mmpretrain/models/multimodal/blip/blip_retrieval.py

# Copyright (c) OpenMMLab. All rights reserved.
from collections import ChainMap
from copy import deepcopy
from typing import Dict, List, Optional, Tuple, Union

import mmengine.dist as dist
import torch
import torch.nn as nn
import torch.nn.functional as F
from mmengine.model import BaseModel
from torch import distributed as torch_dist

from mmpretrain.registry import MODELS, TOKENIZER
from mmpretrain.structures import DataSample
from mmpretrain.utils import track_on_main_process


def all_gather_concat(data: torch.Tensor) -> torch.Tensor:
    """Gather tensors with different first-dimension size and concat to one
    tenosr.

    Note:
        Only the first dimension should be different.

    Args:
        data (Tensor): Tensor to be gathered.

    Returns:
        torch.Tensor: The concatenated tenosr.
    """
    if dist.get_world_size() == 1:
        return data

    data_size = torch.tensor(data.size(0), device=data.device)
    sizes_list = dist.all_gather(data_size)

    max_length = max(sizes_list)
    size_diff = max_length.item() - data_size.item()
    if size_diff:
        padding = torch.zeros(
            size_diff, *data.size()[1:], device=data.device, dtype=data.dtype)
        data = torch.cat((data, padding))

    gather_list = dist.all_gather(data)

    all_data = []
    for tensor, size in zip(gather_list, sizes_list):

        all_data.append(tensor[:size])

    return torch.concat(all_data)


@MODELS.register_module()
class BlipRetrieval(BaseModel):
    """BLIP Retriever.

    Args:
        vision_backbone (dict): Backbone for extracting image features.
        text_backbone (dict): Backbone for extracting text features.
        multimodal_backbone (Optional[dict]): Backbone for extracting
            multi-modal features.
        vision_neck (Optional[dict]): The neck module to process image features
            from vision backbone. Defaults to None.
        text_neck (Optional[dict]): The neck module to process text features
            from text backbone. Defaults to None.
        head (Optional[Union[List[dict], dict]]): The head module to calculate
            loss from processed single modality features.
            See :mod:`mmmultimodal.models.heads`.
            Notice that if the head is not set, `loss` method cannot be used.
            Defaults to None.
        multimodal_head (Optional[Union[List[dict], dict]]): The multi-modal
            head module to calculate loss from processed multimodal features.
            See :mod:`mmmultimodal.models.heads`.
            Notice that if the head is not set, `loss` method cannot be used.
            Defaults to None.
        momentum (float): Momentum used for momentum contrast.
            Defaults to .995.
        negative_all_rank (bool): Whether to sample negative data from all
            ranks for image text matching in training. Defaults to True.
        temperature (float): Temperature parameter that controls the
            concentration level of the distribution. Defaults to 0.07.
        fast_match (bool): If False, select topk similarity as candidates and
            compute the matching score. If True, return the similarity as the
            matching score directly. Defaults to False.
        topk (int): Select topk similarity as candidates for compute matching
            scores. Notice that this is not the topk in evaluation.
            Defaults to 256.
        data_preprocessor (Optional[dict]): The config for preprocessing input
            data. If None or no specified type, it will use
            "MutimodalDataPreprocessor" as type.
            See :class:`MutimodalDataPreprocessor` for more details.
            Defaults to None.
        init_cfg (Optional[dict]): the config to control the initialization.
            Defaults to None.
    """

    def __init__(self,
                 vision_backbone: dict,
                 text_backbone: dict,
                 multimodal_backbone: Optional[dict] = None,
                 vision_neck: Optional[dict] = None,
                 text_neck: Optional[dict] = None,
                 head: Optional[Union[List[dict], dict]] = None,
                 multimodal_head: Optional[Union[List[dict], dict]] = None,
                 tokenizer: Optional[dict] = None,
                 momentum: float = .995,
                 negative_all_rank: bool = True,
                 temperature: float = 0.07,
                 fast_match: bool = False,
                 topk: int = 256,
                 max_txt_len: int = 20,
                 data_preprocessor: Optional[dict] = None,
                 init_cfg: Optional[dict] = None):
        if data_preprocessor is None:
            data_preprocessor = {}
        if isinstance(data_preprocessor, dict):
            data_preprocessor.setdefault('type', 'MultiModalDataPreprocessor')
            data_preprocessor = MODELS.build(data_preprocessor)

