我不知道unsorted_segment_mean的细分ID形状，也找不到其他选择。 在documentation of unsorted_segment_sum (same behavior)中写为：
segment_ids：张量。必须是以下类型之一：int32，int64。张量，其形状是data.shape的前缀。
“形状是data.shape的前缀”是什么意思？ 我试图动态设置形状，但是不起作用：
a, b = 2, 2 # Will be some input variables
x = tf.placeholder(tf.float32)
# Add padding
size = tf.shape(x)
batchSize, inputSize = size[0], size[1]
paddingLength=tf.cast((inputSize/a)%b, dtype=tf.int32)
T_paddingSize = tf.scatter_nd([[1,1]], [paddingLength], [2,2])
x = tf.pad(x, T_paddingSize, 'SYMMETRIC')
# Generate segment IDs
size = tf.shape(x)
batchSize, inputSize = size[0], size[1]
T_segSize = tf.cast(tf.ceil([a/b]), dtype=tf.int32)
T_segments = tf.tile(tf.range(0, a), T_segSize)
# first TRY: ValueError: Cannot convert a partially known TensorShape to a Tensor: (?,)
T_segments = tf.reshape(T_segments, T_segSize*a)
# second TRY: TypeError: Tensor objects are not iterable when eager execution is not enabled.
T_segments = T_segments.set_shape(T_segSize*a)
# Shall compute the mean over unsorted segments
# ValueError: Cannot convert a partially known TensorShape to a Tensor: (?,)
y = tf.matrix_transpose(tf.unsorted_segment_mean(tf.matrix_transpose(x), T_segments, T_segSize[0]))
我该怎么办？ 感谢您的帮助！