modelling_cxrrg.py

import functools
import os
from typing import Optional, Tuple, Union

import torch
import transformers
from torch.nn import CrossEntropyLoss, Linear
from transformers import PreTrainedTokenizerFast, VisionEncoderDecoderModel
from transformers.configuration_utils import PretrainedConfig
from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
from transformers.modeling_utils import PreTrainedModel
from transformers.models.vision_encoder_decoder.configuration_vision_encoder_decoder import (
    VisionEncoderDecoderConfig,
)
from transformers.utils import logging

from .modelling_uniformer import MultiUniFormerWithProjectionHead

logger = logging.get_logger(__name__)


class CXRRGModel(VisionEncoderDecoderModel):

    config_class = VisionEncoderDecoderConfig
    base_model_prefix = "vision_encoder_decoder"
    main_input_name = "pixel_values"
    supports_gradient_checkpointing = True

    def __init__(        
        self,
        config: Optional[PretrainedConfig] = None,
        encoder: Optional[PreTrainedModel] = None,
        decoder: Optional[PreTrainedModel] = None,
        DefaultEncoderClass = MultiUniFormerWithProjectionHead,
        DefaultDecoderClass = transformers.LlamaForCausalLM,
    ):

        if decoder:
            assert not decoder.config.add_cross_attention, '"add_cross_attention" must be False for the given decoder'
            assert decoder.config.is_decoder, '"is_decoder" must be True for the given decoder'

        if config is None and (encoder is None or decoder is None):
            raise ValueError("Either a configuration or an encoder and a decoder has to be provided.")
        if config is None:
            config = VisionEncoderDecoderConfig.from_encoder_decoder_configs(encoder.config, decoder.config)
        else:
            if not isinstance(config, self.config_class):
                raise ValueError(f"Config: {config} has to be of type {self.config_class}")

        config.tie_word_embeddings = False

        # Initialize with config:
        PreTrainedModel.__init__(self, config)

        # Encoder:
        if encoder is None:
            encoder = DefaultEncoderClass(config=config.encoder)

        # Decoder:
        if decoder is None:
            assert not config.decoder.add_cross_attention
            decoder = DefaultDecoderClass(config=config.decoder)

        self.encoder = encoder
        self.decoder = decoder

        if self.encoder.config.to_dict() != self.config.encoder.to_dict():
            logger.warning(
                f"Config of the encoder: {self.encoder.__class__} is overwritten by shared encoder config:"
                f" {self.config.encoder}"
            )
        if self.decoder.config.to_dict() != self.config.decoder.to_dict():
            logger.warning(
                f"Config of the decoder: {self.decoder.__class__} is overwritten by shared decoder config:"
                f" {self.config.decoder}"
            )

        self.encoder.config = self.config.encoder
        self.decoder.config = self.config.decoder

        assert config.decoder.is_decoder
        assert 'img_token_id' in self.decoder.config.__dict__
        assert 'pad_token_id' in self.decoder.config.__dict__
        assert 'token_type_embeddings' in self.decoder.config.__dict__

        if self.decoder.config.token_type_embeddings == 'add':
            self.token_type_embeddings = torch.nn.Embedding(self.decoder.config.num_token_types, self.decoder.config.hidden_size)

    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.FloatTensor] = None,
        decoder_token_type_ids: Optional[torch.LongTensor] = None,
        encoder_outputs: Optional[Tuple[torch.FloatTensor]] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_position_ids: Optional[torch.LongTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs,
    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        kwargs_encoder = {argument: value for argument, value in kwargs.items() if not argument.startswith("decoder_")}

        kwargs_decoder = {
            argument[len("decoder_") :]: value for argument, value in kwargs.items() if argument.startswith("decoder_")
        }

        if decoder_inputs_embeds is None:
            decoder_inputs_embeds = self.decoder.get_input_embeddings()(decoder_input_ids)

        if encoder_outputs is None:  # Ths is for when generate() is not called; for generation, see prepare_inputs_for_generation():
            if pixel_values is None:
                raise ValueError("You have to specify pixel_values")

            encoder_outputs = self.encoder(
                pixel_values,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
                **kwargs_encoder,
            )  # UniFormer does not support output_attentions.

