tf.dynamic_partition()用法

转自TensorFlow官网

tf.dynamic_partition(
    data,
    partitions,
    num_partitions,
    name=None
)

Partitions data into num_partitions tensors using indices from partitions.

使用分区中的索引将数据分区到num_partitions张量中。 

For each index tuple js of size partitions.ndim, the slice data[js, ...] becomes part of outputs[partitions[js]]. The slices with partitions[js] = i are placed in outputs[i] in lexicographic order of js, and the first dimension of outputs[i] is the number of entries in partitionsequal to i. In detail,

    outputs[i].shape = [sum(partitions == i)] + data.shape[partitions.ndim:]

    outputs[i] = pack([data[js, ...] for js if partitions[js] == i])

For example: 

    # Scalar partitions.
    partitions = 1
    num_partitions = 2
    data = [10, 20]
    outputs[0] = []  # Empty with shape [0, 2]
    outputs[1] = [[10, 20]]

    # Vector partitions.
    partitions = [0, 0, 1, 1, 0]
    num_partitions = 2
    data = [10, 20, 30, 40, 50]
    outputs[0] = [10, 20, 50]
    outputs[1] = [30, 40]