sagcopy.py

'''
Copied over from https://huggingface.co/transformers/_modules/transformers/models/bart/modeling_bart.html
'''
from typing import Optional, Tuple
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import BartForConditionalGeneration, BartModel, BartConfig
from transformers.models.bart.modeling_bart import BartDecoder, BartDecoderLayer, BartAttention, _expand_mask
from transformers.modeling_outputs import Seq2SeqLMOutput,  BaseModelOutputWithPastAndCrossAttentions, Seq2SeqModelOutput, BaseModelOutput
class SAGCopyBartForConditionalGeneration(BartForConditionalGeneration):
    
    def __init__(self, config: BartConfig):
        super().__init__(config)
        self.model = SAGCopyBartModel(config)

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        encoder_outputs=None,
        past_key_values=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Labels for computing the masked language modeling loss. Indices should either be in ``[0, ...,
            config.vocab_size]`` or -100 (see ``input_ids`` docstring). Tokens with indices set to ``-100`` are ignored
            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``.

        Returns:
        """
        assert output_attentions or self.model.config.output_attentions, "output_attentions must be true"
        
        # original outputs
        outputs = self.model(input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            encoder_outputs=encoder_outputs,
            decoder_attention_mask=decoder_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,)
        
        if input_ids is None:
            input_ids = self._cache_input_ids
        
        encoder_last_hidden_state = outputs.encoder_last_hidden_state # (batch, seq, hidden)
        decoder_last_hidden_state = outputs[0] #(batch, decoding_seq, hidden )
        
        
        # compute lm logits based on attention
        last_cross_attentions = outputs.cross_attentions[-1] # (batch_size, num_heads, decoding_seq_length, encoding_seq_length).
        
        
        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias  #(batch_size, decoding_seq_length, emb_dim)
        cross_attentions_aggregate = last_cross_attentions.mean(dim=1) #(batch_size, decoding_seq_length, encoding_seq_length)
        dummy_input_ids = input_ids.unsqueeze(-1).expand(-1, -1, lm_logits.size(1)).transpose(1,2) # (batch, decoding_seq_length, encoding_seq_length)
        copy_logits = torch.zeros_like(lm_logits) # (batch, decoding_seq_length, emb_dim)
        copy_logits.scatter_add_(dim=2, index=dummy_input_ids, src=cross_attentions_aggregate)   
        
        
        p_gen = torch.bmm(decoder_last_hidden_state, encoder_last_hidden_state.mean(dim=1).unsqueeze(dim=-1)) # (batch, decoding_seq, 1)
        p_gen = torch.sigmoid(p_gen)


        lm_logits = F.softmax(lm_logits, dim=-1) * p_gen + copy_logits * (1 - p_gen)#(batch_size, decoding_seq_length, emb_dim)
        
        


        masked_lm_loss = None
        if labels is not None:
            # compute loss mask and fill -100 with 0
            loss_mask = labels != -100
            labels.masked_fill_(~loss_mask, 0)
            # use negative log likelihood
            gold_probs = torch.gather(lm_logits, 2, labels.unsqueeze(2)).squeeze(2)
            eps = 1e-7 # for safe log
            masked_lm_loss = - torch.log(gold_probs + eps) * self._loss_weight[labels]
            masked_lm_loss = (masked_lm_loss * loss_mask).mean()


        if not return_dict:
            output = (lm_logits,) + outputs[1:]
            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output

        
        return Seq2SeqLMOutput(
            loss=masked_lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )


class SAGCopyBartModel(BartModel):
    def __init__(self, config: BartConfig):
        super().__init__(config)
        self.decoder = SAGCopyBartDecoder(config, self.shared)
        
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        decoder_input_ids=None,
        decoder_attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        encoder_outputs=None,
        past_key_values=None,
        inputs_embeds=None,
        decoder_inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):

        # different to other models, Bart automatically creates decoder_input_ids from
        # input_ids if no decoder_input_ids are provided
        if decoder_input_ids is None and decoder_inputs_embeds is None:
            decoder_input_ids = shift_tokens_right(
                input_ids, self.config.pad_token_id, self.config.decoder_start_token_id
            )

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        
        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                head_mask=head_mask,
                inputs_embeds=inputs_embeds,
                output_attentions=True,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        
        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )
        
        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_outputs[0],
            encoder_attention_mask=attention_mask,
            head_mask=decoder_head_mask,
            encoder_head_mask=head_mask,
            past_key_values=past_key_values,
            inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            encoder_attentions=encoder_outputs[2]
        )

        if not return_dict:
            return decoder_outputs + encoder_outputs

        return Seq2SeqModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )    

class SAGCopyBartDecoder(BartDecoder):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a :class:`BartDecoderLayer`

    Args:
        config: BartConfig
        embed_tokens (torch.nn.Embedding): output embedding
    """

    def __init__(self, config: BartConfig, embed_tokens: Optional[nn.Embedding] = None):
        super().__init__(config)
        self.layers = nn.ModuleList([BartDecoderLayer(config) for _ in range(config.decoder_layers-1)] + [SAGCopyBARTDecoderLayer(config)])

    

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
        head_mask=None,
        encoder_head_mask=None,
        past_key_values=None,
        inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
        encoder_attentions=None
    ):
        r"""
        Args:
            input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

                Indices can be obtained using :class:`~transformers.BartTokenizer`. See
                :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__`
                for details.

