即生成的图像被判别为真的损失值
mse_loss: tl.cost.mean_squared_error(net_g.outputs, t_target_image) # 计算真实值与生成值之间的像素差
vgg_loss: tl.cost.mean_squared_error(vgg_predict_emb.outputs, vgg_target_emb.outputs) # 用于计算生成图片和真实图片经过vgg19的卷积层后，特征图之间的差异，用来获得特征细节的差异性
训练说明：
首先进行100次迭代，用来优化生成网络，使用tf.train.AdamOptimer(lr_v, beta1=beta1).minimize(mse_loss, var_list=g_var)
等生成网络迭代好以后，开始迭代生成网络和判别网络，以及VGG19的损失值缩小
生成网络：使用了16个残差模块，在残差模块的输入与下一层的输出之间又进行一次残差直连
判别网络：使用的是feature_map递增的卷积层构造成的判别网路
代码说明：
第一步：将参数从config中导入到main.py
第二步：使用tl.file.exists_or_mkdir() 构造用于储存图片的文件夹，同时定义checkpoint的文件夹
第三步：使用sorted(tl.files.load_file_list) 生成图片的列表， 使用tl.vis.read_images() 进行图片的读入
第四步：构建模型的构架Model
第一步：定义输入参数t_image = tf.placeholder('float32', [batch_size, 96, 96, 3]), t_target_image = tf.placeholder('float32', [batch_size, 384, 384, 3])
第二步： 使用SGRAN_g 用来生成最终的生成网络，net_g, 输入参数为t_image, is_training, reuse
第三步： 使用SGRAN_d 用来生成判别网络，输出结果为net_d网络架构，logits_real, 输入参数为t_target_image, is_training, reuse, 同理输入t_image, 获得logits_fake
第四步： 使用net_g.print_params(False) 和 net_g.print_layers() 不打印参数，打印每一层
第五步：将net_g.outputs即生成的结果和t_target_image即目标图像的结果输入到Vgg_19_simple_api, 获得vgg_net, 以及conv第五层的输出结果
第一步：tf.image.resize_images()进行图片的维度变换，为了可以使得其能输入到VGG_19中
第二步：将变化了维度的t_target_image 输入到Vgg_19_simple_api, 获得net_vgg, 和 vgg_target_emb即第五层卷积的输出结果
第三步：将变化了维度的net_g.outputs 输入到Vgg_19_simple_api, 获得 vgg_pred_emb即第五层卷积的输出结果
第六步： 构造net_g_test = SGRAN_g(t_image, False, True) 用于进行训练中的测试图片
第五步：构造模型loss，还有trian_ops操作
第一步： loss的构造， d_loss 和 g_loss的构造
第一步： d_loss的构造， d_loss_1 + d_loss_2
第一步： d_loss_1: 构造真实图片的判别损失值，即tl.cost.softmax_cross_entropy(logits_real, tf.ones_like(logits_real))
第二步： d_loss_2: 构造生成图片的判别损失值， 即tl.cost.softmax_cross_entropy(logits_fake, tf.ones_like(logits_fake))
第二步： g_loss的构造，g_gan_loss, mse_loss, vgg_loss
第一步： g_gan_loss, 生成网络被判别网络判别为真的概率，使用tl.cost.softmax_cross_entropy(logits_fake, tf.ones_like(logits_fake))
第二步：mse_loss 生成图像与目标图像之间的像素点差值，使用tl.cost.mean_squared_error(t_target_image, net_g.outputs)
第三步：vgg_loss 将vgg_target_emb.outputs与vgg_pred_emb.outputs获得第五层卷积层输出的mse_loss
第二步：构造train_op，包括 g_optim_init用预训练, 构造g_optim, d_optim
第一步：g_var = tl.layers.get_variables_with_name(‘SGRAN_g') 生成网络的参数获得
第二步： d_var = tl.layers.get_variable_with_name('SGRAN_d') 判别网络的参数获得
第三步： 使用with tf.variable_scope('learning_rate'): 使用lr_v = tf.Variable(lr_init)
第四步：定义train_op, g_optim_init, g_optim, d_optim
第一步：构造g_optim_init 使用tf.train.Adaoptimer(lr_v, beta1=betal).minimize(mse_loss, var_list=g_var)
第二步：构造g_optim 使用tf.train.Adaoptimer(lr_v, beta1=betal).minimize(g_loss, var_list=g_var)
第三步：构造d_optim 使用tf.train.Adaoptimer(lr_v, beta1=betal).minimize(d_loss, var_list=d_var)
第六步：使用tl.files.load_and_assign_npz() 载入训练好的sess参数
第一步： 使用tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False))
第二步： tl.layers.initialize_global_variables(sess)
第三步： 使用tf.file.load_and_assign_npz 进行g_net的参数下载， 否者就下载g_{}_init的参数下载
第四步：使用tf.file.load_and_assgin_npz进行d_net的参数下载
第七步：下载VGG网络，将其运用到net_vgg
第一步：使用np.load(path, encoding='latin1').item() 下载参数
第二步：循环sorted(npz.items()) 进行参数循环，将其添加到params
第三步：使用tl.files.assign_params(sess, params, net_vgg) 将参数运用到net_vgg
第八步：进行参数的训练操作
第一步：从图片中跳出一个batch_size的数据构成测试集
第一步： 使用tl.prepro.threading_data fn = crop_sub_imgs_fn， 使用crop进行裁剪操作
第二步： 使用tl.prepro.threading_data fn = downsample 使用imresize进行图片的维度压缩
第二步：进行预训练操作
第一步：循环迭代， 获得一个batch的数据，使用crop_sub_imgs_fn 和 downsample构造低水平的数据和高水平的数据
