本文主要是介绍【Tensor'flow】第一个FCN网络，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！
学习Tensorflow，写一个超级简单的全卷积，效果没有，只是能跑通，没有dropout。
#!/usr/bin/env python
#coding:utf-8
from __future__ import absolute_import
from __future__ import divisionimport os,cv2
import numpy as np
import time
import tensorflow as tf
def weight_variable(shape):# 使用截断的正态分布初始权重initial = tf.truncated_normal(shape, stddev = 0.01)return tf.Variable(initial)def bias_variable(shape):return tf.Variable(tf.constant(0.0, shape = shape))def conv_layer(x, W, b):# W的尺寸是[ksize, ksize, input, output]conv = tf.nn.conv2d(x, W, strides = [1, 1, 1, 1], padding = 'SAME')conv_b = tf.nn.bias_add(conv, b)conv_relu = tf.nn.relu(conv_b)return conv_reludef max_pool_layer(x):return tf.nn.max_pool(x, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1], padding = 'SAME')def deconv_layer(x, W, output_shape, b):# strides = 2 两倍上卷积# output_shape = [batch_size, output_width, output_height, output_channel],注意第一个是batch_size# 权重W = [ksize, ksize, output, input]后两位和卷积相反deconv = tf.nn.conv2d_transpose(x, W, output_shape, strides = [1, 2, 2, 1], padding = 'SAME')return tf.nn.bias_add(deconv, b)# 获取数据
def get_data(image_path, label_path):image_list = os.listdir(image_path)label_list = os.listdir(label_path)image_list_arr = []label_list_arr = []for file in image_list:if file[-3:] == 'png':# cv2.imread('', -1)保持原始数据读入；如果没有-1会以图片形式读入，变成三通道image = cv2.imread(os.path.join(image_path,file)，-1)#image = transform.resize(image, (512,512))image_list_arr.append(image)for file in label_list:if file[-3:] == 'png':label = cv2.imread(os.path.join(label_path,file)， -1)label_list_arr.append(label)return (image_list_arr, label_list_arr)# 读取下一个batch数据
def next_batch(images, labels, batch_size, shuffle = False):assert len(images) == len(labels)if shuffle:indices = np.arange(len(images))np.random.shuffle(indices)for start_idx in range(0, len(images) - batch_size + 1, batch_size):if shuffle:exceprt = indices[start_idx : start_idx + batch_size]else:exceprt = slice(start_idx, start_idx + batch_size)yield np.array(images)[exceprt], np.array(labels)[exceprt]def main():# 尽量写相对路径image_path = './data/mri'label_path = './data/labels'# 如果内存耗尽可以考虑将batch减小batch_size = 4n_epoch = 2lr = 0.01images, labels = get_data(image_path, label_path)ratio = 0.8length = len(images)s = np.int(length * ratio)x_train = images[: s]y_train = labels[: s]x_val = images[s: ]y_val = labels[s:]keep_prob = tf.placeholder(tf.float32)# None代表样本数量不固定x = tf.placeholder(tf.float32, shape = [None, 256, 256, 3])y = tf.placeholder(tf.float32, shape = [None, 256, 256, 3])# input 256*256# weight([ksize, ksize, input, output])weight1 = weight_variable([3, 3, 3, 64])bias1 = bias_variable([64])conv1 = conv_layer(x, weight1, bias1)# input 256*256# output 128*128weight2 = weight_variable([3, 3, 64, 128])bias2 = bias_variable([128])conv2 = conv_layer(conv1, weight2, bias2)pool1 = max_pool_layer(conv2)# input 128*128# output 64*64weight3 = weight_variable([3, 3, 128, 256])bias3 = bias_variable([256])conv3 = conv_layer(pool1, weight3, bias3)pool2 = max_pool_layer(conv3)# deconv1# weight([ksize, ksize, output, input])# 64*64->128*128(pool1)deconv_weight1 = weight_variable([3, 3, 128, 256])deconv_b1 = bias_variable([128])deconv1 = deconv_layer(pool2, deconv_weight1, [batch_size, 128, 128, 128], deconv_b1)# 与pool1融合，使用add的话deconv和pool的output channel要一致fuse_pool1 = tf.add(deconv1, pool1)# deconv2# 128*128->256*256(input)deconv_weight2 = weight_variable([3, 3, 64, 128])deconv_b2 = bias_variable([64])deconv2 = deconv_layer(fuse_pool1, deconv_weight2, [batch_size, 256, 256, 64], deconv_b2)# 转换成与输入标签相同的size，获得最后结果weight16 = weight_variable([3, 3, 64, 3])bias16 = bias_variable([3])conv16 = tf.nn.conv2d(deconv2, weight16, strides = [1, 1, 1, 1], padding = 'SAME')conv16_b = tf.nn.bias_add(conv16, bias16)logits16 = conv16_b# lossloss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits16, labels=y))opt = tf.train.AdamOptimizer(1e-4).minimize(loss)sess = tf.Session()sess.run(tf.global_variables_initializer())for epoch in range(n_epoch):# trainfor x_train_batch, y_train_batch in next_batch(x_train, y_train, batch_size, shuffle = True):_, train_loss = sess.run([opt, loss], feed_dict = {x: x_train_batch, y: y_train_batch})print ("------trian loss: %f" % train_loss)# valval_loss = 0for x_val_batch, y_val_batch in next_batch(x_val, y_val, batch_size, shuffle = True):val_loss = sess.run([loss], feed_dict={x: x_val_batch, y: y_val_batch})print("------val loss : %f" % val_loss)sess.close()if __name__ == '__main__':main()
这篇关于【Tensor'flow】第一个FCN网络的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！