        super().__init__(
            init_cfg=init_cfg, data_preprocessor=data_preprocessor)

        self.vision_backbone = MODELS.build(vision_backbone)
        self.text_backbone = MODELS.build(text_backbone)

        if multimodal_backbone is not None:
            self.multimodal_backbone = MODELS.build(multimodal_backbone)

        if vision_neck is not None:
            self.vision_neck = MODELS.build(vision_neck)

        if text_neck is not None:
            self.text_neck = MODELS.build(text_neck)

        if head is not None:
            self.head = MODELS.build(head)

        if multimodal_head is not None:
            self.multimodal_head = MODELS.build(multimodal_head)

        if tokenizer is not None:
            self.tokenizer = TOKENIZER.build(tokenizer)

        self.momentum = momentum
        self.negative_all_rank = negative_all_rank
        self.temp = nn.Parameter(temperature * torch.ones([]))
        # Shares the same para
        self.head.temp = self.temp

        # create the momentum encoder
        self.vision_backbone_m = deepcopy(self.vision_backbone)
        self.text_backbone_m = deepcopy(self.text_backbone)

        self.vision_neck_m = deepcopy(self.vision_neck)
        self.text_neck_m = deepcopy(self.text_neck)

        self.model_pairs = [
            [self.vision_backbone, self.vision_backbone_m],
            [self.text_backbone, self.text_backbone_m],
            [self.vision_neck, self.vision_neck_m],
            [self.text_neck, self.text_neck_m],
        ]
        self.copy_params()

        # multimodal backbone shares weights with text backbone in BLIP
        # No need to set up

        # Notice that this topk is used for select k candidate to compute
        # image-text score, but not the final metric topk in evaluation.
        self.fast_match = fast_match
        self.topk = topk

        self.max_txt_len = max_txt_len

    @property
    def device(self):
        return next(self.parameters()).device

    def preprocess_text(self, data_samples):
        sample_item = data_samples[0]

        if sample_item is not None and 'text' in sample_item:
            if isinstance(sample_item.get('text'), (list, tuple)):
                texts = []
                for sample in data_samples:
                    texts.extend(sample.get('text'))
            elif isinstance(sample_item.get('text'), str):
                texts = [sample.get('text') for sample in data_samples]
            else:
                raise TypeError('text must be a string or a list of strings')
        else:
            return None

        # perform tokenize first if satisfied conditions
        texts = self.tokenizer(
            texts,
            padding='max_length',
            truncation=True,
            max_length=self.max_txt_len,
            return_tensors='pt',
        ).to(self.device)

        return texts

    def forward(self,
                images: torch.tensor = None,
                data_samples: Optional[List[DataSample]] = None,
                mode: str = 'tensor') -> Union[Tuple, dict]:
        """The unified entry for a forward process in both training and test.
        The method should accept two modes: "tensor", and "loss":

        - "tensor": Forward the whole network and return tensor without any
          post-processing, same as a common nn.Module.
        - "loss": Forward and return a dict of losses according to the given
          inputs and data samples.

        Note that this method doesn't handle neither back propagation nor
        optimizer updating, which are done in the :meth:`train_step`.

        For unified "predict" mode in other mm repos. It is noticed that
        image-text retrieval cannot perform batch prediction since it will go
        through all the samples. A standard process of retrieval evaluation is
        to extract and collect all feats, and then predict all samples.
        Therefore the `predict` mode here is remained as a trigger
        to inform use to choose the right configurations.

        Args:
            images (torch.Tensor): The input inputs tensor of shape
                (N, C, ...) in general.
            data_samples (List[DataSample], optional): The annotation
                data of every samples. It's required if ``mode="loss"``.
                Defaults to None.
            mode (str): Return what kind of value. Defaults to 'tensor'.

