一、背景
MARTA-GAN全称为multiple-layer feature-matching generative adversarial networks，是Daoyu Lin等人于16年12月发表的文章。这篇文章算是我看到的GAN最早应用在遥感领域的文章了，所以就打算来实现一下。作者这篇文章的题目为《MARTA GANs: Unsupervised Representation Learning for Remote Sensing Image Classification》，不过我感觉与其说是图像分类，不如说是场景实现更准确一些。
MARTA-GAN是基于DCGAN做的，我实验的结果非常不理想，所以这篇文章重点就放在对MARTA-GAN的解读了。不过实验过程还是要写的，可以为后面做这个模型的人提供参考。
[1]文章链接：https://arxiv.org/pdf/1612.08879.pdf
二、MARTA-GAN解读
搜了一下，网上几乎没有对这篇论文详细解读的文章，那我就根据自己的理解来写写。
先从摘要部分看MARTA-GAN的目的，作者提到现有模型的最大限制是非常有限的标记样本，因此作者提出了一个非监督模型MARTA-GAN，利用无标记样本来实现不同场景。
遥感图像与普通的数字图像还是有区别的，遥感图像具有非常强的空间性，不同地物会表现出不同的大小、颜色。早先用于场景分类的方法包括BoVW（bag of visual words）、SPM（spatial pyramid matching）、CNN等，但这些方法都是基于手工选取的特征或需要大量标记样本。而GAN是一种非监督的方法，因此作者基于DCGAN做了改进，提出MARTA-GAN，作者的改进之处主要表现在：
（1）在生成器中，DCGAN生成64*64的图，但是MARTA-GAN生成256*256的图
（2）核的大小不同，DCGAN为5*5，MARTA-GAN为4*4
（3）MARTA-GAN使用多特征层（multi-feature layer）来整合中级、高级信息（mid- and high-level）
（4）引入两种loss函数：感知loss（perceptual loss）和特征匹配loss（feature matching loss）
作者的主要贡献在于：
1）To our knowledge, this is the first time that GANs have been applied to classify unsupervised remote sensing images. （首次将GAN用于遥感图像）
2）The results of experiments on the UC-Merced Landuse and Brazilian Coffee Scenes datasets showed that the proposed algorithm outperforms state-of-the-art unsupervised algorithms in terms of overall classification accuracy. （效果很好）
3）We propose a multi-feature layer by combining perceptual loss and loss of feature matching to learn better image representations.（结合了感知loss和特征匹配loss的多特征层）
先来看一下作者提出MARTA-GAN的流程图：
流程图还是非常小清新的。模型的驱动函数这里不再过多介绍。关于网络结构，示意图为：
除过上图中标记的网络参数，作者还提到，激活函数使用了leakyReLU，leak值设为0.2，kernel sizes为4*4，步长为2，在判别器和生成器中使用了BN（batch normalization）层，衰减因子设置为0.9。正式进行实验之前，还需要对数据做预处理，将图像的值归一化至[-1 1]，batch size设置为64，使用SGD（随机梯度下降），优化器为Adam，学习率为0.00002，动量为0.5。作者的实现过程是用的tensorlayer。
作者将MARTA-GAN用于两个不同的数据集： the UC Merced Land Use Dataset和the Brazilian Coffee Scenes Dataset，并都取得了不错的表现。我们的实验使用的是the UC Merced Land Use Dataset数据集，后面会详细介绍；关于the Brazilian Coffee Scenes Dataset数据集，一共有2876张影像，均来自SPOT卫星，图像尺寸为64*64，包含1438张Coffee影像和1438张non-Coffee影像。
先来看看在the UC Merced Land Use Dataset数据集上的表现：
再来看看MARTA-GAN在the Brazilian Coffee Scenes Dataset数据集上的表现
虽然the Brazilian Coffee Scenes Dataset的结果比较抽象，不过原始数据集似乎效果也不是很好，所以作者还针对the UC Merced Land Use Dataset数据集的结果做了精度评定，并与DCGAN做了对比：
该数据集一共是21个类别，MARTA-GAN的平均精度约比DCGAN高7%，作者提到这是因为MARTA-GAN中引入了中级和高级特征信息。
最后作者还比对了不同方法用于这两个数据集的效果，可以直观的看到MARTA-GAN的参数量最少且效果较好;
三、MARTA-GAN实现过程
关于MARTA-GAN的实现过程，我主要参考了两个代码：
[2]https://github.com/BUPTLdy/MARTA-GAN
[3]https://github.com/ualiawan/MARTA-GANs
由于[2]是用tensorlayer写的，我的tensorlayer有点问题，而[3]是用tensorflow写的，但是有点小bug并少了一些功能，所以我主要使用[3]的代码，并做了一些改进。
1. 所有文件结构
所有文件结构为：
-- utilities.py
-- network.py
-- train_net.py
-- features.py
-- train_svm.py
-- data # 数据集需要自己准备
|------ uc_test_256
|------ image01.jpg
|------ image02.jpg
|------ ......