            assert decoder_inputs_embeds is not None
            decoder_inputs_embeds = torch.cat([encoder_outputs[0], decoder_inputs_embeds], dim=1)

            # Add image token type identifiers:
            decoder_token_type_ids = torch.cat(
                [
                    torch.full(
                        encoder_outputs[0].shape[:-1], 
                        self.decoder.config.img_token_id, 
                        dtype=decoder_token_type_ids.dtype, 
                        device=decoder_token_type_ids.device,
                    ), 
                    decoder_token_type_ids
                ], 
                dim=1,
            )  
                    
            # Position identifiers accounting for padding:
            report_position_ids = decoder_attention_mask.cumsum(-1) + encoder_outputs[1].max(dim=1).values[:, None]
            report_position_ids.masked_fill_(decoder_attention_mask == 0, 1)
            decoder_position_ids = torch.cat([encoder_outputs[1], report_position_ids], dim=1)

            # 4D attention mask:
            decoder_attention_mask = self.create_4d_attention_mask_mixed_causality(encoder_outputs[1], decoder_attention_mask)

        assert decoder_position_ids is not None
        assert decoder_attention_mask is not None
        assert decoder_token_type_ids is not None

        if self.decoder.config.token_type_embeddings == 'add':
            decoder_inputs_embeds += self.token_type_embeddings(decoder_token_type_ids)
        elif self.decoder.config.token_type_embeddings == 'inbuilt':
            kwargs_decoder['token_type_ids'] = decoder_token_type_ids

        # Forward:
        decoder_outputs = self.decoder(
            inputs_embeds=decoder_inputs_embeds,
            attention_mask=decoder_attention_mask,
            position_ids=decoder_position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            use_cache=use_cache,
            past_key_values=past_key_values,
            return_dict=return_dict,
            **kwargs_decoder,
        )

        # Loss:
        loss = None
        if labels is not None:
            logits = decoder_outputs.logits if return_dict else decoder_outputs[0]
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.reshape(-1, self.decoder.config.vocab_size), labels.reshape(-1))

        if not return_dict:
            if loss is not None:
                return (loss,) + decoder_outputs + encoder_outputs
            else:
                return decoder_outputs + encoder_outputs

        encoder_hidden_states = encoder_outputs[0]

        return Seq2SeqLMOutput(
            loss=loss,
            logits=decoder_outputs.logits,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            encoder_last_hidden_state=encoder_hidden_states, 
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        special_token_ids,
        token_type_id_sections=None,
        past_key_values=None,
        use_cache=None,
        encoder_outputs=None,
        **kwargs,
    ):
        """
        Modification of: 
            https://github.com/huggingface/transformers/blob/main/src/transformers/models/encoder_decoder/modeling_encoder_decoder.py#L660
        """

        report_attention_mask = (input_ids != self.decoder.config.pad_token_id).long()

        if past_key_values is None:
            
            # 4D attention mask:
            decoder_attention_mask = self.create_4d_attention_mask_mixed_causality(encoder_outputs[1], report_attention_mask)
            
            # Position identifiers accounting for padding:
            report_position_ids = report_attention_mask.cumsum(-1) + encoder_outputs[1].max(dim=1).values[:, None]
            report_position_ids.masked_fill_(report_attention_mask == 0, 1)
            decoder_position_ids = torch.cat([encoder_outputs[1], report_position_ids], dim=1)
            
            # `inputs_embeds` are only to be used in the 1st generation step:
            inputs_embeds = torch.cat([encoder_outputs[0], self.decoder.get_input_embeddings()(input_ids)], dim=1)

            decoder_token_type_ids = self.token_ids_to_token_type_ids(input_ids, special_token_ids, token_type_id_sections)
            decoder_token_type_ids = torch.cat(
                [
                    torch.full(
                        encoder_outputs[0].shape[:-1], 
                        self.decoder.config.img_token_id, 
                        dtype=decoder_token_type_ids.dtype, 
                        device=decoder_token_type_ids.device,
                    ), 
                    decoder_token_type_ids,
                ], 
                dim=1,
            )  # Add image token type identifiers.

