def attention_mask(attention_states_sequence_len, input, name=None):
r"""AttentionMask
Args:
attention_states_sequence_len: A `Tensor` of type `int64`.
input: A `Tensor` of type `float32`.
fill_value: A `float`.
name: A name for the operation (optional).
Returns:
A `Tensor` of type `float32`.
"""
return gen_attention_mask_ops._attention_mask(
attention_states_sequence_len=attention_states_sequence_len, input=input,
fill_value=-np.finfo(np.float32).max, name=name)
def _AttentionMaskGrad(op, *grad):
attention_mask_grad = gen_attention_mask_ops._attention_mask(
attention_states_sequence_len=op.inputs[0], input=grad[0], fill_value=0.0)
return [None] + [attention_mask_grad]
