# Copyright (c) OpenMMLab. All rights reserved. import warnings from typing import Dict, List, Optional, Sequence, Tuple import torch from mmdet.structures.bbox import bbox2roi from mmengine.model import BaseModel from torch import nn from mmocr.registry import MODELS, TASK_UTILS from mmocr.structures import KIEDataSample @MODELS.register_module() class SDMGR(BaseModel): """The implementation of the paper: Spatial Dual-Modality Graph Reasoning for Key Information Extraction. https://arxiv.org/abs/2103.14470. Args: backbone (dict, optional): Config of backbone. If None, None will be passed to kie_head during training and testing. Defaults to None. roi_extractor (dict, optional): Config of roi extractor. Only applicable when backbone is not None. Defaults to None. neck (dict, optional): Config of neck. Defaults to None. kie_head (dict): Config of KIE head. Defaults to None. dictionary (dict, optional): Config of dictionary. Defaults to None. data_preprocessor (dict or ConfigDict, optional): The pre-process config of :class:`BaseDataPreprocessor`. it usually includes, ``pad_size_divisor``, ``pad_value``, ``mean`` and ``std``. It has to be None when working in non-visual mode. Defaults to None. init_cfg (dict or list[dict], optional): Initialization configs. Defaults to None. """ def __init__(self, backbone: Optional[Dict] = None, roi_extractor: Optional[Dict] = None, neck: Optional[Dict] = None, kie_head: Dict = None, dictionary: Optional[Dict] = None, data_preprocessor: Optional[Dict] = None, init_cfg: Optional[Dict] = None) -> None: super().__init__( data_preprocessor=data_preprocessor, init_cfg=init_cfg) if dictionary is not None: self.dictionary = TASK_UTILS.build(dictionary) if kie_head.get('dictionary', None) is None: kie_head.update(dictionary=self.dictionary) else: warnings.warn(f"Using dictionary {kie_head['dictionary']} " "in kie_head's config.") if backbone is not None: self.backbone = MODELS.build(backbone) self.extractor = MODELS.build({ **roi_extractor, 'out_channels': self.backbone.base_channels }) self.maxpool = nn.MaxPool2d( roi_extractor['roi_layer']['output_size']) if neck is not None: self.neck = MODELS.build(neck) self.kie_head = MODELS.build(kie_head) def extract_feat(self, img: torch.Tensor, gt_bboxes: List[torch.Tensor]) -> torch.Tensor: """Extract features from images if self.backbone is not None. It returns None otherwise. Args: img (torch.Tensor): The input image with shape (N, C, H, W). gt_bboxes (list[torch.Tensor)): A list of ground truth bounding boxes, each of shape :math:`(N_i, 4)`. Returns: torch.Tensor: The extracted features with shape (N, E). """ if not hasattr(self, 'backbone'): return None x = self.backbone(img) if hasattr(self, 'neck'): x = self.neck(x) x = x[-1] feats = self.maxpool(self.extractor([x], bbox2roi(gt_bboxes))) return feats.view(feats.size(0), -1) def forward(self, inputs: torch.Tensor, data_samples: Sequence[KIEDataSample] = None, mode: str = 'tensor', **kwargs) -> torch.Tensor: """The unified entry for a forward process in both training and test. The method should accept three modes: "tensor", "predict" and "loss": - "tensor": Forward the whole network and return tensor or tuple of tensor without any post-processing, same as a common nn.Module. - "predict": Forward and return the predictions, which are fully processed to a list of :obj:`DetDataSample`. - "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`. Args: inputs (torch.Tensor): The input tensor with shape (N, C, ...) in general. data_samples (list[:obj:`DetDataSample`], optional): The annotation data of every samples. 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 tensor or a tuple of tensor. - If ``mode="predict"``, return a list of :obj:`DetDataSample`. - If ``mode="loss"``, return a dict of tensor. """ if mode == 'loss': return self.loss(inputs, data_samples, **kwargs) elif mode == 'predict': return self.predict(inputs, data_samples, **kwargs) elif mode == 'tensor': return self._forward(inputs, data_samples, **kwargs) else: raise RuntimeError(f'Invalid mode "{mode}". ' 'Only supports loss, predict and tensor mode') def loss(self, inputs: torch.Tensor, data_samples: Sequence[KIEDataSample], **kwargs) -> dict: """Calculate losses from a batch of inputs and data samples. Args: inputs (torch.Tensor): Input images of shape (N, C, H, W). Typically these should be mean centered and std scaled. data_samples (list[KIEDataSample]): A list of N datasamples, containing meta information and gold annotations for each of the images. Returns: dict[str, Tensor]: A dictionary of loss components. """ x = self.extract_feat( inputs, [data_sample.gt_instances.bboxes for data_sample in data_samples]) return self.kie_head.loss(x, data_samples) def predict(self, inputs: torch.Tensor, data_samples: Sequence[KIEDataSample], **kwargs) -> List[KIEDataSample]: """Predict results from a batch of inputs and data samples with post- processing. Args: inputs (torch.Tensor): Input images of shape (N, C, H, W). Typically these should be mean centered and std scaled. data_samples (list[KIEDataSample]): A list of N datasamples, containing meta information and gold annotations for each of the images. Returns: List[KIEDataSample]: A list of datasamples of prediction results. Results are stored in ``pred_instances.labels`` and ``pred_instances.edge_labels``. """ x = self.extract_feat( inputs, [data_sample.gt_instances.bboxes for data_sample in data_samples]) return self.kie_head.predict(x, data_samples) def _forward(self, inputs: torch.Tensor, data_samples: Sequence[KIEDataSample], **kwargs) -> Tuple[torch.Tensor, torch.Tensor]: """Get the raw tensor outputs from backbone and head without any post- processing. Args: inputs (torch.Tensor): Input images of shape (N, C, H, W). Typically these should be mean centered and std scaled. data_samples (list[KIEDataSample]): A list of N datasamples, containing meta information and gold annotations for each of the images. Returns: tuple(torch.Tensor, torch.Tensor): Tensor output from head. - node_cls (torch.Tensor): Node classification output. - edge_cls (torch.Tensor): Edge classification output. """ x = self.extract_feat( inputs, [data_sample.gt_instances.bboxes for data_sample in data_samples]) return self.kie_head(x, data_samples)