# Copyright (c) OpenMMLab. All rights reserved. import random from typing import Dict, Sequence, Union import numpy as np import torch from torch import nn from mmocr.models.common.dictionary import Dictionary from mmocr.models.textrecog.module_losses import CEModuleLoss from mmocr.registry import MODELS from mmocr.structures import TextSpottingDataSample @MODELS.register_module() class SPTSModuleLoss(CEModuleLoss): """Implementation of loss module for SPTS with CrossEntropy loss. Args: dictionary (dict or :obj:`Dictionary`): The config for `Dictionary` or the instance of `Dictionary`. num_bins (int): Number of bins dividing the image. Defaults to 1000. seq_eos_coef (float): The loss weight coefficient of seq_eos token. Defaults to 0.01. max_seq_len (int): Maximum sequence length. In SPTS, a sequence encodes all the text instances in a sample. Defaults to 40, which will be overridden by SPTSDecoder. max_text_len (int): Maximum length for each text instance in a sequence. Defaults to 25. letter_case (str): There are three options to alter the letter cases of gt texts: - unchanged: Do not change gt texts. - upper: Convert gt texts into uppercase characters. - lower: Convert gt texts into lowercase characters. Usually, it only works for English characters. Defaults to 'unchanged'. pad_with (str): The padding strategy for ``gt_text.padded_indexes``. Defaults to 'auto'. Options are: - 'auto': Use dictionary.padding_idx to pad gt texts, or dictionary.end_idx if dictionary.padding_idx is None. - 'padding': Always use dictionary.padding_idx to pad gt texts. - 'end': Always use dictionary.end_idx to pad gt texts. - 'none': Do not pad gt texts. ignore_char (int or str): Specifies a target value that is ignored and does not contribute to the input gradient. ignore_char can be int or str. If int, it is the index of the ignored char. If str, it is the character to ignore. Apart from single characters, each item can be one of the following reversed keywords: 'padding', 'start', 'end', and 'unknown', which refer to their corresponding special tokens in the dictionary. It will not ignore any special tokens when ignore_char == -1 or 'none'. Defaults to 'padding'. flatten (bool): Whether to flatten the output and target before computing CE loss. Defaults to False. reduction (str): Specifies the reduction to apply to the output, should be one of the following: ('none', 'mean', 'sum'). Defaults to 'none'. ignore_first_char (bool): Whether to ignore the first token in target ( usually the start token). Defaults to ``True``. flatten (bool): Whether to flatten the vectors for loss computation. Defaults to False. """ def __init__(self, dictionary: Union[Dict, Dictionary], num_bins: int, seq_eos_coef: float = 0.01, max_seq_len: int = 40, max_text_len: int = 25, letter_case: str = 'unchanged', pad_with: str = 'auto', ignore_char: Union[int, str] = 'padding', flatten: bool = False, reduction: str = 'none', ignore_first_char: bool = True): super().__init__(dictionary, max_seq_len, letter_case, pad_with, ignore_char, flatten, reduction, ignore_first_char) # TODO: fix hardcode self.max_text_len = max_text_len self.max_num_text = (self.max_seq_len - 1) // (2 + max_text_len) self.num_bins = num_bins weights = torch.ones(self.dictionary.num_classes, dtype=torch.float32) weights[self.dictionary.seq_end_idx] = seq_eos_coef weights.requires_grad_ = False self.loss_ce = nn.CrossEntropyLoss( ignore_index=self.ignore_index, reduction=reduction, weight=weights) def get_targets( self, data_samples: Sequence[TextSpottingDataSample] ) -> Sequence[TextSpottingDataSample]: """Target generator. Args: data_samples (list[TextSpottingDataSample]): It usually includes ``gt_instances`` information. Returns: list[TextSpottingDataSample]: Updated data_samples. Two keys will be added to data_sample: - indexes (torch.LongTensor): Character indexes representing gt texts. All special