fix

liuzh47 · Jul 29, 2020 · 995e5d7 · 995e5d7
1 parent 9623240
commit 995e5d7
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/src/gluonnlp/models/transformer.py b/src/gluonnlp/models/transformer.py
@@ -673,7 +673,7 @@ def incremental_decode(self, F, data, states, mem, mem_valid_length, mem_attn_ma
         step_value = F.npx.reshape(step_value, (-2, -2, self._num_heads, -1))
         new_key = F.np.concatenate([prev_key, step_key], axis=time_axis)
         new_value = F.np.concatenate([prev_value, step_value], axis=time_axis)
-        out, _ = self.self_attention(step_query, new_key, new_value, None)
+        out, [_, attn_weight] = self.self_attention(step_query, new_key, new_value, None)
         out = self.proj_in(out)
         out = self.dropout_layer(out)
         out = out + data
@@ -1209,7 +1209,7 @@ def decode_seq(self, F, tgt_data, tgt_valid_length, mem_data, mem_valid_length):
                 tgt_data = tgt_data + F.np.expand_dims(self.tgt_pos_embed_layer(
                     F.npx.arange_like(tgt_data, axis=0)), axis=1)
         if self.layernorm_embedding:
-            tgt_data = self.src_embed_ln(tgt_data)
+            tgt_data = self.tgt_embed_ln(tgt_data)
         dec_out = self.decoder(tgt_data, tgt_valid_length, mem_data, mem_valid_length)
         return dec_out