|------ uc_train_256_data
|------ image01.jpg
|------ image02.jpg
|------ ......
|------ uc_train_256_feat
|------ image01.jpg
|------ image02.jpg
|------ ......2. 数据准备
我们所使用的数据集为the UC Merced Land Use Dataset，论文中作者也对该数据集进行了介绍，下面引入论文中的介绍：
这个数据集有21种土地利用类型（21 land-use class），每种100张，尺寸为256*256。由于数据集比较少，我们使用数据增广的方式增加数据，这里使用了水平裁剪和垂直裁剪，90°旋转的方法，在GTX 1080的配置下用了4个小时。
那么，在哪里才能下载到这个数据集呢？作者给出了这个数据集的百度云下载地址：https://pan.baidu.com/s/1i5zQNdj
所以，我们可以直接下载好数据集，然后解压，放在'data/'路径下，一定要按照上面提到的文件结构放置才行。弄好了之后，我们可以大致来看看这个数据集：
可以看到很清晰的飞机还有棒球场，当然还有其他类别的：
建筑物和丛林也都非常清晰，数据集的质量还是不错的。
3. 辅助函数文件utilities.py文件
这个文件中主要放置了保存图像等相关操作，下面直接给出代码：
import scipy.misc
import numpy as np
def center_crop(x, crop_h, crop_w=None, resize_w=64):
if crop_w is None:
crop_w = crop_h
h, w = x.shape[:2]
j = int(round((h - crop_h)/2.))
i = int(round((w - crop_w)/2.))
return scipy.misc.imresize(x[j:j+crop_h, i:i+crop_w],
[resize_w, resize_w])
def load_image(image_path, image_size, is_crop=True, resize_w=64):
image = scipy.misc.imread(image_path).astype(np.float)
if is_crop:
cropped_image = center_crop(image, image_size, resize_w=resize_w)
else:
cropped_image = image
return np.array(cropped_image)/127.5 - 1.
def get_labels(num_labels, lables_file):
style_labels = list(np.loadtxt(lables_file, str, delimiter='\n'))
if num_labels > 0:
style_labels = style_labels[:num_labels]
return style_labels
# 用于得到batch张图像，关键函数get_image
def imread(path, is_grayscale = False):
if is_grayscale:
return scipy.misc.imread(path, flatten = True).astype(np.float)
else:
return scipy.misc.imread(path).astype(np.float)
def transform(image, npx=64, is_crop=True, resize_w=64):
if is_crop:
cropped_image = center_crop(image, npx, resize_w=resize_w)
else:
cropped_image = image
return np.array(cropped_image)/127.5 - 1.
