對於去噪自編碼器，網上好多都是利用全連接神經網絡進行構建，我自己寫了一個卷積、反捲積形式的去噪自編碼器，其中的參數調優如果有興趣的話，可以自行修改查看結果。
數據集使用最簡單的mnist：
網絡結構：
mnist輸入（28*28=784向量） => 28*28*1矩陣 => 卷積層1 => 14*14*64 => 卷積層2 => 7*7*64 => 卷積層3 => 4*4*32 => 反捲積層1 => 7×7*32 => 反捲積層2 => 14*14*64 => 反捲積層3 => 28*28*64 => 卷積層X => 28×28*1
訓練：
我用train集訓練train_epochs輪，然後用test集對訓練好的模型進行評測，同時保存加噪圖像及對應的去噪圖像。
Code:
#! -*- coding: utf-8 -*-
## by Colie (lijixiang)
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np
from PIL import Image
train_epochs = 35 ## int(1e5+1)
INPUT_HEIGHT = 28
INPUT_WIDTH = 28
batch_size = 256
noise_factor = 0.5 ## (0~1)
## 原始輸入是28×28*3
input_x = tf.placeholder(tf.float32, [None, INPUT_HEIGHT * INPUT_WIDTH], name='input_with_noise')
input_matrix = tf.reshape(input_x, shape=[-1, INPUT_HEIGHT, INPUT_WIDTH, 1])
input_raw = tf.placeholder(tf.float32, shape=[None, INPUT_HEIGHT * INPUT_WIDTH], name='input_without_noise')
## 1 conv layer
## 輸入28*28*3
## 經過卷積、激活、池化，輸出14*14*64
weight_1 = tf.Variable(tf.truncated_normal(shape=[3, 3, 1, 64], stddev=0.1, name = 'weight_1'))
bias_1 = tf.Variable(tf.constant(0.0, shape=[64], name='bias_1'))
conv1 = tf.nn.conv2d(input=input_matrix, filter=weight_1, strides=[1, 1, 1, 1], padding='SAME')
conv1 = tf.nn.bias_add(conv1, bias_1, name='conv_1')
acti1 = tf.nn.relu(conv1, name='acti_1')
pool1 = tf.nn.max_pool(value=acti1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='max_pool_1')
## 2 conv layer
## 輸入14*14*64
## 經過卷積、激活、池化，輸出7×7×64
weight_2 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 64], stddev=0.1, name='weight_2'))
bias_2 = tf.Variable(tf.constant(0.0, shape=[64], name='bias_2'))
conv2 = tf.nn.conv2d(input=pool1, filter=weight_2, strides=[1, 1, 1, 1], padding='SAME')
conv2 = tf.nn.bias_add(conv2, bias_2, name='conv_2')
acti2 = tf.nn.relu(conv2, name='acti_2')
pool2 = tf.nn.max_pool(value=acti2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='max_pool_2')
## 3 conv layer
## 輸入7*7*64
## 經過卷積、激活、池化，輸出4×4×32
## 原始輸入是28*28*3=2352，轉化爲4*4*32=512，大量噪聲會在網絡中過濾掉
weight_3 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 32], stddev=0.1, name='weight_3'))
bias_3 = tf.Variable(tf.constant(0.0, shape=[32]))
conv3 = tf.nn.conv2d(input=pool2, filter=weight_3, strides=[1, 1, 1, 1], padding='SAME')
conv3 = tf.nn.bias_add(conv3, bias_3)
acti3 = tf.nn.relu(conv3, name='acti_3')
pool3 = tf.nn.max_pool(value=acti3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='max_pool_3')
## 1 deconv layer
## 輸入4*4*32
## 經過反捲積，輸出7*7*32
deconv_weight_1 = tf.Variable(tf.truncated_normal(shape=[3, 3, 32, 32], stddev=0.1), name='deconv_weight_1')
deconv1 = tf.nn.conv2d_transpose(value=pool3, filter=deconv_weight_1, output_shape=[batch_size, 7, 7, 32], strides=[1, 2, 2, 1], padding='SAME', name='deconv_1')
## 2 deconv layer
## 輸入7*7*32
## 經過反捲積，輸出14*14*64
deconv_weight_2 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 32], stddev=0.1), name='deconv_weight_2')
