Adding transpose to att_output before concatenate all attention heads

Transformer_Implementation_Tutorial.ipynb

@@ -310,6 +310,7 @@
     "        att_output = torch.matmul(att_score, V_state)\n",
     "    \n",
     "        # concatinate all attention heads\n",
+    "        att_output = att_output.transpose(1, 2)\n",
     "        att_output = att_output.contiguous().view(batch_size, seq_len, self.num_head*self.d_head) \n",
     "    \n",
     "        # final linear transformation to the concatenated output\n",
@@ -792,7 +793,6 @@
     "        self.num_head = num_head\n",
     "        self.dropout = dropout\n",
     "        self.bias = bias\n",
-    "\n",
     "        \n",
     "        # Encoder stack\n",
     "        self.encoder_stack = nn.ModuleList([ TransformerEncoder(\n",
@@ -820,7 +820,6 @@
     "            encoder_output = encoder(embed_input = encoder_output, padding_mask = padding_mask)\n",
     "            decoder_output = decoder(embed_input = decoder_output, cross_input =encoder_output, padding_mask=padding_mask)\n",
     "       \n",
-    "        \n",
     "        return decoder_output"
    ]
   },