        Returns:
            The return type depends on ``mode``.
            - If ``mode="tensor"``, return a tuple.
            - If ``mode="loss"``, return a dict of tensor.
        """
        if mode == 'tensor':
            return self.extract_feat(images, data_samples)
        elif mode == 'loss':
            return self.loss(images, data_samples)
        elif mode == 'predict':
            return self.predict(images, data_samples)
        else:
            raise RuntimeError(f'Invalid mode "{mode}".')

    def extract_feat(
        self,
        images: torch.Tensor = None,
        data_samples: List[DataSample] = None,
        return_texts=True,
        return_embeds=None,
    ) -> Dict[str, torch.Tensor]:
        """Extract features from the input dict.

        Args:
            images (tensor, optional): The images to extract features.
                Defaults to None.
            data_samples (list, optional): The data samples containing texts
                to extract features. Defaults to None.
            return_texts (bool): Whether to return the tokenized text and the
                corresponding attention masks. Defaults to True.
            return_embeds (bool): Whether to return the text embedding and
                image embedding. Defaults to None, which means to use
                ``self.fast_match``.

        Returns:
            Tuple[torch.Tensor]: The output features.
                If multimodal_backbone is not exist, tuple of torch.Tensor
                will be returned.
        """
        if data_samples is not None:
            texts = self.preprocess_text(data_samples)
        else:
            texts = None

        assert images is not None or texts is not None, \
            'At least single modality should be passed as inputs.'

        results = {}
        if texts is not None and return_texts:
            results.update({
                'text_ids': texts.input_ids,
                'text_attn_mask': texts.attention_mask,
            })

        if return_embeds is None:
            return_embeds = not self.fast_match

        # extract image features
        if images is not None:
            output = self._extract_feat(images, modality='images')
            results['image_feat'] = output['image_feat']
            if return_embeds:
                results['image_embeds'] = output['image_embeds']

        # extract text features
        if texts is not None:
            output = self._extract_feat(texts, modality='texts')
            results['text_feat'] = output['text_feat']
            if return_embeds:
                results['text_embeds'] = output['text_embeds']

        return results

    def _extract_feat(self, inputs: Union[torch.Tensor, dict],
                      modality: str) -> Tuple[torch.Tensor]:
        """Extract features from the single modality.

        Args:
            inputs (Union[torch.Tensor, dict]): A batch of inputs.
                For image, a tensor of shape (N, C, ...) in general.
                For text, a dict of tokenized text inputs.
            modality (str): Modality feature to be extracted. Only two
                options are supported.

                - ``images``: Only extract image features, mostly used for
                    inference.
                - ``texts``: Only extract text features, mostly used for
                    inference.

        Returns:
            Tuple[torch.Tensor]: The output features.
        """

        if modality == 'images':
            # extract image features
            image_embeds = self.vision_backbone(inputs)[0]
            image_feat = F.normalize(
                self.vision_neck(image_embeds[:, 0, :]), dim=-1)
            return {'image_embeds': image_embeds, 'image_feat': image_feat}
        elif modality == 'texts':
            # extract text features
            text_output = self.text_backbone(
                inputs.input_ids,
                attention_mask=inputs.attention_mask,
                token_type_ids=None,
                return_dict=True,
                mode='text',
            )
            text_embeds = text_output.last_hidden_state
            text_feat = F.normalize(
                self.text_neck(text_embeds[:, 0, :]), dim=-1)
            return {'text_embeds': text_embeds, 'text_feat': text_feat}
        else:
            raise RuntimeError(f'Invalid modality "{modality}".')

    def loss(
        self,
        images: torch.Tensor,
        data_samples: Optional[List[DataSample]] = None,
    ) -> Dict[str, torch.tensor]:
        """Calculate losses from a batch of inputs and data samples.

        Args:
            inputs (dict): A batch of inputs. The input tensor with of
                at least one modality. For image, the value is a tensor
                of shape (N, C, ...) in general.
                For text, the value is a dict of tokenized text inputs.
            data_samples (Optional[List[DataSample]]):
                The annotation data of every samples. Defaults to None.

        Returns:
            Dict[str, torch.tensor]: a dictionary of loss components of
                both head and multimodal head.
        """
        output = self.extract_feat(images, data_samples, return_embeds=True)

        text_ids = output['text_ids']
        text_attn_mask = output['text_attn_mask']
        image_embeds = output['image_embeds']
        image_feat = output['image_feat']
        text_feat = output['text_feat']

        image_atts = torch.ones(
            image_embeds.size()[:-1], dtype=torch.long).to(self.device)

        # get momentum features
        with torch.no_grad():
            self._momentum_update()
            image_embeds_m = self.vision_backbone_m(images)[0]
            image_feat_m = F.normalize(
                self.vision_neck_m(image_embeds_m[:, 0, :]), dim=-1)

            text_output_m = self.text_backbone_m(
                text_ids,
                attention_mask=text_attn_mask,
                token_type_ids=None,
                return_dict=True,
                mode='text',
            )
            text_embeds_m = text_output_m.last_hidden_state
            text_feat_m = F.normalize(
                self.text_neck_m(text_embeds_m[:, 0, :]), dim=-1)

        loss = self.head.loss(
            ([image_feat, text_feat, image_feat_m, text_feat_m], ),
            data_samples)

        # prepare for itm
        encoder_input_ids = text_ids.clone()
        encoder_input_ids[:,
                          0] = self.tokenizer.additional_special_tokens_ids[0]
        output_pos = self.text_backbone(
            encoder_input_ids,
            attention_mask=text_attn_mask,
            encoder_hidden_states=image_embeds,
            encoder_attention_mask=image_atts,
            return_dict=True,
        )

        idx = torch.tensor([i.image_id for i in data_samples]).view(-1, 1)
        bs = idx.size(0)
        idxs = torch.cat(dist.all_gather(idx))
        if self.negative_all_rank:
            # compute sample similarity
            with torch.no_grad():
                mask = torch.eq(idx, idxs.t()).to(self.device)

                image_feat_world = torch.cat(dist.all_gather(image_feat))
                text_feat_world = torch.cat(dist.all_gather(text_feat))

                sim_i2t = image_feat @ text_feat_world.t() / self.temp
                sim_t2i = text_feat @ image_feat_world.t() / self.temp

                weights_i2t = F.softmax(sim_i2t, dim=1)
                weights_i2t.masked_fill_(mask, 0)

                weights_t2i = F.softmax(sim_t2i, dim=1)
                weights_t2i.masked_fill_(mask, 0)

            world_size = dist.get_world_size()
            if world_size == 1:
                image_embeds_world = image_embeds
            else:
                image_embeds_world = torch.cat(
                    torch_dist.nn.all_gather(image_embeds))

            # select a negative image (from all ranks) for each text
            image_embeds_neg = []
            for b in range(bs):
                neg_idx = torch.multinomial(weights_t2i[b], 1).item()
                image_embeds_neg.append(image_embeds_world[neg_idx])
            image_embeds_neg = torch.stack(image_embeds_neg, dim=0)

            # select a negative text (from all ranks) for each image
            input_ids_world = torch.cat(dist.all_gather(encoder_input_ids))
            att_mask_world = torch.cat(dist.all_gather(text_attn_mask))

            text_ids_neg = []
            text_atts_neg = []
            for b in range(bs):
                neg_idx = torch.multinomial(weights_i2t[b], 1).item()
                text_ids_neg.append(input_ids_world[neg_idx])
                text_atts_neg.append(att_mask_world[neg_idx])

        text_ids_neg = torch.stack(text_ids_neg, dim=0)
        text_atts_neg = torch.stack(text_atts_neg, dim=0)

        text_ids_all = torch.cat([encoder_input_ids, text_ids_neg], dim=0)
        text_atts_all = torch.cat([text_attn_mask, text_atts_neg], dim=0)

        image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)
        image_atts_all = torch.cat([image_atts, image_atts], dim=0)

        output_neg = self.text_backbone(
            text_ids_all,
            attention_mask=text_atts_all,
            encoder_hidden_states=image_embeds_all,
            encoder_attention_mask=image_atts_all,
            return_dict=True,
        )

        vl_embeddings = torch.cat(
            [
                output_pos.last_hidden_state[:, 0, :],
                output_neg.last_hidden_state[:, 0, :],
            ],
            dim=0,
        )

