TF girls系列（3）之CNN网络的搭建

CNN的网络结构：
conv+relu–> conv+relu+pool–>conv+relu–> conv+relu+pool–>fc+relu–>fc 。

# -*- coding: utf-8 -*-
"""
Created on Wed Mar 27 19:38:44 2019

@author: macheng
"""

from __future__ import print_function, division
import tensorflow as tf
from sklearn.metrics import confusion_matrix
import numpy as np
import load

#train_samples 是维度为[num_images, image_size, image_size, num_channels]大小的4维矩阵
#train_labels 是维度为[num_images, 10]大小的2维矩阵
train_samples, train_labels = load._train_samples, load._train_labels
test_samples, test_labels = load._test_samples,  load._test_labels

print('Training set', train_samples.shape, train_labels.shape)
print('    Test set', test_samples.shape, test_labels.shape)

image_size = load.image_size       #32
num_labels = load.num_labels       #10
num_channels = load.num_channels   #channel为1，灰度图


def get_chunk(samples, labels, chunkSize):

	if len(samples) != len(labels):
		raise Exception('Length of samples and labels must equal')
	stepStart = 0  # initial step
	i = 0
	while stepStart < len(samples):
		stepEnd = stepStart + chunkSize
		if stepEnd < len(samples):
			yield i, samples[stepStart:stepEnd], labels[stepStart:stepEnd]
			i += 1
		stepStart = stepEnd


class Network():
	def __init__(self, num_hidden, batch_size, conv_depth, kernel_size, pooling_scale):

		self.batch_size = batch_size
		self.test_batch_size = 500

		# Hyper Parameters
		self.num_hidden = num_hidden
		self.kernel_size = kernel_size     # 卷积核的大小
		self.conv1_depth = conv_depth      # 卷积的深度（也就是卷积核的个数）
		self.conv2_depth = conv_depth
		self.conv3_depth = conv_depth
		self.conv4_depth = conv_depth
		self.last_conv_depth = self.conv4_depth
		self.pooling_scale = pooling_scale
		self.pooling_stride = self.pooling_scale  # Max Pooling Stride

		# Graph Related
		self.graph = tf.Graph()
		self.tf_train_samples = None
		self.tf_train_labels = None
		self.tf_test_samples = None
		self.tf_test_labels = None
		self.tf_test_prediction = None

		# 统计
		self.merged = None
		self.train_summaries = []
		self.test_summaries = []

		"""
			添加summary.FileWriter
		"""
		self.define_graph()
		gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
		self.session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options), graph=self.graph)
		self.writer = tf.summary.FileWriter('./board', self.graph)

	def define_graph(self):
		'''
			定义我的计算图谱
		'''
		with self.graph.as_default():
			# 这里只是定义图谱中的各种变量
			with tf.name_scope('inputs'):
				self.tf_train_samples = tf.placeholder(
					tf.float32, shape=(self.batch_size, image_size, image_size, num_channels), name='tf_train_samples'
				)
				self.tf_train_labels  = tf.placeholder(
					tf.float32, shape=(self.batch_size, num_labels), name='tf_train_labels'
				)
				self.tf_test_samples  = tf.placeholder(
					tf.float32, shape=(self.test_batch_size, image_size, image_size, num_channels), name='tf_test_samples'
				)

			#  CNN不是全连接的，第一层卷积核参数的设定：长度、宽度、输入深度（通道数）、输出深度
			#  全连接神经网络是一个神经元对应一个bias，而CNN中是每一个卷积核对应一个bias。
			with tf.name_scope('conv1'):
				conv1_weights = tf.Variable(
					tf.truncated_normal([self.kernel_size, self.kernel_size, num_channels, self.conv1_depth], stddev=0.1))
				conv1_biases = tf.Variable(tf.zeros([self.conv1_depth]))

			#  第二层卷积核参数的设定：长度、宽度、输入深度、输出深度
			with tf.name_scope('conv2'):
				conv2_weights = tf.Variable(
					tf.truncated_normal([self.kernel_size, self.kernel_size, self.conv1_depth, self.conv2_depth], stddev=0.1))
				conv2_biases = tf.Variable(tf.constant(0.1, shape=[self.conv2_depth]))

			with tf.name_scope('conv3'):
				conv3_weights = tf.Variable(
					tf.truncated_normal([self.kernel_size, self.kernel_size, self.conv2_depth, self.conv3_depth], stddev=0.1))
				conv3_biases = tf.Variable(tf.constant(0.1, shape=[self.conv3_depth]))

