# -*- coding: utf-8 -*- import math import re from dataclasses import dataclass from typing import Optional, Tuple, List, Union, Dict import numpy as np import torch import torch.nn as nn from transformers import ErnieModel, ErniePreTrainedModel, PretrainedConfig, PreTrainedTokenizerFast from transformers.utils import ModelOutput @dataclass class UIEModelOutput(ModelOutput): """ Output class for outputs of UIE. Args: loss (`torch.FloatTensor` of shape `(1),`, *optional*, returned when `labels` is provided): Total span extraction loss is the sum of a Cross-Entropy for the start and end positions. start_prob (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): Span-start scores (after Sigmoid). end_prob (`torch.FloatTensor` of shape `(batch_size, sequence_length)`): Span-end scores (after Sigmoid). hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. """ loss: Optional[torch.FloatTensor] = None start_prob: torch.FloatTensor = None end_prob: torch.FloatTensor = None hidden_states: Optional[Tuple[torch.FloatTensor]] = None attentions: Optional[Tuple[torch.FloatTensor]] = None class UIE(ErniePreTrainedModel): """ UIE model based on Bert model. This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`PretrainedConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. """ def __init__(self, config: PretrainedConfig): super(UIE, self).__init__(config) self.encoder = ErnieModel(config) self.config = config hidden_size = self.config.hidden_size self.linear_start = nn.Linear(hidden_size, 1) self.linear_end = nn.Linear(hidden_size, 1) self.sigmoid = nn.Sigmoid() self.post_init() def forward(self, input_ids: Optional[torch.Tensor] = None, token_type_ids: Optional[torch.Tensor] = None, position_ids: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, head_mask: Optional[torch.Tensor] = None, inputs_embeds: Optional[torch.Tensor] = None, start_positions: Optional[torch.Tensor] = None, end_positions: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None ): """ Args: input_ids (`torch.LongTensor` of shape `({0})`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`BertTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*): Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - 0 corresponds to a *sentence A* token, - 1 corresponds to a *sentence B* token. [What are token type IDs?](../glossary#token-type-ids) position_ids (`torch.LongTensor` of shape `({0})`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict outputs = self.encoder( input_ids=input_ids, token_type_ids=token_type_ids, position_ids=position_ids, attention_mask=attention_mask, head_mask=head_mask, inputs_embeds=inputs_embeds, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict ) sequence_output = outputs[0] start_logits = self.linear_start(sequence_output) start_logits = torch.squeeze(start_logits, -1) start_prob = self.sigmoid(start_logits) end_logits = self.linear_end(sequence_output) end_logits = torch.squeeze(end_logits, -1) end_prob = self.sigmoid(end_logits) total_loss = None if start_positions is not None and end_positions is not None: loss_fct = nn.BCELoss() start_loss = loss_fct(start_prob, start_positions) end_loss = loss_fct(end_prob, end_positions) total_loss = (start_loss + end_loss) / 2.0 if not return_dict: output = (start_prob, end_prob) + outputs[2:] return ((total_loss,) + output) if total_loss is not None else output return UIEModelOutput( loss=total_loss, start_prob=start_prob, end_prob=end_prob, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def predict(self, schema: Union[Dict, List[str], str], input_texts: Union[List[str], str], tokenizer: PreTrainedTokenizerFast, max_length: int = 512, batch_size: int = 32, position_prob: int = 0.5, progress_hook=None) -> List[Dict]: """ Args: schema (Union[Dict, List[str], str]): 抽取目标 input_texts (input_texts: Union[List[str], str]): 待抽取文本 tokenizer (PreTrainedTokenizerFast): max_length (int): batch_size (int): position_prob (float): progress_hook: Returns: result (List[Dict]): """ predictor = UIEPredictor(self, tokenizer=tokenizer, schema=schema, max_length=max_length, position_prob=position_prob, batch_size=batch_size, hook=progress_hook) input_texts = [input_texts] if isinstance(input_texts, str) else input_texts return predictor.predict(input_texts) class