仅乘以张量流数组的某些列

Question

我目前正在修改我的一个对象检测神经网络的损失函数。我基本上有两个数组；

y_true：预测标签。形状为 (x, y, z) 的 tf 张量 y_pred：预测值。形状 (x, y, z) 的 tf 张量 - x 维度是批量大小，y 维度是图像中预测对象的数量，z 维度包含类的 one-hot 编码以及边界所述类的盒子。

现在回到真正的问题：我想要做的基本上是将 y_pred 中的前 5 个 z 值与 y_true 中的前 5 个 z 值相乘。所有其他值应保持不受影响。在 numpy 中，它非常简单；

y_pred[:,:,:5] *= y_true[:,:,:5]

我发现在 tensorflow 中很难做到这一点，因为我无法将值分配给原始张量，并且我想保持所有其他值相同。我该如何在 tensorflow 中执行此操作？

Answer 1

从 v1.1 开始，Tensorflow 涵盖了这种类似 Numpy 的索引，请参阅 Tensor.getitem。

import tensorflow as tf

with tf.Session() as sess:
    y_pred = tf.constant([[[1,2,3,4,5,6,7,8,9,10], [10,20,30,40,50,60,70,80,90,100]]])
    y_true = tf.constant([[[1,2,3,4,5,6,7,8,9,10], [10,20,30,40,50,60,70,80,90,100]]])
    print((y_pred[:,:,:5] * y_true[:,:,:5]).eval()) 
    # [[[   1    4    9   16   25]
    #   [ 100  400  900 1600 2500]]]

评论后编辑：

现在，问题是“*=”部分，即项目分配。这在 Tensorflow 中并不是一个简单的操作。但是，在您的情况下，这可以使用tf.concat 或tf.where 轻松解决（tf.dynamic_partition + tf.dynamic_stitch 可用于更复杂的情况）。

在下面找到前两个解决方案的快速实现。

使用 Tensor.getitem 和 tf.concat 的解决方案：

import tensorflow as tf

with tf.Session() as sess:
    y_pred = tf.constant([[[1,2,3,4,5,6,7,8,9,10]]])
    y_true = tf.constant([[[1,2,3,4,5,6,7,8,9,10]]])

    # tf.where can't apply the condition to any axis (see doc).
    # In your case (condition on 2nd axis), we need either to manually broadcast the
    # condition tensor, or transpose the target tensors.
    # Here is a quick demonstration with the 2nd solution:

    y_pred_edit = y_pred[:,:,:5] * y_true[:,:,:5]
    y_pred_rest = y_pred[:,:,4:]

    y_pred = tf.concat((y_pred_edit, y_pred_rest), axis=2)
    print(y_pred.eval())
    # [[[ 1  4  9 16 25  6  7  8  9 10]]]

使用 tf.where 的解决方案：

import tensorflow as tf

def select_n_fist_indices(n, batch_size):
    """ Return a list of length batch_size with the n first elements True
        and the rest False, i.e. [*[[True] * n], *[[False] * (batch_size - n)]]. 
    """
    n_ones = tf.ones((n))
    rest_zeros = tf.zeros((batch_size - n))
    indices = tf.cast(tf.concat((n_ones, rest_zeros), axis=0), dtype=tf.bool)

    return indices

with tf.Session() as sess:
    y_pred = tf.constant([[[1,2,3,4,5,6,7,8,9,10]]])
    y_true = tf.constant([[[1,2,3,4,5,6,7,8,9,10]]])

    # tf.where can't apply the condition to any axis (see doc).
    # In your case (condition on 2nd axis), we need either to manually broadcast the 
    # condition tensor, or transpose the target tensors.
    # Here is a quick demonstration with the 2nd solution:
    y_pred_tranposed = tf.transpose(y_pred, [2, 0, 1])
    y_true_tranposed = tf.transpose(y_true, [2, 0, 1])

    edit_indices = select_n_fist_indices(5, tf.shape(y_pred_tranposed)[0])

    y_pred_tranposed = tf.where(condition=edit_indices, 
                                x=y_pred_tranposed * y_true_tranposed, y=y_pred_tranposed)

    # Transpose back:  
    y_pred = tf.transpose(y_pred_tranposed, [1, 2, 0])
    print(y_pred.eval())
    # [[[ 1  4  9 16 25  6  7  8  9 10]]]