			with tf.name_scope('conv4'):
				conv4_weights = tf.Variable(
					tf.truncated_normal([self.kernel_size, self.kernel_size, self.conv3_depth, self.conv4_depth], stddev=0.1))
				conv4_biases = tf.Variable(tf.constant(0.1, shape=[self.conv4_depth]))

			with tf.name_scope('fc1'):
				down_scale = self.pooling_scale ** 2  # because we do 2 times pooling of stride 2
				# 下面乘了self.last_conv_dept是因为要将其扁平化
				fc1_weights = tf.Variable(
					tf.truncated_normal([(image_size // down_scale) * (image_size // down_scale) * self.last_conv_depth, self.num_hidden], stddev=0.1)
				)
				fc1_biases = tf.Variable(tf.constant(0.1, shape=[self.num_hidden]))
				self.train_summaries.append(tf.summary.histogram('fc1_weights', fc1_weights))
				self.train_summaries.append(tf.summary.histogram('fc1_biases', fc1_biases))

			with tf.name_scope('fc2'):
				fc2_weights = tf.Variable(
					tf.truncated_normal([self.num_hidden, num_labels], stddev=0.1), name='fc2_weights'
				)
				fc2_biases = tf.Variable(tf.constant(0.1, shape=[num_labels]), name='fc2_biases')
				self.train_summaries.append(tf.summary.histogram('fc2_weights', fc2_weights))
				self.train_summaries.append(tf.summary.histogram('fc2_biases', fc2_biases))

			# 定义图谱的运算
			def model(data, train=True):
				"""
					第一层卷积层是与输入层连接的但不是全连接，而是根据卷积核大小部分连接
				"""
				with tf.name_scope('conv1_model'):
					with tf.name_scope('convolution'):
						conv1 = tf.nn.conv2d(data, filter=conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
						addition = conv1 + conv1_biases
					hidden = tf.nn.relu(addition)

					if not train:
						# transpose the output of an activation to image
						# conv1_activation_relu shape: (8, 32, 32, 64)
						# 64 filter maps from this convolution, that's 64 grayscale images
						# image size is 32x32
						# 8 is the batch_size, which means 8 times of convolution was performed
						# just use the last one (index 7) as record

						filter_map = hidden[-1]
						filter_map = tf.transpose(filter_map, perm=[2, 0, 1])
						filter_map = tf.reshape(filter_map, (self.conv1_depth, 32, 32, 1))
						self.test_summaries.append(
							tf.summary.image('conv1_relu', tensor=filter_map, max_outputs=self.conv1_depth))

				with tf.name_scope('conv2_model'):
					with tf.name_scope('convolution'):
						conv2 = tf.nn.conv2d(hidden, filter=conv2_weights, strides=[1, 1, 1, 1], padding='SAME')
						addition = conv2 + conv2_biases
					hidden = tf.nn.relu(addition)
					hidden = tf.nn.max_pool(
						hidden,
						ksize=[1, self.pooling_scale, self.pooling_scale, 1],
						strides=[1, self.pooling_stride, self.pooling_stride, 1],
						padding='SAME')

				with tf.name_scope('conv3_model'):
					with tf.name_scope('convolution'):
						conv3 = tf.nn.conv2d(hidden, filter=conv3_weights, strides=[1, 1, 1, 1], padding='SAME')
						addition = conv3 + conv3_biases
					hidden = tf.nn.relu(addition)

				with tf.name_scope('conv4_model'):
					with tf.name_scope('convolution'):
						conv4 = tf.nn.conv2d(hidden, filter=conv4_weights, strides=[1, 1, 1, 1], padding='SAME')
						addition = conv4 + conv4_biases
					hidden = tf.nn.relu(addition)
					# if not train:
					# 	filter_map = hidden[-1]
					# 	filter_map = tf.transpose(filter_map, perm=[2, 0, 1])
					# 	filter_map = tf.reshape(filter_map, (self.conv4_depth, 16, 16, 1))
					# 	tf.image_summary('conv4_relu', tensor=filter_map, max_images=self.conv4_depth)
					hidden = tf.nn.max_pool(
						hidden,
						ksize=[1, self.pooling_scale, self.pooling_scale, 1],
						strides=[1, self.pooling_stride, self.pooling_stride, 1],
						padding='SAME')

				# fully connected layer 1
				shape = hidden.get_shape().as_list()
				reshape = tf.reshape(hidden, [shape[0], shape[1] * shape[2] * shape[3]])

				with tf.name_scope('fc1_model'):
					fc1_model = tf.matmul(reshape, fc1_weights) + fc1_biases
					hidden = tf.nn.relu(fc1_model)