            input_dict = {
                'decoder_input_ids': input_ids, 
                'decoder_inputs_embeds': inputs_embeds, 
                'decoder_token_type_ids': decoder_token_type_ids,
            }
        else:
            
            # 4D attention mask:
            decoder_attention_mask = self.create_4d_attention_mask_mixed_causality_past_key_values(encoder_outputs[1], report_attention_mask)

            # Position identifiers accounting for padding:
            decoder_position_ids = report_attention_mask.cumsum(-1) + encoder_outputs[1].max(dim=1).values[:, None]
            decoder_position_ids.masked_fill_(report_attention_mask == 0, 1)
            
            # Always place token_ids_to_token_type_ids_past before input_ids = input_ids[:, remove_prefix_length:]:
            decoder_token_type_ids = self.token_ids_to_token_type_ids_past(input_ids, special_token_ids, token_type_id_sections)
            decoder_position_ids = decoder_position_ids[:, -1:]

            past_length = past_key_values[0][0].shape[2]

            # Some generation methods only pass the last input ID:
            if input_ids.shape[1] > past_length:
                remove_prefix_length = past_length
            else:
                # Keep only the final ID:
                remove_prefix_length = input_ids.shape[1] - 1

            input_ids = input_ids[:, remove_prefix_length:]

            input_dict = {'decoder_input_ids': input_ids, 'decoder_token_type_ids': decoder_token_type_ids}

        input_dict.update(
            {
                'decoder_attention_mask': decoder_attention_mask,
                'decoder_position_ids': decoder_position_ids,
                'encoder_outputs': encoder_outputs,
                'past_key_values': past_key_values,
                'use_cache': use_cache,
            }
        )
        return input_dict
        
    def token_ids_to_token_type_ids(self, token_ids, special_token_ids, token_type_id_sections=None):
        """
        Extract token type identifiers from the token identifiers.

        Argument/s:
            token_ids - token identifiers.
            special_token_ids - special token identifiers that indicate the separation between sections.
            token_type_id_section - token type identifier for each section.

        Returns:
            token_type_ids - token type identifiers.
        """

        token_type_id_sections = token_type_id_sections if token_type_id_sections is not None else list(range(len(special_token_ids) + 1))

        mbatch_size, seq_len = token_ids.shape
        token_type_ids = torch.full_like(token_ids, token_type_id_sections[0], dtype=torch.long, device=token_ids.device)

        for i, j in enumerate(special_token_ids):
            # Find first occurrence of special tokens that indicate the boundary between sections:
            cols = (token_ids == j).int().argmax(dim=1)
            rows = torch.arange(mbatch_size, device=token_ids.device)

            # https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer.create_token_type_ids_from_sequences.example
            cols += 1

            # Ensure that the column index is not out of bounds. If 0, then token_id not present.
            # This is safe as index 0 is always a special token (now equal to 1 due to +1):
            rows = rows[torch.logical_and(cols != 1, cols < seq_len)]
            cols = cols[torch.logical_and(cols != 1, cols < seq_len)]

            # Indices to that correspond to the second sequence:
            if rows.nelement() != 0:
                ids = torch.stack([
                    torch.stack([x, z]) for (x, y) in zip(rows, cols) for z in torch.arange(
                        y, seq_len, device=token_ids.device,
                    )
                ])

                token_type_ids[ids[:, 0], ids[:, 1]] = token_type_id_sections[i + 1]

        return token_type_ids

    def token_ids_to_token_type_ids_past(self, token_ids, special_token_ids, token_type_id_sections=None):
        """
        Extract token type identifiers from the token identifiers if past != None. Make sure to input all the
        token_ids (e.g., do not input input_ids = input_ids[:, remove_prefix_length:] from prepare_inputs_for_generation).

        Argument/s:
            token_ids - token identifiers.
            special_token_ids - special token identifiers that indicate the separation between sections.