def get_image(image_path, image_size, is_crop=True, resize_w=64, is_grayscale=False):
return transform(imread(image_path, is_grayscale), image_size, is_crop, resize_w)
# 用于保存图像，关键函数save_images
def merge(images, size):
h, w = images.shape[1], images.shape[2]
img = np.zeros((h * size[0], w * size[1], 3))
for idx, image in enumerate(images):
i = idx % size[1]
j = idx // size[1]
img[j*h:j*h+h, i*w:i*w+w, :] = image
return img
def inverse_transform(images):
return (images+1.)/2. * 255
def imsave(images, size, path):
return scipy.misc.imsave(path, merge(images, size))
def save_images(images, size, image_path):
return imsave(inverse_transform(images), size, image_path)4. 网络结构network.py文件
这个文件定义了生成器和判别器，下面直接给出代码：
import tensorflow as tf
def generator(inputs, is_train=True, reuse=False):
output_size = 256
kernel = 4
batch_size = 64
gf_dim = 16
c_dim = 3
weight_init = tf.random_normal_initializer(stddev=0.01)
s2, s4, s8, s16, s32, s64 = int(output_size/2), int(output_size/4), int(output_size/8), int(output_size/16), int(output_size/32), int(output_size/64)
with tf.variable_scope('generator', reuse=reuse):
h0 = tf.layers.dense(inputs, units=gf_dim*32*s64*s64, activation=tf.identity, kernel_initializer=weight_init)
h0 = tf.reshape(h0, [-1, s64, s64, gf_dim*32])
h0 = tf.contrib.layers.batch_norm(h0, scale=True, is_training=is_train, scope="g_bn0")
h0 = tf.nn.relu(h0)
output1_shape = [batch_size, s32, s32, gf_dim*16]
w_h1 = tf.get_variable('g_w_h1', [kernel, kernel, output1_shape[-1], int(h0.get_shape()[-1])],
initializer=weight_init)
b_h1 = tf.get_variable('g_b_h1', [output1_shape[-1]], initializer=tf.constant_initializer(0))
h1 = tf.nn.conv2d_transpose(h0, w_h1, output_shape=output1_shape, strides=[1, 2, 2, 1],
padding='SAME', name='g_h1_deconv2d') + b_h1
h1 = tf.contrib.layers.batch_norm(h1, scale=True, is_training=is_train, scope="g_bn1")
h1 = tf.nn.relu(h1)
output2_shape = [batch_size, s16, s16, gf_dim*8]
w_h2 = tf.get_variable('g_w_h2', [kernel, kernel, output2_shape[-1], int(h1.get_shape()[-1])],
initializer=weight_init)
b_h2 = tf.get_variable('g_b_h2', [output2_shape[-1]], initializer=tf.constant_initializer(0))
h2 = tf.nn.conv2d_transpose(h1, w_h2, output_shape=output2_shape, strides=[1, 2, 2, 1],
padding='SAME', name='g_h2_deconv2d') + b_h2
h2 = tf.contrib.layers.batch_norm(h2, scale=True, is_training=is_train, scope="g_bn2")
h2 = tf.nn.relu(h2)
output3_shape = [batch_size, s8, s8, gf_dim*4]
w_h3 = tf.get_variable('g_w_h3', [kernel, kernel, output3_shape[-1], int(h2.get_shape()[-1])],
initializer=weight_init)
b_h3 = tf.get_variable('g_b_h3', [output3_shape[-1]], initializer=tf.constant_initializer(0))
h3 = tf.nn.conv2d_transpose(h2, w_h3, output_shape=output3_shape, strides=[1, 2, 2, 1],
padding='SAME', name='g_h3_deconv2d') + b_h3
h3 = tf.contrib.layers.batch_norm(h3, scale=True, is_training=is_train, scope="g_bn3")
h3 = tf.nn.relu(h3)
output4_shape = [batch_size, s4, s4, gf_dim*2]
w_h4 = tf.get_variable('g_w_h4', [kernel, kernel, output4_shape[-1], int(h3.get_shape()[-1])],
