Implementation of the WeCNLP abstract "Cross+Self-Attention for Transformer Models" (#1097)

Summary:
This PR implements a new attention module which combines cross-attention (encoder-decoder attention) and the decoder self-attention. This work was accepted as an abstract at WeCNLP 2019 (https://www.wecnlp.ai/wecnlp-2019).

Cross+Self-Attention reduces the amount of parameter and increases the inference speed without any degradation in translation quality.
More details can be found in the attached [abstract](https://github.com/pytorch/fairseq/files/3561282/paper.pdf)
Pull Request resolved: #1097

Differential Revision: D17653168

Pulled By: myleott

fbshipit-source-id: deb834c2c78a229d7418ffbfea20ba3ce252991c

Author: stephanpeitz

fairseq/models/transformer.py

@@ -122,6 +122,13 @@ def add_args(parser):
                                  'Must be used with adaptive_loss criterion'),
         parser.add_argument('--adaptive-softmax-dropout', type=float, metavar='D',
                             help='sets adaptive softmax dropout for the tail projections')
+        # args for "Cross+Self-Attention for Transformer Models" (Peitz et al., 2019)
+        parser.add_argument('--no-cross-attention', default=False, action='store_true',
+                            help='do not perform cross-attention')
+        parser.add_argument('--cross-self-attention', default=False, action='store_true',
+                            help='perform cross+self-attention')
+        parser.add_argument('--layer-wise-attention', default=False, action='store_true',
+                            help='perform layer-wise attention (cross-attention or cross+self-attention)')
         # fmt: on
 
     @classmethod
@@ -180,7 +187,12 @@ def build_encoder(cls, args, src_dict, embed_tokens):
 
     @classmethod
     def build_decoder(cls, args, tgt_dict, embed_tokens):
-        return TransformerDecoder(args, tgt_dict, embed_tokens)
+        return TransformerDecoder(
+            args,
+            tgt_dict,
+            embed_tokens,
+            no_encoder_attn=getattr(args, 'no_cross_attention', False),
+        )
 
 
 class TransformerEncoder(FairseqEncoder):
@@ -211,6 +223,8 @@ def __init__(self, args, dictionary, embed_tokens):
             learned=args.encoder_learned_pos,
         ) if not args.no_token_positional_embeddings else None
 
+        self.layer_wise_attention = getattr(args, 'layer_wise_attention', False)
+
         self.layers = nn.ModuleList([])
         self.layers.extend([
             TransformerEncoderLayer(args)
@@ -230,21 +244,29 @@ def forward_embedding(self, src_tokens):
         x = F.dropout(x, p=self.dropout, training=self.training)
         return x, embed
 
-    def forward(self, src_tokens, src_lengths, cls_input=None):
+    def forward(self, src_tokens, src_lengths, cls_input=None, return_all_hiddens=False):
         """
         Args:
             src_tokens (LongTensor): tokens in the source language of shape
                 `(batch, src_len)`
             src_lengths (torch.LongTensor): lengths of each source sentence of
                 shape `(batch)`
+            return_all_hiddens (bool, optional): also return all of the
+                intermediate hidden states (default: False).
 
         Returns:
             dict:
                 - **encoder_out** (Tensor): the last encoder layer's output of
                   shape `(src_len, batch, embed_dim)`
                 - **encoder_padding_mask** (ByteTensor): the positions of
                   padding elements of shape `(batch, src_len)`
+                - **encoder_states** (List[Tensor]): all intermediate
+                  hidden states of shape `(src_len, batch, embed_dim)`.
+                  Only populated if *return_all_hiddens* is True.
         """
+        if self.layer_wise_attention:
+            return_all_hiddens = True
+
         x, encoder_embedding = self.forward_embedding(src_tokens)
 
         # B x T x C -> T x B x C
@@ -255,17 +277,24 @@ def forward(self, src_tokens, src_lengths, cls_input=None):
         if not encoder_padding_mask.any():
             encoder_padding_mask = None
 
+        encoder_states = [] if return_all_hiddens else None
+
         # encoder layers
         for layer in self.layers:
             x = layer(x, encoder_padding_mask)
+            if return_all_hiddens:
+                encoder_states.append(x)
 
         if self.layer_norm:
             x = self.layer_norm(x)
+            if return_all_hiddens:
+                encoder_states[-1] = x
 
         return {
             'encoder_out': x,  # T x B x C
             'encoder_padding_mask': encoder_padding_mask,  # B x T
             'encoder_embedding': encoder_embedding,  # B x T x C
+            'encoder_states': encoder_states,  # List[T x B x C]
         }
 
     def reorder_encoder_out(self, encoder_out, new_order):
@@ -285,6 +314,9 @@ def reorder_encoder_out(self, encoder_out, new_order):
         if encoder_out['encoder_padding_mask'] is not None:
             encoder_out['encoder_padding_mask'] = \
                 encoder_out['encoder_padding_mask'].index_select(0, new_order)
+        if encoder_out.get('encoder_states', None) is not None:
+            for idx, state in enumerate(encoder_out['encoder_states']):
+                encoder_out['encoder_states'][idx] = state.index_select(1, new_order)
         return encoder_out
 
     def max_positions(self):
@@ -293,6 +325,14 @@ def max_positions(self):
             return self.max_source_positions
         return min(self.max_source_positions, self.embed_positions.max_positions())
 
