您的位置:首页 > 其它

tf中的tf.sparse_to_dense的用法

2018-01-09 23:46 369 查看

tf.sparse_to_dense(sparse_indices, output_shape, sparse_values, default_value, name=None)

除去name参数用以指定该操作的name,与方法有关的一共四个参数
第一个参数sparse_indices:稀疏矩阵中那些个别元素对应的索引值。
     有三种情况:
    
sparse_indices
是个数,那么它只能指定一维矩阵的某一个元素
    
sparse_indices
是个向量,那么它可以指定一维矩阵的多个元素
    
sparse_indices
是个矩阵,那么它可以指定二维矩阵的多个元素
第二个参数output_shape:输出的稀疏矩阵的shape
第三个参数sparse_values:个别元素的值。
     分为两种情况:
     sparse_values
是个数:所有索引指定的位置都用这个数
     sparse_values
是个向量:输出矩阵的某一行向量里某一行对应的数(所以这里向量的长度应该和输出矩阵的行数对应,不然报错)
第四个参数default_value:未指定元素的默认值,一般如果是稀疏矩阵的话就是0了
举一个例子:
在mnist里面有一个把数字标签转化成onehot标签的操作,所谓onehot标签
就是:
如果标签是6那么对应onehot就是[ 0.  0.  0.  0.  0.  0.  1.  0.  0.  0.]
如果标签是1那么对应onehot就是[0.  1.  0.  0.  0.  0.  0.  0.  0.  0.]
如果标签是0那么对应onehot就是[1.  0.  0.  0.  0.  0.  0.  0.  0.  0.]
就是把标签变为适用于神经网络输出的形式。
[python] viewplain copyBATCHSIZE=6  label=tf.expand_dims(tf.constant([0,2,3,6,7,9]),1)  假设一个batch有6个样本,每个样本的label是0,2,3,6,7,9[python] viewplain copyindex=tf.expand_dims(tf.range(0, BATCHSIZE),1)  生成一个index表明一个batch里面每个样本对应的序号(至于这里为什么要调用tf.expand_dims我在上一篇博客已经解释过了,链接:http://blog.csdn.net/mao_xiao_feng/article/details/53366163)[python] viewplain copyconcated = tf.concat(1, [index, label])  最后把他们两个矩阵进行连接,连接以后的矩阵是这样的[python] viewplain copy[[0 0]   [1 2]   [2 3]   [3 6]   [4 7]   [5 9]]  [python] viewplain copyonehot_labels = tf.sparse_to_dense(concated, tf.pack([BATCHSIZE,10]), 1.0, 0.0)  最后一步,调用tf.sparse_to_dense输出一个onehot标签的矩阵,输出的shape就是行数为BATCHSIZE,列数为10的矩阵,指定元素值为1.0,其余元素值为0.0程序源码:[python] viewplain copyimport tensorflow as tf  import numpy      BATCHSIZE=6  label=tf.expand_dims(tf.constant([0,2,3,6,7,9]),1)  index=tf.expand_dims(tf.range(0, BATCHSIZE),1)  #use a matrix  concated = tf.concat(1, [index, label])  onehot_labels = tf.sparse_to_dense(concated, tf.pack([BATCHSIZE,10]), 1.0, 0.0)    #use a vector  concated2=tf.constant([1,3,4])  #onehot_labels2 = tf.sparse_to_dense(concated2, tf.pack([BATCHSIZE,10]), 1.0, 0.0)#cant use ,because output_shape is not a vector  onehot_labels2 = tf.sparse_to_dense(concated2, tf.pack([10]), 1.0, 0.0)#can use    #use a scalar  concated3=tf.constant(5)  onehot_labels3 = tf.sparse_to_dense(concated3, tf.pack([10]), 1.0, 0.0)    with tf.Session() as sess:      result1=sess.run(onehot_labels)      result2 = sess.run(onehot_labels2)      result3 = sess.run(onehot_labels3)      print ("This is result1:")      print (result1)      print ("This is result2:")      print (result2)      print ("This is result3:")      print (result3)  输出结果:[python] viewplain copyThis is result1:  [[ 1.  0.  0.  0.  0.  0.  0.  0.  0.  0.]   [ 0.  0.  1.  0.  0.  0.  0.  0.  0.  0.]   [ 0.  0.  0.  1.  0.  0.  0.  0.  0.  0.]   [ 0.  0.  0.  0.  0.  0.  1.  0.  0.  0.]   [ 0.  0.  0.  0.  0.  0.  0.  1.  0.  0.]   [ 0.  0.  0.  0.  0.  0.  0.  0.  0.  1.]]  This is result2:  [ 0.  1.  0.  1.  1.  0.  0.  0.  0.  0.]  This is result3:  [ 0.  0.  0.  0.  0.  1.  0.  0.  0.  0.] 
                                            
内容来自用户分享和网络整理,不保证内容的准确性,如有侵权内容,可联系管理员处理 点击这里给我发消息
标签: