【Tensor'flow】第一個FCN網絡

學習Tensorflow，寫一個超級簡單的全卷積，效果沒有，只是能跑通，沒有dropout。

#!/usr/bin/env python
#coding:utf-8
from __future__ import absolute_import
from __future__ import division

import os,cv2
import numpy as np
import time
import tensorflow as tf
def weight_variable(shape):
	# 使用截斷的正態分佈初始權重
	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_relu
	
def 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 兩倍上卷積
	# output_shape = [batch_size, output_width, output_height, output_channel],注意第一個是batch_size
	# 權重W = [ksize, ksize, output, input]後兩位和卷積相反
	deconv = tf.nn.conv2d_transpose(x, W, output_shape,  strides = [1, 2, 2, 1], padding = 'SAME')
	return tf.nn.bias_add(deconv, b)
	
# 獲取數據
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)保持原始數據讀入；如果沒有-1會以圖片形式讀入，變成三通道
			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)
	
# 讀取下一個batch數據
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():

	# 儘量寫相對路徑
	image_path = './data/mri'
	label_path =  './data/labels'
	# 如果內存耗盡可以考慮將batch減小
	batch_size = 4
	n_epoch = 2
	lr = 0.01
	images, labels = get_data(image_path, label_path)
	ratio = 0.8
	length = 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代表樣本數量不固定
	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*128
	weight2 = 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*64
	weight3 = 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)
	# 與pool1融合，使用add的話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)
	
	
	
	# 轉換成與輸入標籤相同的size，獲得最後結果
	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
	# loss
	loss = 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):
					
		# train
			
		for 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)
			
		# val
		val_loss = 0
		for 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()