import numpy as np import paddle def shuffle_dim(x, axis, perm=None): """Permute input tensor along aixs given the permutation or randomly. Args: x (Tensor): shape(*, d_{axis}, *), the input tensor. axis (int): the axis to shuffle. perm (list[int], ndarray, optional): a permutation of [0, d_{axis}), the order to reorder the tensor along the `axis`-th dimension, if not provided, randomly shuffle the `axis`-th dimension. Defaults to None. Returns: Tensor: the shuffled tensor, it has the same shape as x does. """ size = x.shape[axis] if perm is not None and len(perm) != size: raise ValueError("length of permutation should equals the input " "tensor's axis-th dimension's size") if perm is not None: perm = np.array(perm) else: perm = np.random.permutation(size) perm = paddle.to_tensor(perm) out = paddle.gather(x, perm, axis) return out