        # create false data samples
        data_samples.extend(
            [DataSample(is_matched=False) for _ in range(2 * bs)])
        loss_multimodal = self.multimodal_head.loss((vl_embeddings, ),
                                                    data_samples)

        return dict(ChainMap(loss, loss_multimodal))

    def predict(self, images, data_samples, cal_i2t=True, cal_t2i=True):
        feats = self.extract_feat(images, data_samples)

        return self.predict_all(
            feats, data_samples, cal_i2t=cal_i2t, cal_t2i=cal_t2i)

    def predict_all(self,
                    feats,
                    data_samples,
                    num_images=None,
                    num_texts=None,
                    cal_i2t=True,
                    cal_t2i=True):
        text_ids = feats['text_ids']
        text_ids[:, 0] = self.tokenizer.additional_special_tokens_ids[0]
        text_attn_mask = feats['text_attn_mask']
        image_embeds = feats.get('image_embeds', None)
        image_feat = feats['image_feat']
        text_feat = feats['text_feat']

        num_images = num_images or image_feat.size(0)
        num_texts = num_texts or text_feat.size(0)

        if not self.fast_match:
            image_embeds_all = all_gather_concat(image_embeds)[:num_images]
        else:
            image_embeds_all = None
        image_feat_all = all_gather_concat(image_feat)[:num_images]
        text_feat_all = all_gather_concat(text_feat)[:num_texts]
        text_ids_all = all_gather_concat(text_ids)[:num_texts]
        text_attn_mask_all = all_gather_concat(text_attn_mask)[:num_texts]

        results = []
        if cal_i2t:
            result_i2t = self.compute_score_matrix_i2t(
                image_feat,
                image_embeds,
                text_feat_all,
                text_ids_all,
                text_attn_mask_all,
            )
            results.append(
                self._get_predictions(result_i2t, data_samples, mode='i2t'))
        if cal_t2i:
            result_t2i = self.compute_score_matrix_t2i(
                image_feat_all,
                image_embeds_all,
                text_feat,
                text_ids,
                text_attn_mask,
            )
            results.append(
                self._get_predictions(result_t2i, data_samples, mode='t2i'))
        return tuple(results)

    def compute_score_matrix_i2t(self, img_feats, img_embeds, text_feats,
                                 text_ids, text_atts):
        """Compare the score matrix for image-to-text retrieval. Every image
        should compare to all the text features.

        Args:
            img_feats (torch.Tensor): The input img feats tensor with shape
                (M, C). M stands for numbers of samples on a single GPU.
            img_embeds (torch.Tensor): The input img embeds tensor with shape
                (M, C). M stands for numbers of samples on a single GPU.
            text_feats (torch.Tensor): The input text feats tensor with shape
                (N, C). N stands for numbers of all samples on all GPUs.
            text_ids (torch.Tensor): The input tensor with shape (N, C).
            text_atts (torch.Tensor): The input tensor with shape (N, C).

        Returns:
            torch.Tensor: Score matrix of image-to-text retrieval.
        """

        # compute i2t sim matrix
        sim_matrix_i2t = img_feats @ text_feats.t()
        if self.fast_match:
            return sim_matrix_i2t

        score_matrix_i2t = torch.full((img_feats.size(0), text_feats.size(0)),
                                      -100.0).to(self.device)
        for i in track_on_main_process(
                range(img_feats.size(0)), 'Compute I2T scores...'):
            sims = sim_matrix_i2t[i]
            topk_sim, topk_idx = sims.topk(k=self.topk, dim=0)

            encoder_output = img_embeds[i].repeat(self.topk, 1, 1)
            encoder_att = torch.ones(
                encoder_output.size()[:-1], dtype=torch.long).to(self.device)
            output = self.text_backbone(
                text_ids[topk_idx],
                attention_mask=text_atts[topk_idx],
                encoder_hidden_states=encoder_output,
                encoder_attention_mask=encoder_att,
                return_dict=True,
            )
            score = self.multimodal_head(
                (output.last_hidden_state[:, 0, :], ))[:, 1]
            score_matrix_i2t[i, topk_idx] = score + topk_sim

        return score_matrix_i2t

    def compute_score_matrix_t2i(self, img_feats, img_embeds, text_feats,
                                 text_ids, text_atts):
        """Compare the score matrix for text-to-image retrieval. Every text
        should compare to all the image features.