initializer=weight_init)
b_h4 = tf.get_variable('g_b_h4', [output4_shape[-1]], initializer=tf.constant_initializer(0))
h4 = tf.nn.conv2d_transpose(h3, w_h4, output_shape=output4_shape, strides=[1, 2, 2, 1],
padding='SAME', name='g_h4_deconv2d') + b_h4
h4 = tf.contrib.layers.batch_norm(h4, scale=True, is_training=is_train, scope="g_bn4")
h4 = tf.nn.relu(h4)
output5_shape = [batch_size, s2, s2, gf_dim*1]
w_h5 = tf.get_variable('g_w_h5', [kernel, kernel, output5_shape[-1], int(h4.get_shape()[-1])],
initializer=weight_init)
b_h5 = tf.get_variable('g_b_h5', [output5_shape[-1]], initializer=tf.constant_initializer(0))
h5 = tf.nn.conv2d_transpose(h4, w_h5, output_shape=output5_shape, strides=[1, 2, 2, 1],
padding='SAME', name='g_h5_deconv2d') + b_h5
h5 = tf.contrib.layers.batch_norm(h5, scale=True, is_training=is_train, scope="g_bn5")
h5 = tf.nn.relu(h5)
output6_shape = [batch_size, output_size, output_size, c_dim]
w_h6 = tf.get_variable('g_w_h6', [kernel, kernel, output6_shape[-1], int(h5.get_shape()[-1])],
initializer=weight_init)
b_h6 = tf.get_variable('g_b_h6', [output6_shape[-1]], initializer=tf.constant_initializer(0))
h6 = tf.nn.conv2d_transpose(h5, w_h6, output_shape=output6_shape, strides=[1, 2, 2, 1],
padding='SAME', name='g_h6_deconv2d') + b_h6
#logits = h6.outputs
#h6.outputs = tf.nn.tanh(h6.outputs)
return tf.nn.tanh(h6)
def discriminator(inputs, is_train=True, reuse=False):
kernel = 5
df_dim = 16
weight_init = tf.random_normal_initializer(stddev=0.01)
alpha_lrelu = 0.2
with tf.variable_scope('discriminator', reuse=reuse):
w_h0 = tf.get_variable('d_w_h0', [kernel, kernel, 3, df_dim], initializer=weight_init)
b_h0 = tf.get_variable('d_b_h0', [df_dim], initializer=tf.constant_initializer(0))
h0 = tf.nn.conv2d(inputs, w_h0, strides=[1,2,2,1], padding='SAME', name='d_h0_conv2d') + b_h0
h0 = tf.nn.leaky_relu(h0, alpha_lrelu)
w_h1 = tf.get_variable('d_w_h1', [kernel, kernel, h0.get_shape()[-1], df_dim*2], initializer=weight_init)
b_h1 = tf.get_variable('d_b_h1', [df_dim*2], initializer=tf.constant_initializer(0))
h1 = tf.nn.conv2d(h0, w_h1, strides=[1,2,2,1], padding='SAME', name='d_h1_conv2d') + b_h1
h1 = tf.contrib.layers.batch_norm(h1, is_training=is_train, scope="d_bn1")
h1 = tf.nn.leaky_relu(h1, alpha_lrelu)
w_h2 = tf.get_variable('d_w_h2', [kernel, kernel, h1.get_shape()[-1], df_dim*4], initializer=weight_init)
b_h2 = tf.get_variable('d_b_h2', [df_dim*4], initializer=tf.constant_initializer(0))
h2 = tf.nn.conv2d(h1, w_h2, strides=[1,2,2,1], padding='SAME', name='d_h2_conv2d') + b_h2
h2 = tf.contrib.layers.batch_norm(h2, is_training=is_train, scope="d_bn2")
h2 = tf.nn.leaky_relu(h2, alpha_lrelu)
w_h3 = tf.get_variable('d_w_h3', [kernel, kernel, h2.get_shape()[-1], df_dim*8], initializer=weight_init)
b_h3 = tf.get_variable('d_b_h3', [df_dim*8], initializer=tf.constant_initializer(0))
h3 = tf.nn.conv2d(h2, w_h3, strides=[1,2,2,1], padding='SAME', name='d_h3_conv2d') + b_h3
h3 = tf.contrib.layers.batch_norm(h3, is_training=is_train, scope="d_bn3")
h3 = tf.nn.leaky_relu(h3, alpha_lrelu)