deconv2 = tf.nn.conv2d_transpose(value=deconv1, filter=deconv_weight_2, output_shape=[batch_size, 14, 14, 64], strides=[1, 2, 2, 1], padding='SAME', name='deconv_2')
## 3 deconv layer
## 輸入14*14*64
## 經過反捲積，輸出28*28*64
deconv_weight_3 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 64], stddev=0.1, name='deconv_weight_3'))
deconv3 = tf.nn.conv2d_transpose(value=deconv2, filter=deconv_weight_3, output_shape=[batch_size, 28, 28, 64], strides=[1, 2, 2, 1], padding='SAME', name='deconv_3')
## conv layer
## 輸入28*28*64
## 經過卷積，輸出爲28*28*1
weight_final = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 1], stddev=0.1, name = 'weight_final'))
bias_final = tf.Variable(tf.constant(0.0, shape=[1], name='bias_final'))
conv_final = tf.nn.conv2d(input=deconv3, filter=weight_final, strides=[1, 1, 1, 1], padding='SAME')
conv_final = tf.nn.bias_add(conv_final, bias_final, name='conv_final')
## output
## 輸入28*28*1
## reshape爲28*28
output = tf.reshape(conv_final, shape=[-1, INPUT_HEIGHT * INPUT_WIDTH])
## loss and optimizer
loss = tf.reduce_mean(tf.pow(tf.subtract(output, input_raw), 2.0))
optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)
with tf.Session() as sess:
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
n_samples = int(mnist.train.num_examples)
print('train samples: %d' % n_samples)
print('batch size: %d' % batch_size)
total_batch = int(n_samples / batch_size)
print('total batchs: %d' % total_batch)
init = tf.global_variables_initializer()
sess.run(init)
for epoch in range(train_epochs):
for batch_index in range(total_batch):
batch_x, _ = mnist.train.next_batch(batch_size)
noise_x = batch_x + noise_factor * np.random.randn(*batch_x.shape)
noise_x = np.clip(noise_x, 0., 1.)
_, train_loss = sess.run([optimizer, loss], feed_dict={input_x: noise_x, input_raw: batch_x})
print('epoch: %04d\tbatch: %04d\ttrain loss: %.9f' % (epoch + 1, batch_index + 1, train_loss))
## 訓練結束後，用測試集測試，並保存加噪圖像、去噪圖像
n_test_samples = int(mnist.test.num_examples)
test_total_batch = int(n_test_samples / batch_size)
for i in range(test_total_batch):
batch_test_x, _ = mnist.test.next_batch(batch_size)
noise_test_x = batch_test_x + noise_factor * np.random.randn(*batch_test_x.shape)
noise_test_x = np.clip(noise_test_x, 0., 1.)
test_loss, pred_result = sess.run([loss, conv_final], feed_dict={input_x: noise_test_x, input_raw: batch_test_x})
print('test batch index: %d\ttest loss: %.9f' % (i + 1, test_loss))
for index in range(batch_size):
array = np.reshape(pred_result[index], newshape=[INPUT_HEIGHT, INPUT_WIDTH])
array = array * 255
image = Image.fromarray(array)
if image.mode != 'L':
image = image.convert('L')
image.save('./pred/' + str(i * batch_size + index) + '.png')
array_raw = np.reshape(noise_test_x[index], newshape=[INPUT_HEIGHT, INPUT_WIDTH])
array_raw = array_raw * 255
image_raw = Image.fromarray(array_raw)
if image_raw.mode != 'L':
image_raw = image_raw.convert('L')
image_raw.save('./pred/' + str(i * batch_size + index) + '_raw.png')
#break
去噪效果：