                `What are input IDs? <../glossary.html#input-ids>`__
            attention_mask (:obj:`torch.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
                Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                `What are attention masks? <../glossary.html#attention-mask>`__
            encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, encoder_sequence_length, hidden_size)`, `optional`):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
            encoder_attention_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, encoder_sequence_length)`, `optional`):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
                selected in ``[0, 1]``:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

                `What are attention masks? <../glossary.html#attention-mask>`__
            head_mask (:obj:`torch.Tensor` of shape :obj:`(num_layers, num_heads)`, `optional`):
                Mask to nullify selected heads of the attention modules. Mask values selected in ``[0, 1]``:

                - 1 indicates the head is **not masked**,
                - 0 indicates the heas is **masked**.

            encoder_head_mask (:obj:`torch.Tensor` of shape :obj:`(num_layers, num_heads)`, `optional`):
                Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention
                on hidden heads. Mask values selected in ``[0, 1]``:

                - 1 indicates the head is **not masked**,
                - 0 indicates the heas is **masked**.

            past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers` with each tuple having 2 tuples each of which has 2 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
                Contains precomputed key and value hidden-states of the attention blocks. Can be used to speed up
                decoding.

                If :obj:`past_key_values` are used, the user can optionally input only the last
                :obj:`decoder_input_ids` (those that don't have their past key value states given to this model) of
                shape :obj:`(batch_size, 1)` instead of all :obj:`decoder_input_ids`` of shape :obj:`(batch_size,
                sequence_length)`.
            inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
                Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded
                representation. This is useful if you want more control over how to convert :obj:`input_ids` indices
                into associated vectors than the model's internal embedding lookup matrix.
            output_attentions (:obj:`bool`, `optional`):
                Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under
                returned tensors for more detail.
            output_hidden_states (:obj:`bool`, `optional`):
                Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors
                for more detail.
            return_dict (:obj:`bool`, `optional`):
                Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        # past_key_values_length
        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale

        attention_mask = self._prepare_decoder_attention_mask(
            attention_mask, input_shape, inputs_embeds, past_key_values_length
        )

        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])

        # embed positions
        positions = self.embed_positions(input_shape, past_key_values_length)

        hidden_states = inputs_embeds + positions
        hidden_states = self.layernorm_embedding(hidden_states)

        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
        next_decoder_cache = () if use_cache else None

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
            assert head_mask.size()[0] == (
                len(self.layers)
            ), f"The head_mask should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):
                continue

            past_key_value = past_key_values[idx] if past_key_values is not None else None

            if getattr(self.config, "gradient_checkpointing", False) and self.training:

                if use_cache:
                    logger.warn(
                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
                        "`use_cache=False`..."
                    )
                    use_cache = False

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        # None for past_key_value
                        return module(*inputs, output_attentions, use_cache)
                    return custom_forward
                def create_custom_forward_last(module):
                    def custom_forward_last(*inputs):
                        # None for past_key_value
                        return module(*inputs, output_attentions, use_cache, encoder_attentions)
                    return custom_forward_last

                
                if idx != len(self.layers) -1:
                    layer_outputs = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(decoder_layer),
                        hidden_states,
                        attention_mask,
                        encoder_hidden_states,
                        encoder_attention_mask,
                        head_mask[idx] if head_mask is not None else None,
                        encoder_head_mask[idx] if encoder_head_mask is not None else None,
                        None,
                    )
                else:
                    
                    layer_outputs = torch.utils.checkpoint.checkpoint(
                        create_custom_forward_last(decoder_layer),
                        hidden_states,
                        attention_mask,
                        encoder_hidden_states,
                        encoder_attention_mask,
                        head_mask[idx] if head_mask is not None else None,
                        encoder_head_mask[idx] if encoder_head_mask is not None else None,
                        None,
                    )

            else:
                
                if idx != len(self.layers) -1:
                    layer_outputs = decoder_layer(
                        hidden_states,
                        attention_mask=attention_mask,
                        encoder_hidden_states=encoder_hidden_states,
                        encoder_attention_mask=encoder_attention_mask,
                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                        encoder_layer_head_mask=(encoder_head_mask[idx] if encoder_head_mask is not None else None),
                        past_key_value=past_key_value,
                        output_attentions=output_attentions,
                        use_cache=use_cache,
                    )
                else:
                    
                    layer_outputs = decoder_layer(
                        hidden_states,
                        attention_mask=attention_mask,
                        encoder_hidden_states=encoder_hidden_states,
                        encoder_attention_mask=encoder_attention_mask,
                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                        encoder_layer_head_mask=(encoder_head_mask[idx] if encoder_head_mask is not None else None),
                        past_key_value=past_key_value,
                        output_attentions=output_attentions,
                        use_cache=use_cache,
                        encoder_attentions=encoder_attentions,
                    )
            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

                if encoder_hidden_states is not None:
                    all_cross_attentions += (layer_outputs[2],)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None
        