global_max_h3 = tf.nn.max_pool(h3, [1,4,4,1], strides=[1,4,4,1], padding='SAME', name='d_h3_maxpool')
global_max_h3 = tf.layers.flatten(global_max_h3, name='d_h3_flatten')
w_h4 = tf.get_variable('d_w_h4', [kernel, kernel, h3.get_shape()[-1], df_dim*16], initializer=weight_init)
b_h4 = tf.get_variable('d_b_h4', [df_dim*16], initializer=tf.constant_initializer(0))
h4 = tf.nn.conv2d(h3, w_h4, strides=[1,2,2,1], padding='SAME', name='d_h4_conv2d') + b_h4
h4 = tf.contrib.layers.batch_norm(h4, is_training=is_train, scope="d_bn4")
h4 = tf.nn.leaky_relu(h4, alpha_lrelu)
global_max_h4 = tf.nn.max_pool(h4, [1,2,2,1], strides=[1,2,2,1], padding='SAME', name='d_h4_maxpool')
global_max_h4 = tf.layers.flatten(global_max_h4, name='d_h4_flatten')
w_h5 = tf.get_variable('d_w_h5', [kernel, kernel, h4.get_shape()[-1], df_dim*32], initializer=weight_init)
b_h5 = tf.get_variable('d_b_h5', [df_dim*32], initializer=tf.constant_initializer(0))
h5 = tf.nn.conv2d(h4, w_h5, strides=[1,2,2,1], padding='SAME', name='d_h5_conv2d') + b_h5
h5 = tf.contrib.layers.batch_norm(h5, is_training=is_train, scope="d_bn5")
h5 = tf.nn.leaky_relu(h5, alpha_lrelu)
global_max_h5 = tf.layers.flatten(h5, name='d_h5_flatten')
features = tf.concat([global_max_h3, global_max_h4, global_max_h5], -1, name='d_concat')
h6 = tf.layers.dense(features, units=1, activation=tf.identity, kernel_initializer=weight_init, name='d_h6_dense')
#logits = h6.outputs
#h6.outputs = tf.nn.sigmoid(h6.outputs)
return tf.nn.sigmoid(h6), features5. 训练网络train_net.py文件
该文件用于编写网络结构，训练网络，直接给出代码;
import tensorflow as tf
import network
import sys
import os
import numpy as np
from glob import glob
from random import shuffle
import utilities
import time
flags = tf.app.flags
flags.DEFINE_integer("epoch", 10, "Epoch to train")
flags.DEFINE_float("learning_rate", 0.001, "Learning rate of for adam")
flags.DEFINE_float("beta1", 0.9, "Momentum term of adam")
flags.DEFINE_integer("train_size", sys.maxsize, "The size of train images")
flags.DEFINE_integer("batch_size", 64, "The number of batch images")
flags.DEFINE_integer("image_size", 256, "The size of image to use (will be center cropped)")
flags.DEFINE_integer("output_size", 256, "The size of the output images to produce")
flags.DEFINE_integer("sample_size", 64, "The number of sample images")
flags.DEFINE_integer("c_dim", 3, "Dimension of image color")
flags.DEFINE_integer("z_dim", 100, "Dimensions of input niose to generator")
flags.DEFINE_integer("sample_step", 500, "The interval of generating sample")
flags.DEFINE_string("dataset", "uc_train_256_data", "The name of dataset [celebA, mnist, lsun]")
flags.DEFINE_string("checkpoint_dir", "checkpoint", "Directory name to save the checkpoints [checkpoint]")
flags.DEFINE_string("summaries_dir", "logs", "Directory name to save the summaries")
flags.DEFINE_string("sample_dir", "samples", "Directory name to save the image samples [samples]")
flags.DEFINE_boolean("is_train", True, "True for training, False for testing [False]")
