Tensorboard特征图可视化

秒速五厘米 2022-05-24 10:18 287阅读 0赞

## 前言 ##

实现孪生卷积神经网络时，验证是否共享了参数，进行了特征图可视化。由于卷积层特征图通道都很大，需要进行重新reshape成Tensorboard可以显示的通道数（１或３或４），这里全部变成了单通道的特征图。没有使用tf.reshape()是因为发现有问题，所以还是自己重新利用tf.concat进行reshape。

## 效果 ##

![这里写图片描述][70]

![这里写图片描述][70 1]

## 代码提示 ##

主要是最后两个函数，一个添加Tensorboard，一个对特征图进行reshape,很简单的思路。如果嫌特征图太小了，可以选择每次选一半的特征图进行reshape。见代码注释！

## 代码 ##

# coding=utf-8
    
    """孪生卷积神经网络"""
    
    import tensorflow as tf
    import numpy as np
    import math
    
    
    class SiameseNet(object):
    
        def __init__(self):
            print("正在构建孪生网络...")
            self.opts = {
       'trainWeightDecay': 0.0, 'stddev': 0.01}
    
        def build_siamese_cnn_network(self, input):
            first_image = tf.expand_dims(input[:, :, :, 0], 3)
            second_image = tf.expand_dims(input[:, :, :, 1], 3)
    
            with tf.variable_scope('siamese_CNN') as scope:
                first_output = self._build_branch(first_image, True)
                scope.reuse_variables()
                second_output = self._build_branch(second_image, False)
    
            return first_output, second_output
    
        def _build_branch(self, image, branch):
            print("构建孪生网络分支...")
    
            with tf.variable_scope('conv_block1_1'):
                print("构建conv1,relu1...")
                name = tf.get_variable_scope().name
                outputs = self._conv(image, 64, 3, 1, self.opts['trainWeightDecay'], self.opts['stddev'])
                outputs = tf.nn.relu(outputs)
    
            with tf.variable_scope('conv_block1_2'):
                print("构建conv2,relu2,max-pooling1...")
                name = tf.get_variable_scope().name
                outputs = self._conv(outputs, 64, 3, 1, self.opts['trainWeightDecay'], self.opts['stddev'])
                outputs = tf.nn.relu(outputs)
                outputs = self._maxPool(outputs, 2, 2)
                conv_block1 = outputs
    
            with tf.variable_scope('conv_block2_1'):
                print("构建conv3,relu3...")
                name = tf.get_variable_scope().name
                outputs = self._conv(outputs, 64, 3, 1, self.opts['trainWeightDecay'], self.opts['stddev'])
                outputs = tf.nn.relu(outputs)
    
            with tf.variable_scope('conv_block2_2'):
                print("构建,conv4,relu4,max-pooling2...")
                name = tf.get_variable_scope().name
                outputs = self._conv(outputs, 64, 3, 1, self.opts['trainWeightDecay'], self.opts['stddev'])
                outputs = tf.nn.relu(outputs)
                outputs = self._maxPool(outputs, 2, 2)
                conv_block2 = outputs
    
            with tf.variable_scope('conv_block3_1'):
                print("构建conv5,relu5...")
                name = tf.get_variable_scope().name
                outputs = self._conv(outputs, 128, 3, 1, self.opts['trainWeightDecay'], self.opts['stddev'])
                outputs = tf.nn.relu(outputs)
    
            with tf.variable_scope('conv_block3_2'):
                print("构建conv6,relu6,max-pooling3...")
                name = tf.get_variable_scope().name
                outputs = self._conv(outputs, 128, 3, 1, self.opts['trainWeightDecay'], self.opts['stddev'])
                outputs = tf.nn.relu(outputs)
                outputs = self._maxPool(outputs, 2, 2)
                conv_block3 = outputs
    
            with tf.variable_scope('conv_block4_1'):
                print("构建conv7,relu7...")
                name = tf.get_variable_scope().name
                outputs = self._conv(outputs, 128, 3, 1, self.opts['trainWeightDecay'], self.opts['stddev'])
                outputs = tf.nn.relu(outputs)
    
            with tf.variable_scope('conv_block4_2'):
                print("构建conv8,relu8,max-pooling4...")
                name = tf.get_variable_scope().name
                outputs = self._conv(outputs, 128, 3, 1, self.opts['trainWeightDecay'], self.opts['stddev'])
                outputs = tf.nn.relu(outputs)
    
            if branch:
                self.conv_block1_1 = conv_block1
                self.conv_block1_2 = conv_block2
                self.conv_block1_3 = conv_block3
                self.conv_block1_4 = outputs
            else:
                self.conv_block2_1 = conv_block1
                self.conv_block2_2 = conv_block2
                self.conv_block2_3 = conv_block3
                self.conv_block2_4 = outputs
    