+    def buffered_future_mask(self, tensor):
+        dim = tensor.size(0)
+        if not hasattr(self, '_future_mask') or self._future_mask is None or self._future_mask.device != tensor.device:
+            self._future_mask = torch.triu(utils.fill_with_neg_inf(tensor.new(dim, dim)), 1)
+            if self._future_mask.size(0) < dim:
+                self._future_mask = torch.triu(utils.fill_with_neg_inf(self._future_mask.resize_(dim, dim)), 1)
+        return self._future_mask[:dim, :dim]
+
     def upgrade_state_dict_named(self, state_dict, name):
         """Upgrade a (possibly old) state dict for new versions of fairseq."""
         if isinstance(self.embed_positions, SinusoidalPositionalEmbedding):
@@ -350,6 +390,9 @@ def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False):
             learned=args.decoder_learned_pos,
         ) if not args.no_token_positional_embeddings else None
 
+        self.cross_self_attention = getattr(args, 'cross_self_attention', False)
+        self.layer_wise_attention = getattr(args, 'layer_wise_attention', False)
+
         self.layers = nn.ModuleList([])
         self.layers.extend([
             TransformerDecoderLayer(args, no_encoder_attn)
@@ -435,14 +478,26 @@ def extract_features(self, prev_output_tokens, encoder_out=None, incremental_sta
 
         inner_states = [x]
 
+        self_attn_padding_mask = prev_output_tokens.eq(self.padding_idx)
+        if not self_attn_padding_mask.any() and not self.cross_self_attention:
+            self_attn_padding_mask = None
+
         # decoder layers
-        for layer in self.layers:
+        for idx, layer in enumerate(self.layers):
+            encoder_state = None
+            if encoder_out is not None:
+                if self.layer_wise_attention:
+                    encoder_state = encoder_out['encoder_states'][idx]
+                else:
+                    encoder_state = encoder_out['encoder_out']
+
             x, attn = layer(
                 x,
-                encoder_out['encoder_out'] if encoder_out is not None else None,
+                encoder_state,
                 encoder_out['encoder_padding_mask'] if encoder_out is not None else None,
                 incremental_state,
                 self_attn_mask=self.buffered_future_mask(x) if incremental_state is None else None,
+                self_attn_padding_mask=self_attn_padding_mask,
             )
             inner_states.append(x)
 
@@ -553,6 +608,9 @@ def base_architecture(args):
     args.share_all_embeddings = getattr(args, 'share_all_embeddings', False)
     args.no_token_positional_embeddings = getattr(args, 'no_token_positional_embeddings', False)
     args.adaptive_input = getattr(args, 'adaptive_input', False)
+    args.no_cross_attention = getattr(args, 'no_cross_attention', False)
+    args.cross_self_attention = getattr(args, 'cross_self_attention', False)
+    args.layer_wise_attention = getattr(args, 'layer_wise_attention', False)
 
     args.decoder_output_dim = getattr(args, 'decoder_output_dim', args.decoder_embed_dim)
     args.decoder_input_dim = getattr(args, 'decoder_input_dim', args.decoder_embed_dim)

fairseq/modules/multihead_attention.py

@@ -186,8 +186,15 @@ def forward(self, query, key, value, key_padding_mask=None, incremental_state=No
                     v = prev_value
                 else:
                     v = torch.cat((prev_value, v), dim=1)
+            if 'prev_key_padding_mask' in saved_state and saved_state['prev_key_padding_mask'] is not None:
+                prev_key_padding_mask = saved_state['prev_key_padding_mask']
+                if static_kv:
+                    key_padding_mask = prev_key_padding_mask
+                else:
+                    key_padding_mask = torch.cat((prev_key_padding_mask, key_padding_mask), dim=1)
             saved_state['prev_key'] = k.view(bsz, self.num_heads, -1, self.head_dim)
             saved_state['prev_value'] = v.view(bsz, self.num_heads, -1, self.head_dim)
+            saved_state['prev_key_padding_mask'] = key_padding_mask
 
             self._set_input_buffer(incremental_state, saved_state)
 
@@ -311,7 +318,8 @@ def reorder_incremental_state(self, incremental_state, new_order):
         input_buffer = self._get_input_buffer(incremental_state)
         if input_buffer is not None:
             for k in input_buffer.keys():
-                input_buffer[k] = input_buffer[k].index_select(0, new_order)
+                if input_buffer[k] is not None:
+                    input_buffer[k] = input_buffer[k].index_select(0, new_order)
             self._set_input_buffer(incremental_state, input_buffer)
 
     def _get_input_buffer(self, incremental_state):