flags.DEFINE_boolean("is_crop", False, "True for training, False for testing [False]")
flags.DEFINE_boolean("visualize", False, "True for visualizing, False for nothing [False]")
FLAGS = flags.FLAGS
def main(_):
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
if not os.path.exists(FLAGS.sample_dir):
os.makedirs(FLAGS.sample_dir)
if not os.path.exists(FLAGS.summaries_dir):
os.makedirs(FLAGS.summaries_dir)
with tf.device("/gpu:0"):
#with tf.device("/cpu:0"):
z = tf.placeholder(tf.float32, [FLAGS.batch_size, FLAGS.z_dim], name="g_input_noise")
x = tf.placeholder(tf.float32, [FLAGS.batch_size, FLAGS.output_size, FLAGS.output_size, FLAGS.c_dim], name='d_input_images')
Gz = network.generator(z)
Dx, Dfx = network.discriminator(x)
Dz, Dfz = network.discriminator(Gz, reuse=True)
d_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=Dx, labels=tf.ones_like(Dx)))
d_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=Dz, labels=tf.zeros_like(Dz)))
d_loss = d_loss_real + d_loss_fake
g_loss_perceptual = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = Dz, labels = tf.ones_like(Dz)))
g_loss_features = tf.reduce_mean(tf.nn.l2_loss(Dfx-Dfz))/(FLAGS.image_size*FLAGS.image_size)
g_loss = g_loss_perceptual + g_loss_features
tvars = tf.trainable_variables()
d_vars = [var for var in tvars if 'd_' in var.name]
g_vars = [var for var in tvars if 'g_' in var.name]
print(d_vars)
print("---------------")
print(g_vars)
with tf.variable_scope(tf.get_variable_scope(), reuse=False):
print("reuse or not: {}".format(tf.get_variable_scope().reuse))
assert tf.get_variable_scope().reuse == False, "Houston tengo un problem"
d_trainer = tf.train.AdamOptimizer(FLAGS.learning_rate, FLAGS.beta1).minimize(d_loss, var_list=d_vars)
g_trainer = tf.train.AdamOptimizer(FLAGS.learning_rate, FLAGS.beta1).minimize(g_loss, var_list=g_vars)
tf.summary.scalar("generator_loss_percptual", g_loss_perceptual)
tf.summary.scalar("generator_loss_features", g_loss_features)
tf.summary.scalar("generator_loss_total", g_loss)
tf.summary.scalar("discriminator_loss", d_loss)
tf.summary.scalar("discriminator_loss_real", d_loss_real)
tf.summary.scalar("discriminator_loss_fake", d_loss_fake)
images_for_tensorboard = network.generator(z, reuse=True)
tf.summary.image('Generated_images', images_for_tensorboard, 2)
merged = tf.summary.merge_all()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.30)
gpu_options.allow_growth = True
saver = tf.train.Saver()
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, allow_soft_placement=True)) as sess:
print("starting session")
summary_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train', sess.graph)
sess.run(tf.global_variables_initializer())
data_files = glob(os.path.join("./data", FLAGS.dataset, "*.jpg"))
model_dir = "%s_%s_%s" % (FLAGS.dataset, 64, FLAGS.output_size)
save_dir = os.path.join(FLAGS.checkpoint_dir, model_dir)
sample_seed = np.random.uniform(low=-1, high=1, size=(FLAGS.batch_size, FLAGS.z_dim)).astype(np.float32)
if FLAGS.is_train:
for epoch in range(FLAGS.epoch):
d_total_cost = 0.
g_total_cost = 0.
