[Chainer] Mask sequence with length

sequence_mask.py

def sequence_mask(x, length, value=0.):
    xp = cuda.get_array_module(length.data)
    # create permutation on (length.ndim + 1) dimension and expand dims until it has shame rank as x
    perms = xp.arange(x.shape[length.ndim]).reshape(
        [1] * length.ndim + [-1] + [1] * (x.ndim - length.ndim -1))
    length = length.reshape([1] * (length.ndim - 1) + [-1] + [1] * (x.ndim - length.ndim))
    pad = xp.ones_like(x) * value
    mask = xp.broadcast_to(perms, x.shape) < length
    return F.where(mask, x, pad)