            return outputs
    
        def _conv(self, inputs, filters, size, stride, wd, stddev, name=None):
            channels = int(inputs.get_shape()[-1])
    
            with tf.variable_scope('conv'):
                weights = self.getVariable('weights', shape=[size, size, channels, filters],
                                           initializer=tf.random_normal_initializer(),
                                           weightDecay=wd, dType=tf.float32, trainable=True)
                biases = self.getVariable('biases', shape=[filters, ],
                                          initializer=tf.constant_initializer(dtype=tf.float32),
                                          weightDecay=0.0, dType=tf.float32, trainable=True)
    
            p = np.floor((size - 1) / 2).astype(np.int32)
            p_x = tf.pad(inputs, [[0, 0], [p, p], [p, p], [0, 0]])
            conv = tf.nn.conv2d(p_x, weights, strides=[1, stride, stride, 1], padding='VALID')
            if name is not None:
                conv = tf.add(conv, biases, name=name)
            else:
                conv = tf.add(conv, biases)
    
            print('Layer Type = Conv, Size = %d * %d, Stride = %d, Filters = %d, Input channels = %d' % (
                size, size, stride, filters, channels))
            return conv
    
        def _maxPool(self, inputs, kSize, _stride):
            with tf.variable_scope('poll'):
                p = np.floor((kSize - 1) / 2).astype(np.int32)
                p_x = tf.pad(inputs, [[0, 0], [p, p], [p, p], [0, 0]])
                output = tf.nn.max_pool(p_x, ksize=[1, kSize, kSize, 1], strides=[1, _stride, _stride, 1],
                                        padding='VALID')
            return output
    
        def getVariable(self, name, shape, initializer, weightDecay=0.0, dType=tf.float32, trainable=True):
            if weightDecay > 0:
                regularizer = tf.contrib.layers.l2_regularizer(weightDecay)
            else:
                regularizer = None
    
            return tf.get_variable(name, shape=shape, initializer=initializer, dtype=dType, regularizer=regularizer,
                                   trainable=trainable)
    
        def build_summaries(self):
            with tf.name_scope('CNN_outputs'):
                tf.summary.image('conv_block1_1_1', self._concact_features(self.conv_block1_1[:, :, :, 0:16]), 1)  # 取部分特征图显示
                tf.summary.image('conv_block1_1_2', self._concact_features(self.conv_block1_1[:, :, :, 16:32]), 1)
                tf.summary.image('conv_block1_1_3', self._concact_features(self.conv_block1_1[:, :, :, 32:48]), 1)
                tf.summary.image('conv_block1_1_4', self._concact_features(self.conv_block1_1[:, :, :, 48:64]), 1)
                tf.summary.image('conv_block1_2', self._concact_features(self.conv_block1_2), 1)
                tf.summary.image('conv_block1_3', self._concact_features(self.conv_block1_3), 1)
                tf.summary.image('conv_block1_4', self._concact_features(self.conv_block1_4), 1)
                tf.summary.image('conv_block2_1', self._concact_features(self.conv_block2_1), 1)
                tf.summary.image('conv_block2_2', self._concact_features(self.conv_block2_2), 1)
                tf.summary.image('conv_block2_3', self._concact_features(self.conv_block2_3), 1)
                tf.summary.image('conv_block2_4', self._concact_features(self.conv_block2_4), 1)
    
        def _concact_features(self, conv_output):
            """ 对特征图进行reshape拼接 :param conv_output:输入多通道的特征图 :return: """
            num_or_size_splits = conv_output.get_shape().as_list()[-1]
            each_convs = tf.split(conv_output, num_or_size_splits=num_or_size_splits, axis=3)
            concact_size = int(math.sqrt(num_or_size_splits) / 1)
            all_concact = None
            for i in range(concact_size):
                row_concact = each_convs[i * concact_size]
                for j in range(concact_size - 1):
                    row_concact = tf.concat([row_concact, each_convs[i * concact_size + j + 1]], 1)
                if i == 0:
                    all_concact = row_concact
                else:
                    all_concact = tf.concat([all_concact, row_concact], 2)
    
            return all_concact

## 其他思路 ##

1.  [http://www.360doc.com/content/17/0414/16/10408243\_645594433.shtml][http_www.360doc.com_content_17_0414_16_10408243_645594433.shtml]
2.  [https://blog.csdn.net/u014038273/article/details/78618500][https_blog.csdn.net_u014038273_article_details_78618500]

[70]: /images/20220524/7597ed6d712b4c9aa66a7a491135820c.png
[70 1]: /images/20220524/8a88520f68604c40a333aeffbd447ac1.png
[http_www.360doc.com_content_17_0414_16_10408243_645594433.shtml]: http://www.360doc.com/content/17/0414/16/10408243_645594433.shtml
[https_blog.csdn.net_u014038273_article_details_78618500]: https://blog.csdn.net/u014038273/article/details/78618500