from transformers import PreTrainedModel, VisionEncoderDecoderModel, ViTMAEModel, ConditionalDetrModel from transformers.models.conditional_detr.modeling_conditional_detr import ( ConditionalDetrMLPPredictionHead, ConditionalDetrModelOutput, ConditionalDetrHungarianMatcher, inverse_sigmoid, ) from .configuration_magi import MagiConfig from .processing_magi import MagiProcessor from torch import nn from typing import Optional, List import torch from einops import rearrange, repeat, einsum from .utils import move_to_device, visualise_single_image_prediction, sort_panels, sort_text_boxes_in_reading_order class MagiModel(PreTrainedModel): config_class = MagiConfig def __init__(self, config): super().__init__(config) self.config = config self.processor = MagiProcessor(config) if not config.disable_ocr: self.ocr_model = VisionEncoderDecoderModel(config.ocr_model_config) if not config.disable_crop_embeddings: self.crop_embedding_model = ViTMAEModel(config.crop_embedding_model_config) if not config.disable_detections: self.num_non_obj_tokens = 5 self.detection_transformer = ConditionalDetrModel(config.detection_model_config) self.bbox_predictor = ConditionalDetrMLPPredictionHead( input_dim=config.detection_model_config.d_model, hidden_dim=config.detection_model_config.d_model, output_dim=4, num_layers=3 ) self.is_this_text_a_dialogue = ConditionalDetrMLPPredictionHead( input_dim=config.detection_model_config.d_model, hidden_dim=config.detection_model_config.d_model, output_dim=1, num_layers=3 ) self.character_character_matching_head = ConditionalDetrMLPPredictionHead( input_dim = 3 * config.detection_model_config.d_model + (2 * config.crop_embedding_model_config.hidden_size if not config.disable_crop_embeddings else 0), hidden_dim=config.detection_model_config.d_model, output_dim=1, num_layers=3 ) self.text_character_matching_head = ConditionalDetrMLPPredictionHead( input_dim = 3 * config.detection_model_config.d_model, hidden_dim=config.detection_model_config.d_model, output_dim=1, num_layers=3 ) self.class_labels_classifier = nn.Linear( config.detection_model_config.d_model, config.detection_model_config.num_labels ) self.matcher = ConditionalDetrHungarianMatcher( class_cost=config.detection_model_config.class_cost, bbox_cost=config.detection_model_config.bbox_cost, giou_cost=config.detection_model_config.giou_cost ) def move_to_device(self, input): return move_to_device(input, self.device) def predict_detections_and_associations( self, images, move_to_device_fn=None, character_detection_threshold=0.3, panel_detection_threshold=0.2, text_detection_threshold=0.25, character_character_matching_threshold=0.65, text_character_matching_threshold=0.4, ): assert not self.config.disable_detections move_to_device_fn = self.move_to_device if move_to_device_fn is None else move_to_device_fn inputs_to_detection_transformer = self.processor.preprocess_inputs_for_detection(images) inputs_to_detection_transformer = move_to_device_fn(inputs_to_detection_transformer) detection_transformer_output = self._get_detection_transformer_output(**inputs_to_detection_transformer) predicted_class_scores, predicted_bboxes = self._get_predicted_bboxes_and_classes(detection_transformer_output) # create callback fn def get_character_character_matching_scores(batch_character_indices, batch_bboxes): predicted_obj_tokens_for_batch = self._get_predicted_obj_tokens(detection_transformer_output) predicted_c2c_tokens_for_batch = self._get_predicted_c2c_tokens(detection_transformer_output) crop_bboxes = [batch_bboxes[i][batch_character_indices[i]] for i in range(len(batch_character_indices))] crop_embeddings_for_batch = self.predict_crop_embeddings(images, crop_bboxes, move_to_device_fn) character_obj_tokens_for_batch = [] c2c_tokens_for_batch = [] for