EmbraceRoBERTawithQuery.py

# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch RoBERTa model. """

from __future__ import (absolute_import, division, print_function,
                        unicode_literals)

import logging

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss, MSELoss

from pytorch_transformers.modeling_bert import (BertConfig, BertEmbeddings,
                                                BertLayerNorm, BertModel,
                                                BertPreTrainedModel, gelu)

from pytorch_transformers.modeling_utils import add_start_docstrings
from models.EmbracementLayer import EmbracementLayer
from models.CondensedEmbracementLayer import CondensedEmbracementLayer
from models.AttentionLayer import AttentionLayer
from models.BranchesLayer import BranchesLayer

from models.roberta_utils import (ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
                                  ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING,
                                  RobertaConfig, RobertaModel, RobertaClassificationHead)
# RoBERTa uses the same model as BERT except the embedding layer
from models.SelfAttentionLayer import BertAttentionBertC


logger = logging.getLogger(__name__)


@add_start_docstrings("""RoBERTa Model transformer with a sequence classification/regression head on top (a linear layer 
    on top of the pooled output) e.g. for GLUE tasks. """,
                      ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
class EmbraceRobertaWithQueryForSequenceClassification(BertPreTrainedModel):
    r"""
        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
            Labels for computing the sequence classification/regression loss.
            Indices should be in ``[0, ..., config.num_labels]``.
            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).

    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
            Classification (or regression if config.num_labels==1) loss.
        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
            Classification (or regression if config.num_labels==1) scores (before SoftMax).
        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
            of shape ``(batch_size, sequence_length, hidden_size)``:
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

    Examples::

        tokenizer = RoertaTokenizer.from_pretrained('roberta-base')
        model = RobertaForSequenceClassification.from_pretrained('roberta-base')
        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
        outputs = model(input_ids, labels=labels)
        loss, logits = outputs[:2]

    """
    config_class = RobertaConfig
    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
    base_model_prefix = "roberta"

    def __init__(self, config, dropout_prob, is_condensed=False, add_branches=False,
                 share_branch_weights=False, p='multinomial', max_seq_length=128, extract_key_value_from_bertc=True,
                 dimension_reduction_method='attention', concat_att_with_embracement=False,
                 do_calculate_num_params=False):
        super(EmbraceRobertaWithQueryForSequenceClassification, self).__init__(config)
        self.num_labels = config.num_labels
        self.vocab_size = config.vocab_size
        self.hidden_size = config.hidden_size  # 768
        self.is_condensed = is_condensed
        self.p = p
        self.max_seq_length = max_seq_length  # 128
        self.extract_key_value_from_bertc = extract_key_value_from_bertc
        self.dimension_reduction_method = dimension_reduction_method
        self.concat_att_with_embracement = concat_att_with_embracement
        self.do_calculate_num_params = do_calculate_num_params

        """EmbraceBERT with branches"""
        self.num_labels_evaluator = 2
        self.add_branches = add_branches
        self.share_branch_weights = share_branch_weights
        """END MODIFICATION"""

        self.roberta = RobertaModel(config)
        if not self.is_condensed:
            self.embracement_layer = EmbracementLayer(config, self.hidden_size, self.p, self.max_seq_length)
        else:
            self.embracement_layer = CondensedEmbracementLayer(self.hidden_size, self.p)

        if self.dimension_reduction_method == 'attention':
            self.attention = AttentionLayer(self.hidden_size)
        else:  # projection
            self.projection = nn.Linear(2, 1)
        self.classifier = RobertaClassificationHead(config, dropout_prob)
        # self.classifier = nn.Linear(self.hidden_size, self.num_labels)

        if self.concat_att_with_embracement:  # BERT with proj(T_[CLS], cat(Embrace, att(T_[CLS], T_all)):
            # Att cls layer
            self.embrace_attention = AttentionLayer(self.hidden_size)
            # Projection between 3 inputs: T_[CLS], T_e, T_att
            self.projection = nn.Linear(3, 1)

        # BERTc multihead attention layer
        config.output_attentions = True
        self.multiheadattention = BertAttentionBertC(config, self.hidden_size)

        """EmbraceBERT with branches"""
        if self.add_branches:
            self.branches_layer = BranchesLayer(config, share_branch_weights,
                                                num_labels_evaluator=self.num_labels_evaluator)
        """END MODIFICATION"""

    def forward(self, input_ids, input_bertc=None, token_type_ids=None, attention_mask=None, labels=None,
                position_ids=None, head_mask=None, apply_dropout=False, freeze_bert_weights=False, model_bertc=None,
                is_evaluate=False):

        # Fine-tune with
        if freeze_bert_weights:
            self.roberta.requires_grad = not freeze_bert_weights

        outputs = self.roberta(input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
                               attention_mask=attention_mask, head_mask=head_mask)
        cls_output = outputs[1]  # CLS
        output_tokens_from_bert = outputs[0]

        """EmbraceBERT with branches"""
        if self.add_branches:
            hidden_tokens_from_bert = outputs[2]
            output_tokens_from_bert, attention_mask, logits_branches, logits_branches_evaluator, labels_branch_evaluator = self.branches_layer(
                hidden_tokens_from_bert, output_tokens_from_bert, attention_mask, labels)
        """END MODIFICATION"""

        if self.do_calculate_num_params:  # needed to calcualte the num of params
            output_tokens_from_bert_tmp = self.multiheadattention(output_tokens_from_bert, bert_query=output_tokens_from_bert)
            output_tokens_from_bert_tmp = output_tokens_from_bert_tmp[0]