				# fully connected layer 2
				with tf.name_scope('fc2_model'):
					return tf.matmul(hidden, fc2_weights) + fc2_biases

			# Training computation.
			logits = model(self.tf_train_samples)
			with tf.name_scope('loss'):
				self.loss = tf.reduce_mean(
					tf.nn.softmax_cross_entropy_with_logits(labels=self.tf_train_labels, logits=logits)
				)
				self.train_summaries.append(tf.summary.scalar('Loss', self.loss))

			# Optimizer.
			with tf.name_scope('optimizer'):
				self.optimizer = tf.train.GradientDescentOptimizer(0.0001).minimize(self.loss)

			# Predictions for the training, validation, and test data.
			with tf.name_scope('train'):
				self.train_prediction = tf.nn.softmax(logits, name='train_prediction')
			with tf.name_scope('test'):
				self.test_prediction = tf.nn.softmax(model(self.tf_test_samples, train=False), name='test_prediction')

			self.merged_train_summary = tf.summary.merge(self.train_summaries)
			self.merged_test_summary = tf.summary.merge(self.test_summaries)

	def run(self):
		'''
		用到Session
		'''
		# private function
		def print_confusion_matrix(confusionMatrix):
			print('Confusion    Matrix:')
			for i, line in enumerate(confusionMatrix):
				print(line, line[i]/np.sum(line))
			a = 0
			for i, column in enumerate(np.transpose(confusionMatrix, (1, 0))):
				a += (column[i]/np.sum(column))*(np.sum(column)/26000)
				print(column[i]/np.sum(column),)
			print('\n', np.sum(confusionMatrix), a)

		with self.session as session:
			tf.initialize_all_variables().run()

			# 训练
			print('Start Training')
			for i, samples, labels in get_chunk(train_samples, train_labels, chunkSize=self.batch_size):
				_, l, predictions, summary = session.run(
					[self.optimizer, self.loss, self.train_prediction, self.merged_train_summary],
					feed_dict={self.tf_train_samples: samples, self.tf_train_labels: labels}
				)
				self.writer.add_summary(summary, i)

				# labels is True Labels
				accuracy, _ = self.accuracy(predictions, labels)
				if i % 50 == 0:
					print('Minibatch loss at step %d: %f' % (i, l))
					print('Minibatch accuracy: %.1f%%' % accuracy)
			#

			# 测试
			accuracies = []
			confusionMatrices = []
			for i, samples, labels in get_chunk(test_samples, test_labels, chunkSize=self.test_batch_size):
				result, summary = session.run(
					[self.test_prediction, self.merged_test_summary],
					feed_dict={self.tf_test_samples: samples}
				)

				self.writer.add_summary(summary, i)
				accuracy, cm = self.accuracy(result, labels, need_confusion_matrix=True)
				accuracies.append(accuracy)
				confusionMatrices.append(cm)
				print('%d Test Accuracy: %.1f%%' % (i, accuracy))
			print(' Average  Accuracy:', np.average(accuracies))
			print('Standard Deviation:', np.std(accuracies))
			print_confusion_matrix(np.add.reduce(confusionMatrices))

	def accuracy(self, predictions, labels, need_confusion_matrix=False):
		'''
		计算预测的正确率与召回率
		@return: accuracy and confusionMatrix as a tuple
		'''
		_predictions = np.argmax(predictions, 1)
		_labels = np.argmax(labels, 1)
		cm = confusion_matrix(_labels, _predictions) if need_confusion_matrix else None
		# == is overloaded for numpy array
		accuracy = (100.0 * np.sum(_predictions == _labels) / predictions.shape[0])
		return accuracy, cm


if __name__ == '__main__':
	net = Network(num_hidden=16, batch_size=64, kernel_size=3, conv_depth=16, pooling_scale=2)
	net.run()

CNN网络的正确率好像越来越低了，，，

tensorboard 可视化的网络结构如下图所示

从上面的代码及tensorborad可视化图可以看出，搭建CNN网络的一般步骤是：

1. 定义graph：主要包括输入的定义、卷积层的权重矩阵weights和bias的定义、全连接层的权重矩阵weights和bias的定义。
2. 定义计算模型model：
   （1）对1中的卷积层的运算的定义：使用1中定义的权重矩阵和bias进行卷积运算、激活函数的选择、pooling的选择、
   （2）对1中的全连接层的定义：使用1中定义的权重矩阵和bias进行卷积运算、激活函数的选择、
3. 定义优化器进行优化
4. 定义session进行运算

从代码中可以看出，4个卷积层的定义和运算很相似、2个全连接层的定义和运算也很相似，故下一篇博客会将上面的代码抽象出一个CNN网络的计算框架。