基于tensorflow下的DenseNet 简易搭建

基于tensorflow下的DenseNet 简易搭建

前言

笔者最近在研究深度学习网络，正好看到了DenseNet网络，觉得简易可实现，然后基于了一维数据集完成了DenseNet网络搭建。
DenseNet网络结构

他主要是在每一个Dense_block层中，每一个卷积层都与前面的卷积层相关，具体操作是tf.concat操作，简单理解一下，就是在每个卷积层数据数据，都要并联前几个卷积层数据，这与一般的ResNet不同。从下面一张图中，可以清晰显示出来。

代码实现

我们这里主要分了两部分dense_block与transition_layer，一个实现了denseNet的每一个模块的功能，另一个完成每一个模块的相连（BN+Relu+卷积+池化）
dense_block

def dense_block(p,layers_nums=3):  
    for i in range(layers_nums):
        x = tf.layers.batch_normalization(p)
        x = tf.nn.relu(x)
        x=tf.layers.conv2d(inputs=x, filters=4, kernel_size=[1,3], strides=1, padding='same', kernel_initializer=tf.random_normal_initializer(stddev=0.01))
        x=tf.concat([x, p], axis=3)
        p=x

p代表输入数据，layers_nums代表，、每个模块有多少卷积层。我们默认为3.
transition_layer

def transition_layer(x):
    x = tf.layers.batch_normalization(x)
    x = tf.nn.relu(x)
    n_inchannels = x.get_shape().as_list()[3]
    n_outchannels = int(n_inchannels * 0.5)
    x = tf.layers.conv2d(inputs=x, filters=n_outchannels, kernel_size=1, strides=1, padding='same', kernel_initializer=tf.random_normal_initializer(stddev=0.01))
    x = tf.layers.average_pooling2d(inputs=x, pool_size=[1,2], strides=2)
    return x

完成模块之间的连接
剩下的功能实现与正常CNN相似，就不详细介绍了，接下来贴出详细代码。

完整代码

import tensorflow as tf
import numpy as np
import tensorflow.contrib.slim as slim
import time

matix=np.loadtxt("D:/深度学习/all_feature/15—class_label15_four_train80.txt")  
matix1=np.loadtxt("D:/深度学习/all_feature/15—class_label15_four_test40.txt")
tf.reset_default_graph()
batch_size = 100
learning_rate = 0.001
learning_rate_decay = 0.95
#num_layers_in_dense_block=3
def conv_layer(x):
    return tf.layers.conv2d(inputs=x, filters=4, kernel_size=[1,3], strides=1, padding='same', kernel_initializer=tf.random_normal_initializer(stddev=0.01))

def dense_block(p,layers_nums=3):  
    for i in range(layers_nums):
        x = tf.layers.batch_normalization(p)
        x = tf.nn.relu(x)
        x=tf.layers.conv2d(inputs=x, filters=4, kernel_size=[1,3], strides=1, padding='same', kernel_initializer=tf.random_normal_initializer(stddev=0.01))
        x=tf.concat([x, p], axis=3)
        p=x
    return x
def transition_layer(x):
    x = tf.layers.batch_normalization(x)
    x = tf.nn.relu(x)
    n_inchannels = x.get_shape().as_list()[3]
    n_outchannels = int(n_inchannels * 0.5)
    x = tf.layers.conv2d(inputs=x, filters=n_outchannels, kernel_size=1, strides=1, padding='same', kernel_initializer=tf.random_normal_initializer(stddev=0.01))
    x = tf.layers.average_pooling2d(inputs=x, pool_size=[1,2], strides=2)
    return x
def inference(inputs):
    x = tf.reshape(inputs,[-1,1,1015,1])
    conv_1 = conv_layer(x)
    dense_1 = dense_block(conv_1,3)
    block_1=transition_layer(dense_1)
#    dense_2 = dense_block(block_1,4)
#    block_2=transition_layer(dense_2)

    net_flatten = slim.flatten(block_1,scope='flatten')
    fc_1 = slim.fully_connected(slim.dropout(net_flatten,0.8),200,activation_fn=tf.nn.tanh,scope='fc_1')
    output = slim.fully_connected(slim.dropout(fc_1,0.8),15,activation_fn=None,scope='output_layer')
    return output,net_flatten

def train():
    x = tf.placeholder(tf.float32, [None, 1015])
    y = tf.placeholder(tf.float32, [None, 15])
    y_outputs,net = inference(x)
    global_step = tf.Variable(0, trainable=False)

    entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y_outputs, labels=tf.argmax(y, 1))
    loss = tf.reduce_mean(entropy)
    train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step)

    prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_outputs, 1))
    accuracy = tf.reduce_mean(tf.cast(prediction, tf.float32))

#    saver = tf.train.Saver()
    print('Start to train the DenseNet......')
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        for i in range(3000):
            train_op_, loss_, step = sess.run([train_op, loss, global_step], feed_dict={x: matix[:,:1015], y: matix[:,1015:]})
            if i % 1 == 0:

                
                result = sess.run(loss, feed_dict={x: matix1[:,:1015], y: matix1[:,1015:]})
                acc_1=sess.run(accuracy, feed_dict={x: matix1[:,:1015], y: matix1[:,1015:]})
                
                print(str(i)+' '+str(result)+' '+str(acc_1))
def main(_):
    train()
if __name__ == '__main__':
    t0 = time.clock()
    tf.app.run()
    print(time.clock() - t0, " seconds process time ")