        # ADD: Multi-head attention layer
        if is_evaluate:  # Q, K, V from BERTi
            # if is_condensed:
            # output_tokens_from_bert = self.multiheadattention(output_tokens_from_bert, head_mask=attention_mask, bert_query=output_tokens_from_bert)
            output_tokens_from_bert = self.multiheadattention(output_tokens_from_bert, bert_query=output_tokens_from_bert)
            output_tokens_from_bert = output_tokens_from_bert[0]
        else:  # Q from BERTc, (K,V) from BERTi
            if model_bertc is not None:
                model_bertc.eval()
                # bertc_output = model_bertc.bert(**input_bertc, apply_dropout=apply_dropout)
                bertc_output = model_bertc.bert(input_bertc['input_ids'], token_type_ids=input_bertc['token_type_ids'], attention_mask=input_bertc['attention_mask'])
                bertc_cls_output = bertc_output[1]  # CLS
                bertc_output_tokens_from_bert = bertc_output[0]

                # if is_condensed:
                # output_tokens_from_bert = self.multiheadattention(output_tokens_from_bert, head_mask=attention_mask, bert_query=bertc_output_tokens_from_bert)
                output_tokens_from_bert = self.multiheadattention(output_tokens_from_bert, bert_query=bertc_output_tokens_from_bert, extract_key_value_from_bertc=self.extract_key_value_from_bertc)
                output_tokens_from_bert = output_tokens_from_bert[0]

        # Embracement layer with attention and no docking
        # sequence_output = outputs[0]
        if self.is_condensed:  # Embracement layer with outputs between CLS and SEP only
            embraced_features_token = self.embracement_layer(output_tokens_from_bert, attention_mask)
        else:  # Embracement layer with all outputs (except CLS)
            embraced_features_token = self.embracement_layer(output_tokens_from_bert, cls_token=cls_output)

        if self.concat_att_with_embracement:  # BERT with proj(T_[CLS], cat(Embrace, att(T_[CLS], T_all)):
            # Att cls layer: 'bertwithattclsprojection'
            att_output = self.embrace_attention(cls_output, output_tokens_from_bert)
            att_output = att_output[0]
            if len(att_output.shape) == 1:
                att_output = att_output.unsqueeze(0).unsqueeze(0)
            elif len(att_output.shape) == 2:
                att_output = att_output.unsqueeze(1)
            if self.dimension_reduction_method == 'attention':
                tokens = torch.cat((att_output, embraced_features_token.unsqueeze(1).cuda()), 1)
                embrace_output = self.attention(cls_output.unsqueeze(1), tokens)
                embrace_output = embrace_output[0]
            else:
                tokens = torch.cat((cls_output.unsqueeze(1), att_output, embraced_features_token.unsqueeze(1).cuda()), 1)
                tokens = tokens.permute((0, 2, 1))
                # Projection layer to obtain 1 feature vector for classification
                embrace_output = self.projection(tokens).squeeze()
        else:
            # Last step: Apply attention layer to CLS and embraced_features_token
            # embrace_output = self.embrace_attention(embraced_cls_with_branches, embraced_features_token)
            if self.dimension_reduction_method == 'attention':
                embrace_output = self.attention(cls_output, embraced_features_token)
                embrace_output = embrace_output[0]
            else:  # projection
                tokens = torch.cat((cls_output.unsqueeze(1), embraced_features_token.cuda().unsqueeze(1)), 1)
                tokens = tokens.permute((0, 2, 1))
                # Projection layer to obtain 1 feature vector for classification
                embrace_output = self.projection(tokens).squeeze()

        # Classify, dropout also applied here, but maybe not needed because the embrace layer functions as
        #   a dropout mechanism?
        s = embrace_output.shape
        if len(s) == 1:  # bs=1
            embrace_output = embrace_output.unsqueeze(0).unsqueeze(0)
        else:  # bs != 1
            embrace_output = embrace_output.unsqueeze(1)
        # logits = self.classifier(embrace_output, apply_dropout)
        if self.do_calculate_num_params:
            # args.do_calculate_num_params:
            logits = self.classifier(embrace_output)
        else:  # Regular run:
            logits = self.classifier(embrace_output, apply_dropout)

        size_logits = logits.shape
        if len(size_logits) == 3:  # Should be (bs, num_class)
            logits = logits.squeeze()

        outputs = (logits,) + outputs[2:]
        if labels is not None:
            if self.num_labels == 1:
                #  We are doing regression
                loss_fct = MSELoss()
                loss = loss_fct(logits.view(-1), labels.view(-1))
            else:
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

            """EmbraceBERT with branches: Add losses from all branches"""
            if self.add_branches:
                loss_branches, loss_branches_evaluator = \
                    self.branches_layer.loss_branches_and_evaluator(loss_fct, logits_branches,
                                                                    logits_branches_evaluator, labels,
                                                                    labels_branch_evaluator)
                loss += loss_branches + loss_branches_evaluator
            """END MODIFICATION"""

            outputs = (loss,) + outputs

        return outputs  # (loss), logits, (hidden_states), (attentions)