如何使用 indexlist 对 Tensor 进行切片并组成新的 Tensor

Question

如何使用 tensorflow Circulation 对一个 Tensor 进行切片并组成一个新的 Tensor 只是意味着：

    text_embeding =tf.constant(
                       #index 0       index 1      index 2
                   [[[0.1,0.2,0.3],[0.4,0.5,0.6],[0.1,0.2,0.3]], 
                    [[0.1,0.2,0.3],[0.4,0.5,0.6],[0.1,0.2,0.3]],
                    [[0.1,0.2,0.3],[0.4,0.5,0.6],[0.1,0.2,0.3]]
                   ] 
               )

我想要求batch中的每个张量根据索引的组合得到一个新的值列表 index_list = [[0,0],[1,1],[2,2],[0,1],[1,2],[0,2]]

我想获得价值 '''

[
    [
     [0.1,0.2,0.3 , 0.1,0.2,0.3], index0,0
     [0.4,0.5,0.6 , 0.4,0.5,0.6], index1,1
     [0.1,0.2,0.3 , 0.1,0.2,0.3], index2,2
     [0.1,0.2,0.3 , 0.4,0.5,0.6], index0,1
     [0.4,0.5,0.6 , 0.1,0.2,0.3], index1,2
     [0.1,0.2,0.3 , 0.1,0.2,0.3]  index0,2
    ],
    [
     [0.1,0.2,0.3 , 0.1,0.2,0.3], index0,0
     [0.4,0.5,0.6 , 0.4,0.5,0.6], index1,1
     [0.1,0.2,0.3 , 0.1,0.2,0.3], index2,2
     [0.1,0.2,0.3 , 0.4,0.5,0.6], index0,1
     [0.4,0.5,0.6 , 0.1,0.2,0.3], index1,2
     [0.1,0.2,0.3 , 0.1,0.2,0.3]  index0,2
    ],
    [
     [0.1,0.2,0.3 , 0.1,0.2,0.3], index0,0
     [0.4,0.5,0.6 , 0.4,0.5,0.6], index1,1
     [0.1,0.2,0.3 , 0.1,0.2,0.3], index2,2
     [0.1,0.2,0.3 , 0.4,0.5,0.6], index0,1
     [0.4,0.5,0.6 , 0.1,0.2,0.3], index1,2
     [0.1,0.2,0.3 , 0.1,0.2,0.3]  index0,2
    ]
   ]

''' 我的代码是这样的，但是在会话图准备好之前，batch_size=output_layer_sequence.shape[0] 是 None，这里是错误的！

vsp = tf.batch_gather(output_layer_sequence, tf.tile([[j, j + i]],multiples=[output_layer_sequence.shape[0],1]))  # batch * 2 * hidden_size

for i in range(2):
    for j in range(2):
        vsp = tf.batch_gather(output_layer_sequence, tf.tile([[j, j + i]],multiples=[16,1]))  # batch * 2 * hidden_size
        # vsp = tf.batch_gather(output_layer_sequence, tf.tile([[j, j + i]],multiples=[output_layer_sequence.shape[0],1]))  # batch * 2 * hidden_size
        vsp_start, vsp_end = tf.split(vsp, 2, 1)  # batch * 1 * hiddensize
        vsp_start = tf.squeeze(vsp_start)  # batch  * hiddensize
             
        vsp_end = tf.squeeze(vsp_end)  # batch * hiddensize
        vsp = tf.concat([vsp_start, vsp_end], axis=-1, name='concat')  # [batch ,2*hiddensize]

        span_logits = tf.matmul(vsp, output_span_weight, transpose_b=True)  # output:[batch,class_labels]

        span_logits = tf.nn.bias_add(span_logits, output_span_bias)  # [batch,class_labels]
        span_logit_sum.append(span_logits)

谢谢！

Answer 1

使用tf.gather():

text_embeding =tf.constant(
                    #index 0       index 1      index 2
                [[[0.1,0.2,0.3],[0.4,0.5,0.6],[0.1,0.2,0.3]], 
                [[0.1,0.2,0.3],[0.4,0.5,0.6],[0.1,0.2,0.3]],
                [[0.1,0.2,0.3],[0.4,0.5,0.6],[0.1,0.2,0.3]]
                ] 
            )
index_list = tf.constant([[0,0],[1,1],[2,2],[0,1],[1,2],[0,2]])
output = tf.gather(text_embeding, index_list, axis=2)