        Returns:
            token_type_ids - token type identifiers.
        """

        token_type_id_sections = token_type_id_sections if token_type_id_sections is not None else list(range(len(special_token_ids) + 1))
        token_type_ids = torch.full([token_ids.shape[0], 1], token_type_id_sections[0], dtype=torch.long, device=token_ids.device)

        # https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertTokenizer.create_token_type_ids_from_sequences.example
        token_ids = token_ids[:, :-1]

        for i, j in enumerate(special_token_ids):

            # Find first occurrence of special token, which indicates the boundary between sections:
            exists = torch.any(token_ids == j, dim=1, keepdim=True)
            token_type_ids[exists] = token_type_id_sections[i + 1]

        return token_type_ids
    
    def tokenize_report_teacher_forcing(self, findings: str, impression: str, tokenizer: PreTrainedTokenizerFast, max_len: int):
        """
        Tokenize the reports and creates the inputs and targets for teacher forcing.

        Argument/s:
            findings - findings sections.
            impression - impression sections.
            return_token_type_ids - return the token type identifiers.
            tokenizer - Hugging Face tokenizer.
            max_len - maximum number of tokens.

        Returns:
            decoder_input_ids - the token identifiers for the input of the decoder.
            decoder_attention_mask - the attention mask for the decoder_input_ids.
            label_ids - the label token identifiers for the decoder.
        """

        # Prepare the sections for the tokenizer by placing special tokens between each section:
        reports = [f'{tokenizer.bos_token}{i}{tokenizer.sep_token}{j}{tokenizer.eos_token}' for i, j in
                  zip(findings, impression)]

        # Tokenize the report:
        tokenized = tokenizer(
            reports,
            padding='longest',
            truncation=True,
            max_length=max_len + 1,  # +1 to account for the bias between input and target.
            return_tensors='pt',
            return_token_type_ids=False,
            add_special_tokens=False,
        ).to(self.device)

        # Modify for language modelling:
        batch_dict = {

            # Labels for the decoder (shifted right by one for autoregression):
            'label_ids': tokenized['input_ids'][:, 1:].detach().clone(),

            # Remove last token identifier to match the sequence length of the labels:
            'decoder_input_ids': tokenized['input_ids'][:, :-1],

            # Attention mask for the decoder_input_ids (remove first token so that the eos_token_id is not considered):
            'decoder_attention_mask': tokenized['attention_mask'][:, 1:],
        }

        return batch_dict

    def tokenize_report_teacher_forcing_rev_a(self, tokenizer: PreTrainedTokenizerFast, max_len: int, findings: Optional[str] = None, impression: Optional[str] = None, reports: Optional[str] = None):
        """
        Tokenize the reports and creates the inputs and targets for teacher forcing.

        Argument/s:
            tokenizer - Hugging Face tokenizer.
            max_len - maximum number of tokens.
            findings - findings sections.
            impression - impression sections.
            reports - prepared reports, with special tokens and report sections.

        Returns:
            decoder_input_ids - the token identifiers for the input of the decoder.
            decoder_attention_mask - the attention mask for the decoder_input_ids.
            label_ids - the label token identifiers for the decoder.
        """

        # Prepare the sections for the tokenizer by placing special tokens between each section:
        if reports is None:
            assert findings and impression, "If 'reports' is not defined, 'findings' and 'impression' need to be defined." 
            reports = [f'{tokenizer.bos_token}{i}{tokenizer.sep_token}{j}{tokenizer.eos_token}' for i, j in
                    zip(findings, impression)]

        # Tokenize the report:
        tokenized = tokenizer(
            reports,
            padding='longest',
            truncation=True,
            max_length=max_len + 1,  # +1 to account for the bias between input and target.
            return_tensors='pt',
            return_token_type_ids=False,
            add_special_tokens=False,
        ).to(self.device)

        # Modify for language modelling:
        batch_dict = {

            # Labels for the decoder (shifted right by one for autoregression):
            'label_ids': tokenized['input_ids'][:, 1:].detach().clone(),

            # Remove last token identifier to match the sequence length of the labels:
            'decoder_input_ids': tokenized['input_ids'][:, :-1],

            # Attention mask for the decoder_input_ids (remove first token so that the eos_token_id is not considered):
            'decoder_attention_mask': tokenized['attention_mask'][:, 1:],
        }

        return batch_dict

    def split_and_decode_sections(self, token_ids, special_token_ids, tokenizer: PreTrainedTokenizerFast):
        """
        Split the token identifiers into sections, then convert the token identifiers into strings.

