Removed the inversion of the mask

i6_models/parts/mha.py

@@ -45,10 +45,6 @@ def forward(
             value: torch.Tensor,
             key_padding_mask: torch.Tensor):
 
-        if key_padding_mask is not None:
-            inv_sequence_mask = compat.logical_not(key_padding_mask)
-        else:
-            inv_sequence_mask = None
         assert query is value is key, "only supports self attention for now"
 
         batch_dim , num_tokens, embed_dim = query.shape
@@ -70,9 +66,9 @@ def forward(
         dot = torch.matmul(query, key)  # [B, D//H, T, T]
         dot = dot / self.norm
 
-        if inv_sequence_mask is not None:
-            inv_sequence_mask = inv_sequence_mask.view(batch_dim, 1, 1, inv_sequence_mask.size(1))
-            dot = dot.masked_fill(inv_sequence_mask, -float('inf'))
+        if key_padding_mask is not None:
+            key_padding_mask = key_padding_mask.view(batch_dim, 1, 1, key_padding_mask.size(1))
+            dot = dot.masked_fill(key_padding_mask, -float('inf'))
 
         alpha = self.softmax(dot)# [B, D//H, T, T]
         alpha = self.dropout(alpha)