global_attention_3d error

[chaoshunh]

it should be

new_shape = tf.concat([tf.shape(input=q)[0:-1],[v.shape[-1]]],0)

instead of

new_shape=tf.concat([tf.shape(input=q)[0:-1],[v.shape[-1].value]],0)

[kkedich]

Yeah, v is a tensor at this stage, so I did not found as well v.shape having a value attribute.