        Argument/s:
            token_ids - token identifiers.
            special_token_ids - special token identifiers that indicate the end of each section.
            tokenizer - Hugging Face tokenizer.

        Returns:
            token_type_ids - token type identifiers.
        """

        _, seq_len = token_ids.shape

        # The number of sections is the same as the number of special_token_ids:
        num_sections = len(special_token_ids)

        sections = {k: [] for k in range(num_sections)}

        for i in token_ids:
            prev_col = 0
            for j, k in enumerate(special_token_ids):

                # The maximum sequence length was exceeded, thus no more tokens:
                if prev_col >= seq_len:
                    sections[j].append('')
                    continue

                # Find first occurrence of special tokens that indicate the boundary between sections:
                col = (i == k).int().argmax().item()

                # If equal to 0, token was not found, set the column to the sequence length (as the decoder exceeded
                # the maximum sequence length):
                if col == 0:
                    col = seq_len

                # Extract section token identifiers:
                section_token_ids = i[prev_col:col]
                prev_col = col
                section_string = tokenizer.decode(section_token_ids, skip_special_tokens=True)

                sections[j].append(section_string)

        return tuple(sections.values())

    @staticmethod
    def create_4d_attention_mask_mixed_causality(non_causal_2d_attention_mask, causal_2d_attention_mask):
    
        prompt_seq_len = non_causal_2d_attention_mask.shape[-1] 
        report_seq_len = causal_2d_attention_mask.shape[-1]
        
        non_causal_2d_attention_mask = non_causal_2d_attention_mask[:, None, None, :]
        causal_2d_attention_mask = causal_2d_attention_mask[:, None, None, :]
    
        # Upper left of attention matrix:
        upper_left = non_causal_2d_attention_mask.expand(-1, -1, prompt_seq_len, -1)
        upper_left = upper_left * non_causal_2d_attention_mask
        upper_left = upper_left * non_causal_2d_attention_mask.permute(0, 1, 3, 2)
        
        causal_mask = torch.tril(
            torch.ones(
                (
                    report_seq_len, 
                    report_seq_len,
                ), 
                dtype=torch.long, 
                device=causal_2d_attention_mask.device,
            ),
        )   
        
        # Lower right of attention matrix:
        lower_right = causal_2d_attention_mask.expand(-1, -1, report_seq_len, -1)
        lower_right = lower_right * causal_2d_attention_mask.permute(0, 1, 3, 2)
        lower_right = lower_right * causal_mask
        
        # Upper right of attention matrix:
        upper_right = torch.zeros(
            causal_2d_attention_mask.shape[0], 
            1, 
            prompt_seq_len, 
            report_seq_len, 
            dtype=torch.long, 
            device=causal_2d_attention_mask.device,
        )
        
        # Lower left of attention matrix:
        lower_left = non_causal_2d_attention_mask.expand(-1, -1, report_seq_len, -1)
        lower_left = lower_left * causal_2d_attention_mask.permute(0, 1, 3, 2)
            
        left = torch.cat((upper_left, lower_left), dim=2)
        right = torch.cat((upper_right, lower_right), dim=2)

        mixed_causality_4d_attention_mask = torch.cat((left, right), dim=-1)
        return mixed_causality_4d_attention_mask
    
    @staticmethod
    def create_4d_attention_mask_mixed_causality_past_key_values(non_causal_2d_attention_mask, causal_2d_attention_mask):
    
        non_causal_2d_attention_mask = non_causal_2d_attention_mask[:, None, None, :]
        causal_2d_attention_mask = causal_2d_attention_mask[:, None, None, :]

        mixed_causality_4d_attention_mask = torch.cat((non_causal_2d_attention_mask, causal_2d_attention_mask), dim=-1)
        return mixed_causality_4d_attention_mask