tokens are excluded, except for UKN. - padded_indexes (torch.LongTensor): Character indexes representing gt texts with BOS and EOS if applicable, following several padding indexes until the length reaches ``max_seq_len``. In particular, if ``pad_with='none'``, no padding will be applied. """ batch_max_len = 0 for data_sample in data_samples: if data_sample.get('have_target', False): continue if len(data_sample.gt_instances) > self.max_num_text: keep = random.sample( range(len(data_sample.gt_instances)), self.max_num_text) data_sample.gt_instances = data_sample.gt_instances[keep] gt_instances = data_sample.gt_instances if len(gt_instances) > 0: center_pts = [] # Slightly different from the original implementation # which gets the center points from bezier curves # for bezier_pt in gt_instances.beziers: # bezier_pt = bezier_pt.reshape(8, 2) # mid_pt1 = sample_bezier_curve( # bezier_pt[:4], mid_point=True) # mid_pt2 = sample_bezier_curve( # bezier_pt[4:], mid_point=True) # center_pt = (mid_pt1 + mid_pt2) / 2 for polygon in gt_instances.polygons: center_pt = polygon.reshape(-1, 2).mean(0) center_pts.append(center_pt) center_pts = np.vstack(center_pts) center_pts /= data_sample.img_shape[::-1] center_pts = torch.from_numpy(center_pts).type(torch.float32) else: center_pts = torch.ones(0).reshape(-1, 2).type(torch.float32) center_pts = (center_pts * self.num_bins).floor().type(torch.long) center_pts = torch.clamp(center_pts, min=0, max=self.num_bins - 1) gt_indexes = [] for text in gt_instances.texts: if self.letter_case in ['upper', 'lower']: text = getattr(text, self.letter_case)() indexes = self.dictionary.str2idx(text) indexes_tensor = torch.zeros( self.max_text_len, dtype=torch.long) + self.dictionary.end_idx max_len = min(self.max_text_len - 1, len(indexes)) indexes_tensor[:max_len] = torch.LongTensor(indexes)[:max_len] indexes_tensor = indexes_tensor gt_indexes.append(indexes_tensor) if len(gt_indexes) == 0: gt_indexes = torch.ones(0).reshape(-1, self.max_text_len) else: gt_indexes = torch.vstack(gt_indexes) gt_indexes = torch.cat([center_pts, gt_indexes], dim=-1) gt_indexes = gt_indexes.flatten() if self.dictionary.start_idx is not None: gt_indexes = torch.cat([ torch.LongTensor([self.dictionary.start_idx]), gt_indexes ]) if self.dictionary.seq_end_idx is not None: gt_indexes = torch.cat([ gt_indexes, torch.LongTensor([self.dictionary.seq_end_idx]) ]) batch_max_len = max(batch_max_len, len(gt_indexes)) gt_instances.set_metainfo(dict(indexes=gt_indexes)) # Here we have to have the second pass as we need to know the max # length of the batch to pad the indexes in order to save memory for data_sample in data_samples: if data_sample.get('have_target', False): continue indexes = data_sample.gt_instances.indexes padded_indexes = ( torch.zeros(batch_max_len, dtype=torch.long) + self.dictionary.padding_idx) padded_indexes[:len(indexes)] = indexes data_sample.gt_instances.set_metainfo( dict(padded_indexes=padded_indexes)) data_sample.set_metainfo(dict(have_target=True)) return data_samples def forward(self, outputs: torch.Tensor, data_samples: Sequence[TextSpottingDataSample]) -> Dict: """ Args: outputs (Tensor): A raw logit tensor of shape :math:`(N, T, C)`. data_samples (list[TextSpottingDataSample]): List of ``TextSpottingDataSample`` which are processed by ``get_targets``. Returns: dict: A loss dict with the key ``loss_ce``. """ targets = list() for data_sample in data_samples: targets.append(data_sample.gt_instances.padded_indexes) targets = torch.stack(targets, dim=0).long() if self.ignore_first_char: targets = targets[:, 1:].contiguous() # outputs = outputs[:, :-1, :].contiguous() if self.flatten: outputs = outputs.view(-1, outputs.size(-1)) targets = targets.view(-1) else: outputs = outputs.permute(0, 2, 1).contiguous() loss_ce = self.loss_ce(outputs, targets.to(outputs.device)) losses = dict(loss_ce=loss_ce) return losses