def attention_mask_window(attention_states_sequence_len, index, input, name=None):
r"""AttentionMaskWindow
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_window(
attention_states_sequence_len=attention_states_sequence_len, index=index,
input=input, fill_value=-np.finfo(np.float32).max, name=name)
def _AttentionMaskWindowGrad(op, *grad):
attention_mask_grad = gen_attention_mask_ops._attention_mask_window(
attention_states_sequence_len=op.inputs[0], index=op.get_attr("index"),
input=grad[0], fill_value=0.0)
return [None] + [attention_mask_grad]