UIEPredictor(object): def __init__(self, model, tokenizer, schema, max_length=512, position_prob=0.5, batch_size=32, hook=None): self.model = model self._tokenizer = tokenizer self._position_prob = position_prob self.max_length = max_length self._batch_size = batch_size self._multilingual = getattr(self.model.config, 'multilingual', False) self._schema_tree = self.set_schema(schema) self._hook = hook def set_schema(self, schema): if isinstance(schema, dict) or isinstance(schema, str): schema = [schema] return self._build_tree(schema) @classmethod def _build_tree(cls, schema, name="root"): """ Build the schema tree. """ schema_tree = SchemaTree(name) for s in schema: if isinstance(s, str): schema_tree.add_child(SchemaTree(s)) elif isinstance(s, dict): for k, v in s.items(): if isinstance(v, str): child = [v] elif isinstance(v, list): child = v else: raise TypeError( "Invalid schema, value for each key:value pairs should be list or string" "but {} received".format(type(v)) ) schema_tree.add_child(cls._build_tree(child, name=k)) else: raise TypeError("Invalid schema, element should be string or dict, " "but {} received".format(type(s))) return schema_tree def _single_stage_predict(self, inputs): input_texts = [] prompts = [] for i in range(len(inputs)): input_texts.append(inputs[i]["text"]) prompts.append(inputs[i]["prompt"]) # max predict length should exclude the length of prompt and summary tokens max_predict_len = self.max_length - len(max(prompts)) - 3 short_input_texts, self.input_mapping = Utils.auto_splitter(input_texts, max_predict_len, split_sentence=False) short_texts_prompts = [] for k, v in self.input_mapping.items(): short_texts_prompts.extend([prompts[k] for _ in range(len(v))]) short_inputs = [ {"text": short_input_texts[i], "prompt": short_texts_prompts[i]} for i in range(len(short_input_texts)) ] prompts = [] texts = [] for s in short_inputs: prompts.append(s["prompt"]) texts.append(s["text"]) if self._multilingual: padding_type = "max_length" else: padding_type = "longest" encoded_inputs = self._tokenizer( text=prompts, text_pair=texts, stride=2, truncation=True, max_length=self.max_length, padding=padding_type, add_special_tokens=True, return_offsets_mapping=True, return_tensors="np") offset_maps = encoded_inputs["offset_mapping"] start_probs = [] end_probs = [] for idx in range(0, len(texts), self._batch_size): l, r = idx, idx + self._batch_size input_ids = encoded_inputs["input_ids"][l:r] token_type_ids = encoded_inputs["token_type_ids"][l:r] attention_mask = encoded_inputs["attention_mask"][l:r] if self._multilingual: input_ids = np.array( input_ids, dtype="int64") attention_mask = np.array( attention_mask, dtype="int64") position_ids = (np.cumsum(np.ones_like(input_ids), axis=1) - np.ones_like(input_ids)) * attention_mask input_dict = { "input_ids": input_ids, "attention_mask": attention_mask, "position_ids": position_ids } else: input_dict = { "input_ids": np.array( input_ids, dtype="int64"), "token_type_ids": np.array( token_type_ids, dtype="int64"), "attention_mask": np.array( attention_mask, dtype="int64") } start_prob, end_prob = self._infer(input_dict) start_prob = start_prob.tolist() end_prob = end_prob.tolist() start_probs.extend(start_prob) end_probs.extend(end_prob) if self._hook is not None: self._hook.update(1) start_ids_list = Utils.get_bool_ids_greater_than(start_probs, limit=self._position_prob, return_prob=True) end_ids_list = Utils.get_bool_ids_greater_than(end_probs, limit=self._position_prob, return_prob=True) sentence_ids = [] probs = [] for start_ids, end_ids, offset_map in zip(start_ids_list, end_ids_list, offset_maps.tolist()): span_list = Utils.get_span(start_ids, end_ids, with_prob=True) sentence_id, prob = Utils.get_id_and_prob(span_list, offset_map) sentence_ids.append(sentence_id) probs.append(prob) results = Utils.convert_ids_to_results(short_inputs, sentence_ids, probs) results = Utils.auto_joiner(results, short_input_texts, self.input_mapping) return results def _multi_stage_predict(self, data): """ Traversal the schema tree and do multi-stage prediction. Args: data (list): a list of strings Returns: list: a list of predictions, where the list's length equals to the length of `data` """ results = [{} for _ in range(len(data))] # input check to early