        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )

class SAGCopyBARTDecoderLayer(BartDecoderLayer):
    def __init__(self, config: BartConfig):
        super().__init__(config)

        self.encoder_attn = SAGCopyBartAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        encoder_layer_head_mask: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
        encoder_attentions: torch.Tensor = None,
    ):
        """
        Args:
            hidden_states (:obj:`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
            attention_mask (:obj:`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            encoder_hidden_states (:obj:`torch.FloatTensor`): cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
            encoder_attention_mask (:obj:`torch.FloatTensor`): encoder attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (:obj:`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(config.encoder_attention_heads,)`.
            encoder_layer_head_mask (:obj:`torch.FloatTensor`): mask for encoder attention heads in a given layer of
                size `(config.encoder_attention_heads,)`.
            past_key_value (:obj:`Tuple(torch.FloatTensor)`): cached past key and value projection states
            output_attentions (:obj:`bool`, `optional`):
                Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under
                returned tensors for more detail.
        """
        residual = hidden_states

        # Self Attention
        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
        # add present self-attn cache to positions 1,2 of present_key_value tuple
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=self_attn_past_key_value,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Cross-Attention Block
        cross_attn_present_key_value = None
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states

            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                layer_head_mask=encoder_layer_head_mask,
                past_key_value=cross_attn_past_key_value,
                output_attentions=output_attentions,
                encoder_attentions=encoder_attentions
            )
            hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
            hidden_states = residual + hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # add cross-attn to positions 3,4 of present_key_value tuple
            present_key_value = present_key_value + cross_attn_present_key_value

        # Fully Connected
        residual = hidden_states
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = F.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = F.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.final_layer_norm(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        if use_cache:
            outputs += (present_key_value,)

        return outputs

class SAGCopyBartAttention(BartAttention):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
    ):
        super().__init__(embed_dim, num_heads, dropout, is_decoder, bias)

        # the same w_p as in the paper
        self.w_p = torch.nn.Linear(1, embed_dim, bias=False)

    def compute_centrality_representations(self, encoder_attentions, iterations=0):
        '''
        encoder_attentions: (batch_size, num_heads, seq_len, seq_len)
        '''
        
        encoder_attentions = encoder_attentions[-1].mean(1) #(batch, seq_len, seq_len)
        T = encoder_attentions / encoder_attentions[-1].sum(1).unsqueeze(1) #(batch, seq_len, seq_len)
        scores = T.sum(1) #(batch, seq_len)

        # do TextRank update
        for _ in range(iterations):
            scores = torch.bmm(T, scores.unsqueeze(2)).squeeze(2)
        scores = scores.unsqueeze(2)
        representations = self.w_p(scores)
        return representations

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        encoder_attentions: torch.Tensor = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None
        bsz, tgt_len, embed_dim = hidden_states.size()

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions plus (Whhi + wpscorei) 
            key_states = self._shape(self.k_proj(key_value_states) + self.compute_centrality_representations(encoder_attentions, iterations=3), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        if self.is_decoder:
            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
            # Further calls to cross_attention layer can then reuse all cross-attention
            # key/value_states (first "if" case)
            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
            # all previous decoder key/value_states. Further calls to uni-directional self-attention
            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
            # if encoder bi-directional self-attention `past_key_value` is always `None`
            past_key_value = (key_states, value_states)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
        key_states = key_states.view(*proj_shape)
        value_states = value_states.view(*proj_shape)

        src_len = key_states.size(1)
        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))

        assert attn_weights.size() == (
            bsz * self.num_heads,
            tgt_len,
            src_len,
        ), f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is {attn_weights.size()}"

        if attention_mask is not None:
            assert attention_mask.size() == (
                bsz,
                1,
                tgt_len,
                src_len,
            ), f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        attn_weights = F.softmax(attn_weights, dim=-1)

        if layer_head_mask is not None:
            assert layer_head_mask.size() == (
                self.num_heads,
            ), f"Head mask for a single layer should be of size {(self.num_heads,)}, but is {layer_head_mask.size()}"
            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

        if output_attentions:
            # this operation is a bit akward, but it's required to
            # make sure that attn_weights keeps its gradient.
            # In order to do so, attn_weights have to reshaped
            # twice and have to be reused in the following
            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
        else:
            attn_weights_reshaped = None

        attn_probs = F.dropout(attn_weights, p=self.dropout, training=self.training)

        attn_output = torch.bmm(attn_probs, value_states)

        assert attn_output.size() == (
            bsz * self.num_heads,
            tgt_len,
            self.head_dim,
        ), f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is {attn_output.size()}"

        attn_output = (
            attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
            .transpose(1, 2)
            .reshape(bsz, tgt_len, embed_dim)
        )

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value