The implementation of the Attention function is different from the description in DIEN paper.

In `Model_DIEN` the attention weights are calculated by `din_fcn_attention()`, which seems same as `din_attention()` except for:
```
query = tf.layers.dense(query, facts_size, activation=None, name='f1' + stag)
query = prelu(query)
```
And I think this is different from the description in the DIEN paper:

Start with Model_DIEN and compare din_fcn_attention() with din_attention(), focusing on the query dense layer and prelu call shown in the issue. Read the DIEN paper's attention description and determine whether the implementation should match it; done means the behavior and corresponding validation reflect the paper or the discrepancy is documented.

Written by the indexing model from the issue text.