predicted_obj_tokens, predicted_c2c_tokens, character_indices in zip(predicted_obj_tokens_for_batch, predicted_c2c_tokens_for_batch, batch_character_indices): character_obj_tokens_for_batch.append(predicted_obj_tokens[character_indices]) c2c_tokens_for_batch.append(predicted_c2c_tokens) return self._get_character_character_affinity_matrices( character_obj_tokens_for_batch=character_obj_tokens_for_batch, crop_embeddings_for_batch=crop_embeddings_for_batch, c2c_tokens_for_batch=c2c_tokens_for_batch, apply_sigmoid=True, ) # create callback fn def get_text_character_matching_scores(batch_text_indices, batch_character_indices): predicted_obj_tokens_for_batch = self._get_predicted_obj_tokens(detection_transformer_output) predicted_t2c_tokens_for_batch = self._get_predicted_t2c_tokens(detection_transformer_output) text_obj_tokens_for_batch = [] character_obj_tokens_for_batch = [] t2c_tokens_for_batch = [] for predicted_obj_tokens, predicted_t2c_tokens, text_indices, character_indices in zip(predicted_obj_tokens_for_batch, predicted_t2c_tokens_for_batch, batch_text_indices, batch_character_indices): text_obj_tokens_for_batch.append(predicted_obj_tokens[text_indices]) character_obj_tokens_for_batch.append(predicted_obj_tokens[character_indices]) t2c_tokens_for_batch.append(predicted_t2c_tokens) return self._get_text_character_affinity_matrices( character_obj_tokens_for_batch=character_obj_tokens_for_batch, text_obj_tokens_for_this_batch=text_obj_tokens_for_batch, t2c_tokens_for_batch=t2c_tokens_for_batch, apply_sigmoid=True, ) # create callback fn def get_dialog_confidence_scores(batch_text_indices): predicted_obj_tokens_for_batch = self._get_predicted_obj_tokens(detection_transformer_output) dialog_confidence = [] for predicted_obj_tokens, text_indices in zip(predicted_obj_tokens_for_batch, batch_text_indices): confidence = self.is_this_text_a_dialogue(predicted_obj_tokens[text_indices]).sigmoid() dialog_confidence.append(rearrange(confidence, "i 1 -> i")) return dialog_confidence return self.processor.postprocess_detections_and_associations( predicted_bboxes=predicted_bboxes, predicted_class_scores=predicted_class_scores, original_image_sizes=torch.stack([torch.tensor(img.shape[:2]) for img in images], dim=0).to(predicted_bboxes.device), get_character_character_matching_scores=get_character_character_matching_scores, get_text_character_matching_scores=get_text_character_matching_scores, get_dialog_confidence_scores=get_dialog_confidence_scores, character_detection_threshold=character_detection_threshold, panel_detection_threshold=panel_detection_threshold, text_detection_threshold=text_detection_threshold, character_character_matching_threshold=character_character_matching_threshold, text_character_matching_threshold=text_character_matching_threshold, ) def predict_crop_embeddings(self, images, crop_bboxes, move_to_device_fn=None, mask_ratio=0.0, batch_size=256): if self.config.disable_crop_embeddings: return None assert isinstance(crop_bboxes, List), "please provide a list of bboxes for each image to get embeddings for" move_to_device_fn = self.move_to_device if move_to_device_fn is None else move_to_device_fn # temporarily change the mask ratio from default to the one specified old_mask_ratio = self.crop_embedding_model.embeddings.config.mask_ratio self.crop_embedding_model.embeddings.config.mask_ratio = mask_ratio crops_per_image = [] num_crops_per_batch = [len(bboxes) for bboxes in crop_bboxes] for image, bboxes, num_crops in zip(images, crop_bboxes, num_crops_per_batch): crops = self.processor.crop_image(image, bboxes) assert len(crops) == num_crops crops_per_image.extend(crops) if len(crops_per_image) == 0: return [[] for _ in crop_bboxes] crops_per_image = self.processor.preprocess_inputs_for_crop_embeddings(crops_per_image) crops_per_image = move_to_device_fn(crops_per_image) # process the crops in batches to avoid OOM embeddings = [] for i in range(0, len(crops_per_image), batch_size): crops = crops_per_image[i:i+batch_size] embeddings_per_batch = self.crop_embedding_model(crops).last_hidden_state[:, 0] embeddings.append(embeddings_per_batch) embeddings = torch.cat(embeddings, dim=0) crop_embeddings_for_batch = [] for num_crops in num_crops_per_batch: crop_embeddings_for_batch.append(embeddings[:num_crops]) embeddings = embeddings[num_crops:] # restore the mask ratio to the default self.crop_embedding_model.embeddings.config.mask_ratio = old_mask_ratio return crop_embeddings_for_batch def predict_ocr(self, images, crop_bboxes, move_to_device_fn=None, use_tqdm=False, batch_size=32): assert not self.config.disable_ocr move_to_device_fn = self.move_to_device if move_to_device_fn is None else move_to_device_fn crops_per_image = [] num_crops_per_batch = [len(bboxes) for bboxes in crop_bboxes] for image, bboxes, num_crops in zip(images, crop_bboxes, num_crops_per_batch): crops = self.processor.crop_image(image, bboxes) assert len(crops) == num_crops crops_per_image.extend(crops) if len(crops_per_image) == 0: return [[] for _ in crop_bboxes] crops_per_image = self.processor.preprocess_inputs_for_ocr(crops_per_image) crops_per_image = move_to_device_fn(crops_per_image) # process the crops in batches to avoid OOM all_generated_texts = [] if use_tqdm: from tqdm import tqdm pbar = tqdm(range(0, len(crops_per_image), batch_size)) else: pbar = range(0, len(crops_per_image), batch_size) for i in pbar: crops = crops_per_image[i:i+batch_size] generated_ids = self.ocr_model.generate(crops) generated_texts = self.processor.postprocess_ocr_tokens(generated_ids) all_generated_texts.extend(generated_texts) texts_for_images = [] for num_crops in num_crops_per_batch: texts_for_images.append([x.replace("\n", "") for x in all_generated_texts[:num_crops]]) all_generated_texts = all_generated_texts[num_crops:] return texts_for_images def visualise_single_image_prediction( self, image_as_np_array, predictions, filename=None ): return visualise_single_image_prediction(image_as_np_array, predictions, filename) def generate_transcript_for_single_image( self, predictions, ocr_results, filename=None ): character_clusters = predictions["character_cluster_labels"] text_to_character = predictions["text_character_associations"] text_to_character = {k: v for k, v in text_to_character} transript = " ### Transcript ###\n" for index, text in enumerate(ocr_results): if index in text_to_character: speaker = character_clusters[text_to_character[index]] speaker = f"<{speaker}>" else: speaker = "" transript += f"{speaker}: {text}\n" if filename is not None: with open(filename, "w") as file: file.write(transript) return transript def get_affinity_matrices_given_annotations( self, images, annotations, move_to_device_fn=None, apply_sigmoid=True ): assert not self.config.disable_detections move_to_device_fn = self.move_to_device if move_to_device_fn is None else move_to_device_fn character_bboxes_in_batch = [[bbox for bbox, label in zip(a["bboxes_as_x1y1x2y2"], a["labels"]) if label == 0] for a in annotations] crop_embeddings_for_batch = self.predict_crop_embeddings(images, character_bboxes_in_batch, move_to_device_fn) inputs_to_detection_transformer = self.processor.preprocess_inputs_for_detection(images, annotations) inputs_to_detection_transformer = move_to_device_fn(inputs_to_detection_transformer) processed_targets = inputs_to_detection_transformer.pop("labels") detection_transformer_output = self._get_detection_transformer_output(**inputs_to_detection_transformer) predicted_obj_tokens_for_batch = self._get_predicted_obj_tokens(detection_transformer_output) predicted_t2c_tokens_for_batch = self._get_predicted_t2c_tokens(detection_transformer_output) predicted_c2c_tokens_for_batch = self._get_predicted_c2c_tokens(detection_transformer_output) predicted_class_scores, predicted_bboxes = self._get_predicted_bboxes_and_classes(detection_transformer_output) matching_dict = { "logits": predicted_class_scores, "pred_boxes": predicted_bboxes, } indices = self.matcher(matching_dict, processed_targets) matched_char_obj_tokens_for_batch = [] matched_text_obj_tokens_for_batch = [] t2c_tokens_for_batch = [] c2c_tokens_for_batch = [] text_bboxes_for_batch = [] character_bboxes_for_batch = [] for j, (pred_idx, tgt_idx) in enumerate(indices): target_idx_to_pred_idx = {tgt.item(): pred.item() for pred, tgt in zip(pred_idx, tgt_idx)} targets_for_this_image = processed_targets[j] indices_of_text_boxes_in_annotation = [i for i, label in enumerate(targets_for_this_image["class_labels"]) if label == 1] indices_of_char_boxes_in_annotation = [i for i, label in enumerate(targets_for_this_image["class_labels"]) if label == 0] predicted_text_indices = [target_idx_to_pred_idx[i] for i in indices_of_text_boxes_in_annotation] predicted_char_indices = [target_idx_to_pred_idx[i] for i in indices_of_char_boxes_in_annotation] text_bboxes_for_batch.append( [annotations[j]["bboxes_as_x1y1x2y2"][k] for k in indices_of_text_boxes_in_annotation] ) character_bboxes_for_batch.append( [annotations[j]["bboxes_as_x1y1x2y2"][k] for k in indices_of_char_boxes_in_annotation] ) matched_char_obj_tokens_for_batch.append(predicted_obj_tokens_for_batch[j][predicted_char_indices]) matched_text_obj_tokens_for_batch.append(predicted_obj_tokens_for_batch[j][predicted_text_indices]) t2c_tokens_for_batch.append(predicted_t2c_tokens_for_batch[j]) c2c_tokens_for_batch.append(predicted_c2c_tokens_for_batch[j]) text_character_affinity_matrices = self._get_text_character_affinity_matrices( character_obj_tokens_for_batch=matched_char_obj_tokens_for_batch, text_obj_tokens_for_this_batch=matched_text_obj_tokens_for_batch, t2c_tokens_for_batch=t2c_tokens_for_batch, apply_sigmoid=apply_sigmoid, ) character_character_affinity_matrices = self._get_character_character_affinity_matrices( character_obj_tokens_for_batch=matched_char_obj_tokens_for_batch, crop_embeddings_for_batch=crop_embeddings_for_batch, c2c_tokens_for_batch=c2c_tokens_for_batch, apply_sigmoid=apply_sigmoid, ) return { "text_character_affinity_matrices": text_character_affinity_matrices, "character_character_affinity_matrices": character_character_affinity_matrices, "text_bboxes_for_batch": text_bboxes_for_batch, "character_bboxes_for_batch": character_bboxes_for_batch, } def get_obj_embeddings_corresponding_to_given_annotations( self, images, annotations, move_to_device_fn=None ): assert not self.config.disable_detections move_to_device_fn = self.move_to_device if move_to_device_fn is None else move_to_device_fn inputs_to_detection_transformer = self.processor.preprocess_inputs_for_detection(images, annotations) inputs_to_detection_transformer = move_to_device_fn(inputs_to_detection_transformer) processed_targets = inputs_to_detection_transformer.pop("labels") detection_transformer_output = self._get_detection_transformer_output(**inputs_to_detection_transformer) predicted_obj_tokens_for_batch = self._get_predicted_obj_tokens(detection_transformer_output) predicted_t2c_tokens_for_batch = self._get_predicted_t2c_tokens(detection_transformer_output) predicted_c2c_tokens_for_batch = self._get_predicted_c2c_tokens(detection_transformer_output) predicted_class_scores, predicted_bboxes = self._get_predicted_bboxes_and_classes(detection_transformer_output) matching_dict = { "logits": predicted_class_scores, "pred_boxes": predicted_bboxes, } indices = self.matcher(matching_dict, processed_targets) matched_char_obj_tokens_for_batch = [] matched_text_obj_tokens_for_batch = [] matched_panel_obj_tokens_for_batch = [] t2c_tokens_for_batch = [] c2c_tokens_for_batch = [] for j, (pred_idx, tgt_idx) in enumerate(indices): target_idx_to_pred_idx = {tgt.item(): pred.item() for pred, tgt in zip(pred_idx, tgt_idx)} targets_for_this_image = processed_targets[j] indices_of_char_boxes_in_annotation = [i for i, label in enumerate(targets_for_this_image["class_labels"]) if label == 0] indices_of_text_boxes_in_annotation = [i for i, label in enumerate(targets_for_this_image["class_labels"]) if label == 1] indices_of_panel_boxes_in_annotation = [i for i, label in enumerate(targets_for_this_image["class_labels"]) if label == 2] predicted_text_indices = [target_idx_to_pred_idx[i] for i in indices_of_text_boxes_in_annotation] predicted_char_indices = [target_idx_to_pred_idx[i] for i in indices_of_char_boxes_in_annotation] predicted_panel_indices = [target_idx_to_pred_idx[i] for i in indices_of_panel_boxes_in_annotation] matched_char_obj_tokens_for_batch.append(predicted_obj_tokens_for_batch[j][predicted_char_indices]) matched_text_obj_tokens_for_batch.append(predicted_obj_tokens_for_batch[j][predicted_text_indices]) matched_panel_obj_tokens_for_batch.append(predicted_obj_tokens_for_batch[j][predicted_panel_indices]) t2c_tokens_for_batch.append(predicted_t2c_tokens_for_batch[j]) c2c_tokens_for_batch.append(predicted_c2c_tokens_for_batch[j]) return { "character": matched_char_obj_tokens_for_batch, "text": matched_text_obj_tokens_for_batch, "panel": matched_panel_obj_tokens_for_batch, "t2c": t2c_tokens_for_batch, "c2c": c2c_tokens_for_batch, } def sort_panels_and_text_bboxes_in_reading_order( self, batch_panel_bboxes, batch_text_bboxes, ): batch_sorted_panel_indices = [] batch_sorted_text_indices = [] for batch_index in range(len(batch_text_bboxes)): panel_bboxes = batch_panel_bboxes[batch_index] text_bboxes = batch_text_bboxes[batch_index] sorted_panel_indices = sort_panels(panel_bboxes) sorted_panels = [panel_bboxes[i] for i in sorted_panel_indices] sorted_text_indices = sort_text_boxes_in_reading_order(text_bboxes, sorted_panels) batch_sorted_panel_indices.append(sorted_panel_indices) batch_sorted_text_indices.append(sorted_text_indices) return batch_sorted_panel_indices, batch_sorted_text_indices def _get_detection_transformer_output( self, pixel_values: torch.FloatTensor, pixel_mask: Optional[torch.LongTensor] = None ): if self.config.disable_detections: raise ValueError("Detection model is disabled. Set disable_detections=False in the config.") return self.detection_transformer( pixel_values=pixel_values, pixel_mask=pixel_mask, return_dict=True ) def _get_predicted_obj_tokens( self, detection_transformer_output: ConditionalDetrModelOutput ): return detection_transformer_output.last_hidden_state[:, :-self.num_non_obj_tokens] def _get_predicted_c2c_tokens( self, detection_transformer_output: ConditionalDetrModelOutput ): return detection_transformer_output.last_hidden_state[:, -self.num_non_obj_tokens] def _get_predicted_t2c_tokens( self, detection_transformer_output: ConditionalDetrModelOutput ): return detection_transformer_output.last_hidden_state[:, -self.num_non_obj_tokens+1] def _get_predicted_bboxes_and_classes( self, detection_transformer_output: ConditionalDetrModelOutput, ): if self.config.disable_detections: raise ValueError("Detection model is disabled. Set disable_detections=False in the config.") obj = self._get_predicted_obj_tokens(detection_transformer_output) predicted_class_scores = self.class_labels_classifier(obj) reference = detection_transformer_output.reference_points[:-self.num_non_obj_tokens] reference_before_sigmoid = inverse_sigmoid(reference).transpose(0, 1) predicted_boxes = self.bbox_predictor(obj) predicted_boxes[..., :2] += reference_before_sigmoid predicted_boxes = predicted_boxes.sigmoid() return predicted_class_scores, predicted_boxes def _get_character_character_affinity_matrices( self, character_obj_tokens_for_batch: List[torch.FloatTensor] = None, crop_embeddings_for_batch: List[torch.FloatTensor] = None, c2c_tokens_for_batch: List[torch.FloatTensor] = None, apply_sigmoid=True, ): assert self.config.disable_detections or (character_obj_tokens_for_batch is not None and c2c_tokens_for_batch is not None) assert self.config.disable_crop_embeddings or crop_embeddings_for_batch is not None assert not self.config.disable_detections or not self.config.disable_crop_embeddings if self.config.disable_detections: affinity_matrices = [] for crop_embeddings in crop_embeddings_for_batch: crop_embeddings = crop_embeddings / crop_embeddings.norm(dim=-1, keepdim=True) affinity_matrix = crop_embeddings @ crop_embeddings.T affinity_matrices.append(affinity_matrix) return affinity_matrices affinity_matrices = [] for batch_index, (character_obj_tokens, c2c) in enumerate(zip(character_obj_tokens_for_batch, c2c_tokens_for_batch)): if character_obj_tokens.shape[0] == 0: affinity_matrices.append(torch.zeros(0, 0).type_as(character_obj_tokens)) continue if not self.config.disable_crop_embeddings: crop_embeddings = crop_embeddings_for_batch[batch_index] assert character_obj_tokens.shape[0] == crop_embeddings.shape[0] character_obj_tokens = torch.cat([character_obj_tokens, crop_embeddings], dim=-1) char_i = repeat(character_obj_tokens, "i d -> i repeat d", repeat=character_obj_tokens.shape[0]) char_j = repeat(character_obj_tokens, "j d -> repeat j d", repeat=character_obj_tokens.shape[0]) char_ij = rearrange([char_i, char_j], "two i j d -> (i j) (two d)") c2c = repeat(c2c, "d -> repeat d", repeat = char_ij.shape[0]) char_ij_c2c = torch.cat([char_ij, c2c], dim=-1) character_character_affinities = self.character_character_matching_head(char_ij_c2c) character_character_affinities = rearrange(character_character_affinities, "(i j) 1 -> i j", i=char_i.shape[0]) character_character_affinities = (character_character_affinities + character_character_affinities.T) / 2 if apply_sigmoid: character_character_affinities = character_character_affinities.sigmoid() affinity_matrices.append(character_character_affinities) return affinity_matrices def _get_text_character_affinity_matrices( self, character_obj_tokens_for_batch: List[torch.FloatTensor] = None, text_obj_tokens_for_this_batch: List[torch.FloatTensor] = None, t2c_tokens_for_batch: List[torch.FloatTensor] = None, apply_sigmoid=True, ): assert not self.config.disable_detections assert character_obj_tokens_for_batch is not None and text_obj_tokens_for_this_batch is not None and t2c_tokens_for_batch is not None affinity_matrices = [] for character_obj_tokens, text_obj_tokens, t2c in zip(character_obj_tokens_for_batch, text_obj_tokens_for_this_batch, t2c_tokens_for_batch): if character_obj_tokens.shape[0] == 0 or text_obj_tokens.shape[0] == 0: affinity_matrices.append(torch.zeros(text_obj_tokens.shape[0], character_obj_tokens.shape[0]).type_as(character_obj_tokens)) continue text_i = repeat(text_obj_tokens, "i d -> i repeat d", repeat=character_obj_tokens.shape[0]) char_j = repeat(character_obj_tokens, "j d -> repeat j d", repeat=text_obj_tokens.shape[0]) text_char = rearrange([text_i, char_j], "two i j d -> (i j) (two d)") t2c = repeat(t2c, "d -> repeat d", repeat = text_char.shape[0]) text_char_t2c = torch.cat([text_char, t2c], dim=-1) text_character_affinities = self.text_character_matching_head(text_char_t2c) text_character_affinities = rearrange(text_character_affinities, "(i j) 1 -> i j", i=text_i.shape[0]) if apply_sigmoid: text_character_affinities = text_character_affinities.sigmoid() affinity_matrices.append(text_character_affinities) return affinity_matrices