return if len(data) < 1 or self._schema_tree is None: return results _pre_node_total = len(data) // self._batch_size + (1 if len(data) % self._batch_size else 0) _finish_node = 0 if self._hook is not None: self._hook.reset(total=self._schema_tree.shape * _pre_node_total) # copy to stay `self._schema_tree` unchanged schema_list = self._schema_tree.children[:] while len(schema_list) > 0: node = schema_list.pop(0) examples = [] input_map = {} cnt = 0 idx = 0 if not node.prefix: for one_data in data: examples.append({"text": one_data, "prompt": Utils.dbc2sbc(node.name)}) input_map[cnt] = [idx] idx += 1 cnt += 1 else: for pre, one_data in zip(node.prefix, data): if len(pre) == 0: input_map[cnt] = [] else: for p in pre: examples.append({"text": one_data, "prompt": Utils.dbc2sbc(p + node.name)}) input_map[cnt] = [i + idx for i in range(len(pre))] idx += len(pre) cnt += 1 if len(examples) == 0: result_list = [] else: result_list = self._single_stage_predict(examples) if not node.parent_relations: relations = [[] for _ in range(len(data))] for k, v in input_map.items(): for idx in v: if len(result_list[idx]) == 0: continue if node.name not in results[k].keys(): results[k][node.name] = result_list[idx] else: results[k][node.name].extend(result_list[idx]) if node.name in results[k].keys(): relations[k].extend(results[k][node.name]) else: relations = node.parent_relations for k, v in input_map.items(): for i in range(len(v)): if len(result_list[v[i]]) == 0: continue if "relations" not in relations[k][i].keys(): relations[k][i]["relations"] = {node.name: result_list[v[i]]} elif node.name not in relations[k][i]["relations"].keys(): relations[k][i]["relations"][node.name] = result_list[v[i]] else: relations[k][i]["relations"][node.name].extend(result_list[v[i]]) new_relations = [[] for _ in range(len(data))] for i in range(len(relations)): for j in range(len(relations[i])): if "relations" in relations[i][j].keys() and node.name in relations[i][j]["relations"].keys(): for k in range(len(relations[i][j]["relations"][node.name])): new_relations[i].append(relations[i][j]["relations"][node.name][k]) relations = new_relations prefix = [[] for _ in range(len(data))] for k, v in input_map.items(): for idx in v: for i in range(len(result_list[idx])): prefix[k].append(result_list[idx][i]["text"] + "的") for child in node.children: child.prefix = prefix child.parent_relations = relations schema_list.append(child) _finish_node += 1 if self._hook is not None: self._hook.n = _finish_node * _pre_node_total if self._hook is not None: self._hook.close() return results def _infer(self, input_dict): for input_name, input_value in input_dict.items(): input_dict[input_name] = torch.LongTensor(input_value).to(self.model.device) outputs = self.model(**input_dict) return outputs.start_prob.detach().cpu().numpy(), outputs.end_prob.detach().cpu().numpy() def predict(self, input_data): results = self._multi_stage_predict(data=input_data) return results class SchemaTree(object): """ Implementataion of SchemaTree """ def __init__(self, name="root", children=None): self.name = name self.children = [] self.prefix = None self.parent_relations = None if children is not None: for child in children: self.add_child(child) self._total_nodes = 0 @property def shape(self): return len(self.children) + sum([child.shape for child in self.children]) def __repr__(self): return self.name def add_child(self, node): assert isinstance(node, SchemaTree), "The children of a node should be an instacne of SchemaTree." self._total_nodes += 1 self.children.append(node) class Utils: @classmethod def dbc2sbc(cls, s): rs = "" for char in s: code = ord(char) if code == 0x3000: code = 0x0020 else: code -= 0xFEE0 if not (0x0021 <= code <= 0x7E): rs += char continue rs += chr(code) return rs @classmethod def cut_chinese_sent(cls, para): """ Cut the Chinese sentences more precisely, reference to "https://blog.csdn.net/blmoistawinde/article/details/82379256". """ para = re.sub(r'([。！？?])([^”’])', r"\1\n\2", para) # 单字符断句符 para = re.sub(r'(\.