        Args:
            img_feats (torch.Tensor): The input img feats tensor with shape
                (M, C). M stands for numbers of samples on a single GPU.
            img_embeds (torch.Tensor): The input img embeds tensor with shape
                (M, C). M stands for numbers of samples on a single GPU.
            text_feats (torch.Tensor): The input text feats tensor with shape
                (N, C). N stands for numbers of all samples on all GPUs.
            text_ids (torch.Tensor): The input tensor with shape (M, C).
            text_atts (torch.Tensor): The input tensor with shape (M, C).

        Returns:
            torch.Tensor: Score matrix of text-to-image retrieval.
        """

        # compute t2i sim matrix
        sim_matrix_t2i = text_feats @ img_feats.t()
        if self.fast_match:
            return sim_matrix_t2i

        score_matrix_t2i = torch.full((text_feats.size(0), img_feats.size(0)),
                                      -100.0).to(self.device)
        for i in track_on_main_process(
                range(text_feats.size(0)), 'Compute T2I scores...'):
            sims = sim_matrix_t2i[i]
            topk_sim, topk_idx = sims.topk(k=self.topk, dim=0)

            encoder_output = img_embeds[topk_idx]
            encoder_att = torch.ones(
                encoder_output.size()[:-1], dtype=torch.long).to(self.device)
            output = self.text_backbone(
                text_ids[i].repeat(self.topk, 1),
                attention_mask=text_atts[i].repeat(self.topk, 1),
                encoder_hidden_states=encoder_output,
                encoder_attention_mask=encoder_att,
                return_dict=True,
            )
            score = self.multimodal_head(
                (output.last_hidden_state[:, 0, :], ))[:, 1]
            score_matrix_t2i[i, topk_idx] = score + topk_sim

        return score_matrix_t2i

    def _get_predictions(self,
                         result: torch.Tensor,
                         data_samples: List[DataSample],
                         mode: str = 'i2t'):
        """Post-process the output of retriever.

        Args:
            result (torch.Tensor): Score matrix of single retrieve,
                either from image or text.
            data_samples (List[DataSample], optional): The annotation
                data of every samples.
            mode (str): Retrieve mode, either `i2t` for image to text, or `t2i`
                text to image. Defaults to `i2t`.

        Returns:
            List[DataSample]: the raw data_samples with
                the predicted results.
        """

        # create data sample if not exists
        if data_samples is None:
            data_samples = [DataSample() for _ in range(result.size(0))]
        elif mode == 't2i':
            # Process data samples to align with the num of texts.
            new_data_samples = []
            for sample in data_samples:
                if isinstance(sample.text, (list, tuple)):
                    texts = sample.text
                else:
                    texts = [sample.text]
                for i, text in enumerate(texts):
                    new_sample = DataSample(text=text)
                    if 'gt_image_id' in sample:
                        new_sample.gt_label = sample.gt_image_id[i]
                    new_data_samples.append(new_sample)
            assert len(new_data_samples) == result.size(0)
            data_samples = new_data_samples
        elif mode == 'i2t':
            for sample in data_samples:
                if 'gt_text_id' in sample:
                    sample.gt_label = sample.gt_text_id
        else:
            raise ValueError(f'Type {mode} is not supported.')

        for data_sample, score in zip(data_samples, result):
            idx = score.argmax(keepdim=True).detach()

            data_sample.set_pred_score(score)
            data_sample.set_pred_label(idx)
        return data_samples

    # TODO: add temperaily
    @torch.no_grad()
    def copy_params(self):
        for model_pair in self.model_pairs:
            for param, param_m in zip(model_pair[0].parameters(),
                                      model_pair[1].parameters()):
                param_m.data.copy_(param.data)  # initialize
                param_m.requires_grad = False  # not update by gradient

    @torch.no_grad()
    def _momentum_update(self):
        for model_pair in self.model_pairs:
            for (name,
                 param), (name_m,
                          param_m) in zip(model_pair[0].named_parameters(),
                                          model_pair[1].named_parameters()):
                # hack to behave the same
                if any([i in name for i in ['8', '9', '10', '11']
                        ]) and 'layers' in name and any(
                            [i in name for i in ['attn', 'ffn']]):
                    param_m.data = param.data
                else:
                    param_m.data = param_m.data * self.momentum + \
                        param.data * (1.0 - self.momentum)