shuffle(data_files)
num_batches = min(len(data_files), FLAGS.train_size) // FLAGS.batch_size
#num_batches = 2
for batch_i in range(num_batches):
batch_files = data_files[batch_i*FLAGS.batch_size:(batch_i+1)*FLAGS.batch_size]
batch = [utilities.load_image(batch_file, FLAGS.image_size, is_crop=FLAGS.is_crop, resize_w=FLAGS.output_size) for batch_file in batch_files]
batch_x = np.array(batch).astype(np.float32)
batch_z = np.random.normal(-1, 1, size=[FLAGS.batch_size, FLAGS.z_dim]).astype(np.float32)
start_time = time.time()
# print(batch[0])
d_err, _ = sess.run([d_loss, d_trainer], feed_dict={z: batch_z, x: batch_x})
g_err, _ = sess.run([g_loss, g_trainer], feed_dict={z: batch_z, x: batch_x})
d_total_cost += d_err
g_total_cost += g_err
if batch_i % 10 == 0:
summary = sess.run(merged, feed_dict={x: batch_x, z: batch_z})
summary_writer.add_summary(summary, (epoch-1)*(num_batches/30)+(batch_i/30))
print("Epoch: [%2d/%2d] [%4d/%4d] time: %4.4f, d_loss: %.8f, g_loss: %.8f" % (
epoch, FLAGS.epoch, batch_i, num_batches, time.time() - start_time, d_err, g_err))
# update sample files based on shuffled data
sample_files = batch_files[0:FLAGS.batch_size]
sample = [utilities.get_image(sample_file, FLAGS.image_size,
is_crop=FLAGS.is_crop,
resize_w=FLAGS.output_size,
is_grayscale=0) for sample_file in sample_files]
sample_images = np.array(sample).astype(np.float32)
if np.mod(batch_i, 10) == 0:
# generate and visualize generated images
# img, errD, errG = sess.run([net_g2.outputs, d_loss, g_loss], feed_dict={z : sample_seed, real_images: sample_images})
img = sess.run(Gz, feed_dict={z: sample_seed, x: sample_images})
utilities.save_images(img, [8, 8],
'./{}/train_{:02d}_{:03d}.png'.format(FLAGS.sample_dir, epoch, batch_i))
print("Epoch:", '%04d' % (epoch+1), "d_cost= {:.9f}".format(d_total_cost/num_batches),
"g_cost=", "{:.9f}".format(g_total_cost/num_batches))
save_path = saver.save(sess, save_dir)
print("Model saved in path: %s" % save_path)
sys.stdout.flush()
sess.close()
if __name__ == '__main__':
tf.app.run()
6. 提取特征feature.py文件
该文件用于提取训练好的网络特征，直接给出代码：
import sys
import os
import numpy as np
from glob import glob
from random import shuffle
import utilities
flags = tf.app.flags
flags.DEFINE_integer("train_size", sys.maxsize, "The size of train images")
flags.DEFINE_integer("batch_size", 64, "The number of batch images")
flags.DEFINE_integer("image_size", 256, "The size of image to use (will be center cropped)")
flags.DEFINE_integer("output_size", 256, "The size of the output images to produce [64]")
flags.DEFINE_integer("sample_size", 64, "The number of sample images [64]")
flags.DEFINE_integer("features_size", 14336, "Number of features for one image")
flags.DEFINE_integer("c_dim", 3, "Dimension of image color. [3]")
flags.DEFINE_integer("sample_step", 500, "The interval of generating sample. [500]")
flags.DEFINE_string("train_dataset", "uc_train_256_data", "The name of dataset")
flags.DEFINE_string("test_dataset", "uc_test_256", "The name of dataset")
flags.DEFINE_string("checkpoint_dir", "checkpoint", "Directory name to save the checkpoints [checkpoint]")
flags.DEFINE_string("sample_dir", "samples", "Directory name to save the image samples [samples]")
flags.DEFINE_string("feature_dir", "features", "Directory name to save features")
flags.DEFINE_boolean("is_train", False, "True for training, False for testing [False]")
flags.DEFINE_boolean("is_crop", False, "True for training, False for testing [False]")
flags.DEFINE_boolean("visualize", True, "True for visualizing, False for nothing [False]")
flags.DEFINE_integer("num_labels", 21, "Number of different labels")
