forward attention for TF Tacotron2

tf/layers/common_layers.py

@@ -155,6 +155,23 @@ class Attention(keras.layers.Layer):
         score -= 1.e9 * math_ops.cast(padding_mask, dtype=tf.float32)
         return score
 
+    def apply_forward_attention(self, alignment, alpha):
+        # forward attention
+        fwd_shifted_alpha = tf.pad(alpha[:, :-1], ((0, 0), (1, 0)))
+        # compute transition potentials
+        new_alpha = ((1 - 0.5) * alpha + 0.5 * fwd_shifted_alpha + 1e-8) * alignment
+        # renormalize attention weights
+        new_alpha = new_alpha / tf.reduce_sum(new_alpha, axis=1, keepdims=True)
+        return new_alpha
+
+    def update_states(self, old_states, scores_norm, attn_weights, new_alpha=None):
+        states = []
+        if self.use_loc_attn:
+            states = [old_states[0] + scores_norm, attn_weights]
+        if self.use_forward_attn:
+            states.append(new_alpha)
+        return tuple(states)
+
     def call(self, query, states):
         """
         shapes:
@@ -170,13 +187,19 @@ class Attention(keras.layers.Layer):
         # self.apply_score_masking(score, mask)
         # attn_weights shape == (batch_size, max_length, 1)
 
-        attn_weights = self.norm_func(score)
+        # normalize attention scores
+        scores_norm = self.norm_func(score)
+        attn_weights = scores_norm
 
-        # update attention states
-        if self.use_loc_attn:
-            states = (states[0] + attn_weights, attn_weights)
-        else:
-            states = ()
+        # apply forward attention
+        new_alpha = None
+        if self.use_forward_attn:
+            new_alpha = self.apply_forward_attention(attn_weights, states[-1])
+            attn_weights = new_alpha
+
+        # update states tuple
+        # states = (cum_attn_weights, attn_weights, new_alpha)
+        states = self.update_states(states, scores_norm, attn_weights, new_alpha)
 
         # context_vector shape after sum == (batch_size, hidden_size)
         context_vector = tf.matmul(tf.expand_dims(attn_weights, axis=2), self.values, transpose_a=True, transpose_b=False)