src/calibrator.py

import torch
import torch.nn as nn
import torch.nn.functional as F

from transformers import (
    AutoTokenizer,
    AutoModelForSeq2SeqLM,
    LogitsProcessorList,
    MinLengthLogitsProcessor,
    BeamSearchScorer,
    StoppingCriteriaList,
    MaxLengthCriteria,
    T5ForConditionalGeneration,
    T5Tokenizer
)


class EncoderDecoderCalibrator(nn.Module):

    def __init__(self, model, loss, regularization, beam_size, num_candidates, max_length=16, alpha=0.01):

        super().__init__()
        
        self.model = model 
        
        self.loss = loss
        self.regularization = regularization
        self.alpha = alpha

        assert beam_size >= num_candidates, "num_candidates should be less or equal than beam_size"
        self.beam_size = beam_size
        self.num_candidates = num_candidates
        self.min_length = 0
        self.max_length = max_length

        self.length_penalty = 1.0
        
        self.eos_token_id = self.model.config.eos_token_id
        self.decoder_start_token_id = self.model.config.decoder_start_token_id
        self.pad_token_id = self.model.config.pad_token_id

    def generate_candidates(self, encoder_outputs):

        B, L = encoder_outputs.last_hidden_state.shape[:2]

        beam_scorer = BeamSearchScorer(
            batch_size=B,
            num_beams=self.beam_size,
            device=encoder_outputs.last_hidden_state.device,
            length_penalty=self.length_penalty,
            do_early_stopping=False,
            num_beam_hyps_to_keep=self.num_candidates,
            max_length=self.max_length,
        )

        stopping_criteria = StoppingCriteriaList()

        stopping_criteria.append(
            MaxLengthCriteria(
                max_length=self.max_length,
                max_position_embeddings=self.max_length,
            )
        )

        logits_processor = LogitsProcessorList(
            [
                MinLengthLogitsProcessor(self.min_length, eos_token_id=self.eos_token_id),
            ]
        )

        encoder_outputs.last_hidden_state = encoder_outputs.last_hidden_state.repeat_interleave(self.beam_size, 0)
        
        input_ids = torch.full((B * self.beam_size, 1), self.decoder_start_token_id, device=self.model.device, dtype=torch.long)
        # print(input_ids.shape)
        return self.model.beam_search(
            input_ids,
            beam_scorer,
            logits_processor=logits_processor,
            pad_token_id=self.pad_token_id,
            eos_token_id=self.eos_token_id,
            output_scores=True,
            output_logits=True,
            output_hidden_states=True,
            stopping_criteria=stopping_criteria,
            return_dict_in_generate=True,
            encoder_outputs=encoder_outputs
        )

    def forward(self, input_ids, labels, **kwargs):

        # print(input_ids.shape)
        B, C, L,H = input_ids.shape

        # generate output of encoder
    
        encoder_outputs = self.model.get_encoder()(input_ids, return_dict=True) 

        candidates = self.generate_candidates(encoder_outputs)
        
        sequences = candidates.sequences
        # print(sequences.shape)
        # print(B, self.num_candidates)
        sequences_len = (sequences != 0).sum(-1)
        
        transition_scores = self.model.compute_transition_scores(sequences, candidates.scores, candidates.beam_indices, normalize_logits=False) 

        sequences_scores = transition_scores.sum(-1) / sequences_len

        loss = self.loss(sequences.view(B, self.num_candidates, -1), labels, sequences_scores.view((B, -1)))
        del candidates
        # TODO: investigate if we can use the scores returned by the beam search
        #scores_reg = torch.stack(candidates.scores, dim=1)
        scores_reg = F.log_softmax(self.model(decoder_input_ids=sequences, encoder_outputs=encoder_outputs).logits, dim=-1)
        loss = loss + self.alpha * self.regularization(sequences, scores_reg, labels, encoder_outputs=encoder_outputs)

        return {"loss": loss}

    # def generate(self, input_ids, max_length=None, num_return_sequences=1, **kwargs):
    #     if max_length is None:
    #         max_length = self.max_length

    #     encoder_outputs = self.model.get_encoder()(input_ids, return_dict=True)
    #     print(encoder_outputs)
    #     output_ids = self.model.generate(
    #         encoder_outputs=encoder_outputs,
    #         max_length=max_length,
    #         num_return_sequences=num_return_sequences,
    #         do_sample=True,  # Enable sampling
    #         top_k=50,  # Set the top-k sampling parameter
    #         top_p=0.95,  # Set the top-p (nucleus) sampling parameter
    #         num_beams=4,  # Set the number of beams for beam search
    #         early_stopping=True,  # Enable early stopping
    #         **kwargs
    #     )

    #     return output_ids