flags.DEFINE_string("labels_file", "style_names.txt", "File containing a list of labels")
FLAGS = flags.FLAGS
def main(_):
if not os.path.exists(FLAGS.checkpoint_dir):
print("Houston tengo un problem: No checkPoint directory found")
return 0
if not os.path.exists(FLAGS.feature_dir):
os.makedirs(FLAGS.feature_dir)
if not os.path.exists(FLAGS.sample_dir):
os.makedirs(FLAGS.sample_dir)
#with tf.device("/gpu:0"):
with tf.device("/cpu:0"):
x = tf.placeholder(tf.float32, [FLAGS.batch_size, FLAGS.output_size, FLAGS.output_size, FLAGS.c_dim], name='d_input_images')
d_netx, Dfx = network.discriminator(x, is_train=FLAGS.is_train, reuse=False)
saver = tf.train.Saver()
with tf.Session() as sess:
print("starting session")
sess.run(tf.global_variables_initializer())
model_dir = "%s_%s_%s" % (FLAGS.train_dataset, 64, FLAGS.output_size)
save_dir = os.path.join(FLAGS.checkpoint_dir, model_dir)
labels = utilities.get_labels(FLAGS.num_labels, FLAGS.labels_file)
saver.restore(sess, save_dir)
print("Model restored from file: %s" % save_dir)
#extracting features from train dataset
extract_features(x, labels, sess, Dfx)
#extracting features from test dataset
extract_features(x, labels, sess, Dfx, training=False)
sess.close()
def extract_features(x, labels, sess, Dfx, training=True):
if training:
data_path = FLAGS.train_dataset
features_path = "features_train.npy"
labels_path = "labels_train.npy"
else:
data_path = FLAGS.test_dataset
features_path = "features_test.npy"
labels_path = "labels_test.npy"
data_files = glob(os.path.join("./data", data_path, "*.jpg"))
shuffle(data_files)
num_batches = min(len(data_files), FLAGS.train_size) // FLAGS.batch_size
#num_batches =2
num_examples = num_batches*FLAGS.batch_size
y = np.zeros(num_examples, dtype=np.uint8)
for i in range(num_examples):
for j in range(len(labels)):
if labels[j] in data_files[i]:
y[i] = j
break
features = np.zeros((num_examples, FLAGS.features_size))
for batch_i in range(num_batches):
batch_files = data_files[batch_i*FLAGS.batch_size:(batch_i+1)*FLAGS.batch_size]
batch = [utilities.load_image(batch_file, FLAGS.image_size, is_crop=FLAGS.is_crop, resize_w=FLAGS.output_size) for batch_file in batch_files]
batch_x = np.array(batch).astype(np.float32)
f = sess.run(Dfx, feed_dict={x: batch_x})
begin = FLAGS.batch_size*batch_i
end = FLAGS.batch_size + begin
features[begin:end, ...] = f
print("Features Extracted, Now saving")
np.save(os.path.join(FLAGS.feature_dir, features_path), features)
np.save(os.path.join(FLAGS.feature_dir, labels_path), y)
print("Features Saved")
sys.stdout.flush()
if __name__ == '__main__':
tf.app.run()7. SVM分类文件train_svm.py文件
作者最后是用svm做了一个简单的分类，这里直接给出代码
from sklearn.metrics import accuracy_score
from sklearn import svm
import numpy as np
accuracy = []
x_train = np.load('features/features_train.npy')
y_train = np.load('features/labels_train.npy')
x_test = np.load('features/features_test.npy')
y_test = np.load('features/labels_test.npy')
print("Fitting the classifier to the training set")
C = 1000.0 # SVM regularization parameter
clf = svm.SVC(kernel='linear', C=C).fit(x_train, y_train)
print("Predicting...")
y_pred = clf.predict(x_test)
print("Accuracy: %.3f" % (accuracy_score(y_test, y_pred)))
accuracy.append(accuracy_score(y_test, y_pred))
print(accuracy)四、实验结果
这个实验的实现过程还稍微有点复杂，首先先运行train_net.py文件，运行完毕之后运行feature.py文件，最后在运行train_svm.py文件。这个模型我是用cpu跑了，跑了一天，运行完train_svm.py文件的最终输出为：
精度为71.35%，跟论文里作者的结果差的还蛮多的，后面再看一下我的生成结果，几乎一片噪声：
就完全没有生成啥结果。。。。。
五、分析
1. 目前只有一个想法，我跑完模型没啥效果，有空我再好好检查一下自己的模型。
2. 关于the UC Merced Land Use Dataset数据集，目前用深度学习做遥感图像的用的还挺多，补充一个用caffe做的例子，参考链接：https://github.com/yangxue0827/CNN_UCMerced-LandUse_Caffe
5472
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