第二步：使用sess.run， g_optim_init进行图片的预训练
第三步：进行训练操作
第一步：循环迭代，获得一个batch的数据，使用crop_sub_imgs_fn 和 downsample构造低水平的数据和高水平的数据
第二步：使用sess.run, g_optim 和 d_optim 进行图片的训练操作
第九步：进行evaluate图片的测试阶段
第一步： 构造图片展示的文件夹，使用tf.files.exits_files_mkdir
第二步: 使用tl.files.load_file_list 和 tl.vis.read_images读入图片
第三步：根据索引选择一张图片，/127.5 - 1 进行归一化处理
第四步：使用tf.placeholder('float32', [1, None, None, 3]) 构造输入的t_image
第五步： 使用SGRAN_g(t_image, False, False) 构造net_g
第六步：使用tf.Session() 构造sess，使用tl.files.load_and_assign_npz下载训练好的sess， network=net_g
第七步：使用sess.run([net_g.outputs], feed_dict={t_image:[valid_lr_img]}) 获得图片
第八步：使用tl.vis.save_images(outputs[0])保存图片
第九步：使用scipy.misc.imresize() 将低像素的图片扩大为原来的四倍，与重构的图像作对比
代码： main.py 主函数
import tensorlayer as tl import tensorflow as tf import numpy as np from config import config from model import * import os import time import scipy ## 添加参数 batch_size = config.TRAIN.batch_size lr_init = config.TRAIN.lr_init betal = config.TRAIN.betal ### initialze G n_epoch_init = config.TRAIN.n_epoch_init ### adversarial learning n_epoch = config.TRAIN.n_epoch lr_decay = config.TRAIN.lr_decay decay_every = config.TRAIN.decay_every ni = int(np.sqrt(batch_size)) def train(): # 创建用于进行图片储存的文件 save_dir_ginit = 'sample/{}_ginit'.format(tl.global_flag['mode']) save_dir_gan = 'sample/{}_gan'.format(tl.global_flag['mode']) tl.files.exists_or_mkdir(save_dir_ginit) tl.files.exists_or_mkdir(save_dir_gan) checkpoint = 'checkpoint' tl.files.exists_or_mkdir(checkpoint) train_hr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.hr_img_path, regx='.*.png', printable=False)) train_lr_img_list = sorted(tl.files.load_file_list(path=config.TRAIN.lr_img_path, regx='.*.png', printable=False)) train_hr_img = tl.vis.read_images(train_hr_img_list, path=config.TRAIN.hr_img_path, n_threads=8) train_lr_img = tl.vis.read_images(train_lr_img_list, path=config.TRAIN.lr_img_path, n_threads=8) # 构造输入 t_image = tf.placeholder('float32', [batch_size, 96, 96, 3]) t_target_image = tf.placeholder('float32', [batch_size, 384, 384, 3]) # 构造生成的model，获得生成model的输出net_g net_g = SRGAN_g(t_image, True, False) # 构造判别网络，判别net_g.output, t_target_image, net_d表示整个网络 net_d, logist_real = SRGAN_d(t_target_image, True, False) _, logist_fake = SRGAN_d(net_g.outputs, True, True) # 构造VGG网络 net_g.print_params(False) net_g.print_layers() net_d.print_params(False) net_d.print_layers() # 进行输入数据的维度变换，将其转换为224和224 target_image_224 = tf.image.resize_images(t_target_image, [224, 224], method=0, align_corners=False) pred_image_224 = tf.image.resize_images(net_g.outputs, [224, 224], method=0, align_corners=False) net_vgg, vgg_target_emb = Vgg_19_simple_api((target_image_224 + 1) / 2, reuse=False) _, vgg_pred_emb = Vgg_19_simple_api((net_g + 1) / 2, reuse=True) # 进行训练阶段的测试 net_g_test = SRGAN_g(t_image, False, True) #### ========== DEFINE_TRAIN_OP =================### d_loss_1 = tl.cost.sigmoid_cross_entropy(logist_real, tf.ones_like(logist_real)) d_loss_2 = tl.cost.sigmoid_cross_entropy(logist_fake, tf.zeros_like(logist_fake)) d_loss = d_loss_1 + d_loss_2 g_gan_loss = 1e-3 * tl.cost.sigmoid_cross_entropy(logist_fake, tf.ones_like(logist_fake)) mse_loss = tl.cost.mean_squared_error(net_g.outputs, t_target_image, is_mean=True) vgg_loss = 2e-6 * tl.cost.mean_squared_error(vgg_target_emb.outputs, vgg_pred_emb.outputs, is_mean=True) g_loss = g_gan_loss + mse_loss + vgg_loss g_var = tl.layers.get_variables_with_name('SRGAN_g', True, True) d_var = tl.layers.get_variables_with_name('SRGAN_d', True, True) with tf.variable_scope('learning_rate'): lr_v = tf.Variable(lr_init, trainable=False) g_optim_init = tf.train.AdamOptimizer(lr_v, beta1=betal).minimize(mse_loss, var_list=g_var) g_optim = tf.train.AdamOptimizer(lr_v, beta1=betal).minimize(g_loss, var_list=g_var) d_optim = tf.train.AdamOptimizer(lr_v, beta1=betal).minimize(d_loss, var_list=d_var) ###======================RESTORE_MODEL_SESS ==================### sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False)) tl.layers.initialize_global_variables(sess) if tl.files.load_and_assign_npz(sess, checkpoint + '/g_{}.npz'.format(tl.global_flag['mode'], network=net_g)) is False: tl.files.load_and_assign_npz(sess, checkpoint + '/g_init_{}.npz'.format(tl.global_flag['mode'], network=net_g)) tl.files.load_and_assign_npz(sess, checkpoint + '/d_{}.npz'.format(tl.global_flag['mode'], network=net_d)) ### ================== load vgg params =================== ### vgg_npy_path = 'vgg19.npy' if not os.path.isfile(vgg_npy_path): print('Please download vgg19.npz from : https://github.com/machrisaa/tensorflow-vgg') exit() npz = np.load(vgg_npy_path, encoding='latin1').item() params = [] for var in sorted(npz.items()): W = np.asarray(var[1][0]) b = np.asarray(var[1][1]) params.extend([W, b]) tl.files.assign_params(sess, params, net_vgg) print('ok') ###======================== TRAIN =======================### sample_imgs = train_hr_img[0:batch_size] # 进行随机裁剪，保证其维度为384 sample_imgs_384 = tl.prepro.threading_data(sample_imgs, fn=crop_sub_imgs_fn, is_random=False) # 进行像素的降低 sample_imgs_96 = tl.prepro.threading_data(sample_imgs_384, fn=downsample_fn) # 进行图片的保存 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') ###======================== initial train G =====================### for epoch in range(n_epoch_init): n_iter = 0 init_loss_total = 0 for idx in range(0, len(train_hr_img), batch_size): b_img_384 = tl.prepro.threading_data(train_hr_img[idx:idx+batch_size], fn=crop_sub_imgs_fn, is_random=False) b_img_96 = tl.prepro.threading_data(b_img_384, fn=downsample_fn) _, MSE_LOSS = sess.run([g_optim_init, mse_loss], feed_dict={t_image:b_img_96, t_target_image:b_img_384}) init_loss_total += MSE_LOSS if (epoch != 0) and (epoch % 10 == 0): out = sess.run(net_g_test.outputs, feed_dict={t_image:sample_imgs_96}) print('[*] save image') tl.vis.save_images(out, [ni, ni], save_dir_ginit + '/train_%d.png' % epoch) if (epoch != 0) and (epoch % 10 ==0): tl.files.save_npz(net_g.all_params, name=checkpoint + '/g_init_{}.npz'.format(tl.global_flag['mode'])) ### ======================== train GAN ================== ### for epoch in range(0, n_epoch+1): if epoch != 0 and epoch % decay_every == 0: new_lr = lr_decay ** (epoch // decay_every) sess.run(tf.assign(lr_v, new_lr * lr_v)) log = '** new learning rate: %f(for GAN)' % (lr_init * new_lr) 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, 0 for idx in range(0, len(train_hr_img), batch_size): b_img_384 = tl.prepro.threading_data(train_hr_img[idx:idx+batch_size], fn=crop_sub_imgs_fn, is_random=False) b_img_96 = tl.prepro.threading_data(b_img_384, fn=downsample_fn) _, errD = sess.run([d_optim, d_loss], feed_dict={t_image:b_img_96, t_target_image:b_img_384}) _, errG, errM, errV, errA = sess.run([g_optim, g_loss, mse_loss, vgg_loss, g_gan_loss], feed_dict={t_image:b_img_96, t_target_image:b_img_384}) total_d_loss += errD total_g_loss += errG if epoch != 0 and epoch % 10 == 0: out = sess.run(net_g_test.outputs, feed_dict={t_image:sample_imgs_96}) print('[*] save image') tl.vis.save_images(out, [ni, ni], save_dir_gan + '/train_%d' % epoch) if epoch != 0 and epoch % 10 == 0: tl.files.save_npz(net_g.all_params, name = checkpoint + '/g_{}.npz'.format(tl.global_flag['mode'])) tl.files.save_npz(net_d.all_params, name= checkpoint + '/d_{}.npz'.format(tl.global_flag['mode'])) def evaluate(): save_dir = 'sample/{}'.format(tl.global_flag['mode']) tl.files.exists_or_mkdir(save_dir) checkpoints = 'checkpoints' evaluate_hr_img_list = sorted(tl.files.load_file_list(config.VALID.hr_img_path, regx='.*.png', printable=False)) evaluate_lr_img_list = sorted(tl.files.load_file_list(config.VALID.lr_img_path, regx='.*.png', printable=False)) valid_lr_imgs = tl.vis.read_images(evaluate_lr_img_list, path=config.VALID.lr_img_path, n_threads=8) valid_hr_imgs = tl.vis.read_images(evaluate_hr_img_list, path=config.VALID.hr_img_path, n_threads=8) ### ==================== DEFINE MODEL =================### imid = 64 valid_lr_img = valid_lr_imgs[imid] valid_hr_img = valid_hr_imgs[imid] valid_lr_img = (valid_lr_img / 127.5) - 1 t_image = tf.placeholder('float32', [1, None, None, 3]) net_g = SGRAN_g(t_image, False, False) sess = tf.Session() tl.files.load_and_assign_npz(sess, checkpoints + '/g_{}.npz'.format(tl.global_flag['mode']), network=net_g) output = sess.run([net_g.outputs], feed_dict={t_image:[valid_lr_img]}) tl.vis.save_images(output[0], [ni, ni], save_dir + '/valid_gen.png') tl.vis.save_images(valid_lr_img, [ni, ni], save_dir + '/valid_lr.png') tl.vis.save_images(valid_hr_img, [ni, ni], save_dir + '/valid_hr.png') size = valid_hr_img.shape out_bicu = scipy.misc.imresize(valid_lr_img, [size[0]*4, size[1]*4], interp='bicubic', mode=None) tl.vis.save_images(out_bicu, [ni, ni], save_dir + '/valid_out_bicu.png') if __name__ == '__main__': import argparse parse = argparse.ArgumentParser() parse.add_argument('--mode', type=str, default='srgan', help='srgan evaluate') args = parse.parse_args() tl.global_flag['mode'] = args.mode if tl.global_flag['mode'] == 'srgan': train() elif tl.global_flag['mode'] == 'evaluate': evaluate()
model.py 构建模型
import tensorflow as tf import tensorlayer as tl from tensorlayer.layers import * import time def SRGAN_g(input_image, is_train, reuse): w_init = tf.random_normal_initializer(stddev=0.2) b_init = None g_init = tf.random_normal_initializer(1, 0.02) with tf.variable_scope('SRGAN_g', reuse=reuse): n = InputLayer(input_image, name='in') n = Conv2d(n, 64, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', W_init=w_init, name='n64s1/c') temp = n for i in range(16): nn = Conv2d(n, 64, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='n64s1/c1/%d' % i) nn = BatchNormLayer(nn, act=tf.nn.relu, is_train=is_train, gamma_init=g_init, name='n64s1/b1/%d' % i) nn = Conv2d(nn, 64, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='n64s1/c2/%d' % i) nn = BatchNormLayer(nn, act=None, is_train=is_train, gamma_init=g_init, name='n64s1/b2/%d'%i) nn = ElementwiseLayer([n, nn], tf.add, name='b_residual_add_%d' % i) n = nn n = Conv2d(n, 64, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='n64s1/c3') n = BatchNormLayer(n, act=None, is_train=is_train, gamma_init=g_init, name='n64s1/b3') n = ElementwiseLayer([temp, n], tf.add, name='add3') # 进行反卷积操作 n = Conv2d(n, 256, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='n64s1/c4') n = SubpixelConv2d(n, scale=2, n_out_channel=None, act=tf.nn.relu, name='pixelshuffler2/1') n = Conv2d(n, 256, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='n64s1/c5') n = SubpixelConv2d(n, scale=2, n_out_channel=None, act=tf.nn.relu, name='pixelshuffle2/2') n = Conv2d(n, 3, (1, 1), (1, 1), act=tf.nn.tanh, padding='SAME', W_init=w_init, name='out') return n def SRGAN_d(input_image, is_training=True, reuse=False): w_init = tf.random_normal_initializer(stddev=0.2) b_init = None g_init = tf.random_normal_initializer(1.0, stddev=0.02) lrelu = lambda x: tl.act.lrelu(x, 0.2) df_dim = 64 with tf.variable_scope('SRGAN_d', reuse=reuse): tl.layers.set_name_reuse(reuse) net_in = InputLayer(input_image, name='input/image') 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, name='h1/c') net_h1 = BatchNormLayer(net_h1, act=lrelu, is_train=is_training, gamma_init=g_init, name='h1/bn') net_h2 = Conv2d(net_h1, df_dim*4, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, name='h2/c') net_h2 = BatchNormLayer(net_h2, act=lrelu, is_train=is_training, gamma_init=g_init, name='h2/bn') net_h3 = Conv2d(net_h2, df_dim*8, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, name='h3/c') net_h3 = BatchNormLayer(net_h3, act=lrelu, is_train=is_training, gamma_init=g_init, name='h3/bn') net_h4 = Conv2d(net_h3, df_dim*16, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, name='h4/c') net_h4 = BatchNormLayer(net_h4, act=lrelu, is_train=is_training, gamma_init=g_init, name='h4/bn') net_h5 = Conv2d(net_h4, df_dim*32, (4, 4), (2, 2), act=None, padding='SAME', W_init=w_init, name='h5/c') net_h5 = BatchNormLayer(net_h5, act=lrelu, is_train=is_training, gamma_init=g_init, name='h5/bn') net_h6 = Conv2d(net_h5, df_dim*16, (1, 1), (1, 1), act=None, padding='SAME', W_init=w_init, name='h6/c') net_h6 = BatchNormLayer(net_h6, act=lrelu, is_train=is_training, gamma_init=g_init, name='h6/bn') net_h7 = Conv2d(net_h6, df_dim*8, (1, 1), (1, 1), act=None, padding='SAME', W_init=w_init, name='h7/c') net_h7 = BatchNormLayer(net_h7, act=lrelu, is_train=is_training, gamma_init=g_init, name='h7/bn') net = Conv2d(net_h7, df_dim*2, (1, 1), (1, 1), act=None, padding='SAME', W_init=w_init, name='reg/c') net = BatchNormLayer(net, act=lrelu, is_train=is_training, gamma_init=g_init, name='reg/bn') net = Conv2d(net, df_dim*2, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='reg/c2') net = BatchNormLayer(net, act=lrelu, is_train=is_training, gamma_init=g_init, name='reg/bn2') net = Conv2d(net, df_dim*8, (3, 3), (1, 1), act=None, padding='SAME', W_init=w_init, name='reg/c3') net = BatchNormLayer(net, act=lrelu, is_train=is_training, gamma_init=g_init, name='reg/bn3') net_h8 = ElementwiseLayer([net_h7, net], tf.add, name='red/add') net_h8.outputs = tl.act.lrelu(net_h8.outputs, 0.2) net_ho = FlattenLayer(net_h8, name='ho/flatten') net_ho = DenseLayer(net_ho, n_units=1, act=tf.identity, W_init=w_init, name='ho/dense') logits = net_ho.outputs net_ho.outputs = tf.nn.sigmoid(net_ho.outputs) return net_ho, logits def Vgg_19_simple_api(input_image, reuse): VGG_MEAN = [103.939, 116.779, 123.68] # 将输入的rgb图像转换为bgr with tf.variable_scope('VGG19', reuse=reuse) as vs: start_time = time.time() print('build the model') input_image = input_image * 255 red, green, blue = tf.split(input_image, 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 input_image.get_shape().as_list()[1:] == [224, 224, 3] net_in = InputLayer(bgr, name='input') # 构建网络 """conv1""" network = Conv2d(net_in, 64, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv1_1') network = Conv2d(network, 64, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv1_2') network = MaxPool2d(network, (2, 2), (2, 2), padding='SAME', name='pool1') '''conv2''' network = Conv2d(network, 128, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv2_1') network = Conv2d(network, 128, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv2_2') network = MaxPool2d(network, (2, 2), (2, 2), padding='SAME', name='pool2') '''conv3''' network = Conv2d(network, 256, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv3_1') network = Conv2d(network, 256, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv3_2') network = Conv2d(network, 256, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv3_3') network = Conv2d(network, 256, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv3_4') network = MaxPool2d(network, (2, 2), (2, 2), padding='SAME', name='pool3') '''conv4''' network = Conv2d(network, 512, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv4_1') network = Conv2d(network, 512, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv4_2') network = Conv2d(network, 512, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv4_3') network = Conv2d(network, 512, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv4_4') network = MaxPool2d(network, (2, 2), (2, 2), padding='SAME', name='pool4') '''conv5''' network = Conv2d(network, 512, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv5_1') network = Conv2d(network, 512, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv5_2') network = Conv2d(network, 512, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv5_3') network = Conv2d(network, 512, (3, 3), (1, 1), act=tf.nn.relu, padding='SAME', name='conv5_4') network = MaxPool2d(network, (2, 2), (2, 2), padding='SAME', name='pool5') conv = network """fc6-8""" 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('finish the bulid %fs' % (time.time() - start_time)) return network, conv
config.py 参数文件
from easydict import EasyDict as edict import json config = edict() config.TRAIN = edict() # Adam config.TRAIN.batch_size = 1 config.TRAIN.lr_init = 1e-4 config.TRAIN.betal = 0.9 ### initialize G config.TRAIN.n_epoch_init = 100 ### adversarial_leaning config.TRAIN.n_epoch = 2000 config.TRAIN.lr_decay = 0.1 config.TRAIN.decay_every = int(config.TRAIN.n_epoch / 2) ## train set location config.TRAIN.hr_img_path = r'C:\Users\qq302\Desktop\srdata\DIV2K_train_HR' config.TRAIN.lr_img_path = r'C:\Users\qq302\Desktop\srdata\DIV2K_train_LR_bicubic\X4' # valid set location config.VALID = edict() config.VALID.hr_img_path = r'C:\Users\qq302\Desktop\srdata\DIV2K_valid_HR' config.VALID.lr_img_path = r'C:\Users\qq302\Desktop\srdata\DIV2K_valid_LR_bicubic/X4'
utils.py 操作文件
from tensorlayer.prepro import * def crop_sub_imgs_fn(img, is_random=True): x = crop(img, wrg=384, hrg=384, is_random=is_random) # 进行 -1 - 1 的归一化 x = x / 127.5 - 1 return x def downsample_fn(img): x = imresize(img, [96, 96], interp='bicubic', mode=None) # 存在一定的问题 x = x / 127.5 - 1 return x