fairseq/modules/transformer_layer.py

@@ -3,6 +3,7 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
+import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from fairseq import utils
@@ -134,13 +135,14 @@ class TransformerDecoderLayer(nn.Module):
     def __init__(self, args, no_encoder_attn=False, add_bias_kv=False, add_zero_attn=False):
         super().__init__()
         self.embed_dim = args.decoder_embed_dim
+        self.cross_self_attention = getattr(args, 'cross_self_attention', False)
         self.self_attn = MultiheadAttention(
             embed_dim=self.embed_dim,
             num_heads=args.decoder_attention_heads,
             dropout=args.attention_dropout,
             add_bias_kv=add_bias_kv,
             add_zero_attn=add_zero_attn,
-            self_attention=True
+            self_attention=not self.cross_self_attention,
         )
         self.dropout = args.dropout
         self.activation_fn = utils.get_activation_fn(
@@ -208,13 +210,27 @@ def forward(
         if prev_self_attn_state is not None:
             if incremental_state is None:
                 incremental_state = {}
-            prev_key, prev_value = prev_self_attn_state
+            prev_key, prev_value = prev_self_attn_state[:2]
             saved_state = {"prev_key": prev_key, "prev_value": prev_value}
+            if len(prev_self_attn_state) >= 3:
+                saved_state["prev_key_padding_mask"] = prev_self_attn_state[2]
             self.self_attn._set_input_buffer(incremental_state, saved_state)
+
+        if self.cross_self_attention and not (incremental_state is not None and "prev_key" in self.self_attn._get_input_buffer(incremental_state)):
+            if self_attn_mask is not None:
+                self_attn_mask = torch.cat((x.new(x.size(0), encoder_out.size(0)).zero_(), self_attn_mask), dim=1)
+            if self_attn_padding_mask is not None:
+                if encoder_padding_mask is None:
+                    encoder_padding_mask = self_attn_padding_mask.new(encoder_out.size(1), encoder_out.size(0)).zero_()
+                self_attn_padding_mask = torch.cat((encoder_padding_mask, self_attn_padding_mask), dim=1)
+            y = torch.cat((encoder_out, x), dim=0)
+        else:
+            y = x
+
         x, attn = self.self_attn(
             query=x,
-            key=x,
-            value=x,
+            key=y,
+            value=y,
             key_padding_mask=self_attn_padding_mask,
             incremental_state=incremental_state,
             need_weights=False,
@@ -230,9 +246,12 @@ def forward(
             if prev_attn_state is not None:
                 if incremental_state is None:
                     incremental_state = {}
-                prev_key, prev_value = prev_attn_state
+                prev_key, prev_value = prev_attn_state[:2]
                 saved_state = {"prev_key": prev_key, "prev_value": prev_value}
+                if len(prev_attn_state) >= 3:
+                    saved_state["prev_key_padding_mask"] = prev_attn_state[2]
                 self.encoder_attn._set_input_buffer(incremental_state, saved_state)
+
             x, attn = self.encoder_attn(
                 query=x,
                 key=encoder_out,
@@ -256,7 +275,10 @@ def forward(
         x = self.maybe_layer_norm(self.final_layer_norm, x, after=True)
         if self.onnx_trace and incremental_state is not None:
             saved_state = self.self_attn._get_input_buffer(incremental_state)
-            self_attn_state = saved_state["prev_key"], saved_state["prev_value"]
+            if self_attn_padding_mask is not None:
+                self_attn_state = saved_state["prev_key"], saved_state["prev_value"], saved_state["prev_key_padding_mask"]
+            else:
+                self_attn_state = saved_state["prev_key"], saved_state["prev_value"]
             return x, attn, self_attn_state
         return x, attn
 

tests/test_binaries.py

@@ -154,6 +154,23 @@ def test_transformer(self):
                 ], run_validation=True)
                 generate_main(data_dir)
 
+    def test_transformer_cross_self_attention(self):
+        with contextlib.redirect_stdout(StringIO()):
+            with tempfile.TemporaryDirectory('test_transformer_cross_self_attention') as data_dir:
+                create_dummy_data(data_dir)
+                preprocess_translation_data(data_dir)
+                train_translation_model(data_dir, 'transformer_iwslt_de_en', [
+                    '--encoder-layers', '2',
+                    '--decoder-layers', '2',
+                    '--encoder-embed-dim', '8',
+                    '--decoder-embed-dim', '8',
+                    '--decoder-embed-dim', '8',
+                    '--no-cross-attention',
+                    '--cross-self-attention',
+                    '--layer-wise-attention',
+                ], run_validation=True)
+                generate_main(data_dir, extra_flags=[])
+
     def test_lightconv(self):
         with contextlib.redirect_stdout(StringIO()):
             with tempfile.TemporaryDirectory('test_lightconv') as data_dir:
@@ -543,6 +560,10 @@ def train_translation_model(data_dir, arch, extra_flags=None, task='translation'
 
 
 def generate_main(data_dir, extra_flags=None):
+    if extra_flags is None:
+        extra_flags = [
+            '--print-alignment',
+        ]
     generate_parser = options.get_generation_parser()
     generate_args = options.parse_args_and_arch(
         generate_parser,
@@ -554,7 +575,6 @@ def generate_main(data_dir, extra_flags=None):
             '--max-len-b', '5',
             '--gen-subset', 'valid',
             '--no-progress-bar',
-            '--print-alignment',
         ] + (extra_flags or []),
     )