{6})([^”’])', r"\1\n\2", para) # 英文省略号 para = re.sub(r'(…{2})([^”’])', r"\1\n\2", para) # 中文省略号 para = re.sub(r'([。！？?][”’])([^，。！？?])', r'\1\n\2', para) para = para.rstrip() return para.split("\n") @classmethod def get_bool_ids_greater_than(cls, probs, limit=0.5, return_prob=False): """ Get idx of the last dimension in probability arrays, which is greater than a limitation. Args: probs (List[List[float]]): The input probability arrays. limit (float): The limitation for probability. return_prob (bool): Whether to return the probability Returns: List[List[int]]: The index of the last dimension meet the conditions. """ probs = np.array(probs) dim_len = len(probs.shape) if dim_len > 1: result = [] for p in probs: result.append(cls.get_bool_ids_greater_than(p, limit, return_prob)) return result else: result = [] for i, p in enumerate(probs): if p > limit: if return_prob: result.append((i, p)) else: result.append(i) return result @classmethod def get_span(cls, start_ids, end_ids, with_prob=False): """ Get span set from position start and end list. Args: start_ids (List[int]/List[tuple]): The start index list. end_ids (List[int]/List[tuple]): The end index list. with_prob (bool): If True, each element for start_ids and end_ids is a tuple as like: (index, probability). Returns: set: The span set without overlapping, every id can only be used once . """ if with_prob: start_ids = sorted(start_ids, key=lambda x: x[0]) end_ids = sorted(end_ids, key=lambda x: x[0]) else: start_ids = sorted(start_ids) end_ids = sorted(end_ids) start_pointer = 0 end_pointer = 0 len_start = len(start_ids) len_end = len(end_ids) couple_dict = {} while start_pointer < len_start and end_pointer < len_end: if with_prob: start_id = start_ids[start_pointer][0] end_id = end_ids[end_pointer][0] else: start_id = start_ids[start_pointer] end_id = end_ids[end_pointer] if start_id == end_id: couple_dict[end_ids[end_pointer]] = start_ids[start_pointer] start_pointer += 1 end_pointer += 1 continue if start_id < end_id: couple_dict[end_ids[end_pointer]] = start_ids[start_pointer] start_pointer += 1 continue if start_id > end_id: end_pointer += 1 continue result = [(couple_dict[end], end) for end in couple_dict] result = set(result) return result @classmethod def get_id_and_prob(cls, span_set, offset_mapping: np.array): """ Return text id and probability of predicted spans Args: span_set (set): set of predicted spans. offset_mapping (numpy.array): list of pair preserving the index of start and end char in original text pair (prompt + text) for each token. Returns: sentence_id (list[tuple]): index of start and end char in original text. prob (list[float]): probabilities of predicted spans. """ prompt_end_token_id = offset_mapping[1:].index([0, 0]) bias = offset_mapping[prompt_end_token_id][1] + 1 for index in range(1, prompt_end_token_id + 1): offset_mapping[index][0] -= bias offset_mapping[index][1] -= bias sentence_id = [] prob = [] for start, end in span_set: prob.append(start[1] * end[1]) start_id = offset_mapping[start[0]][0] end_id = offset_mapping[end[0]][1] sentence_id.append((start_id, end_id)) return sentence_id, prob @classmethod def auto_splitter(cls, input_texts, max_text_len, split_sentence=False): """ Split the raw texts automatically for model inference. Args: input_texts (List[str]): input raw texts. max_text_len (int): cutting length. split_sentence (bool): If True, sentence-level split will be performed. return: short_input_texts (List[str]): the short input texts for model inference. input_mapping (dict): mapping between raw text and short input texts. """ input_mapping = {} short_input_texts = [] cnt_org = 0 cnt_short = 0 for text in input_texts: if not split_sentence: sens = [text] else: sens = Utils.cut_chinese_sent(text) for sen in sens: lens = len(sen) if lens <= max_text_len: short_input_texts.append(sen) if cnt_org not in input_mapping.keys(): input_mapping[cnt_org] = [cnt_short] else: input_mapping[cnt_org].append(cnt_short) cnt_short += 1 else: temp_text_list = [sen[i: i + max_text_len] for i in range(0, lens, max_text_len)] short_input_texts.extend(temp_text_list) short_idx = cnt_short cnt_short += math.ceil(lens / max_text_len) temp_text_id = [short_idx + i for i in range(cnt_short - short_idx)] if cnt_org not in input_mapping.keys(): input_mapping[cnt_org] = temp_text_id else: input_mapping[cnt_org].extend(temp_text_id) cnt_org += 1 return short_input_texts, input_mapping @classmethod def convert_ids_to_results(cls, examples, sentence_ids, probs): """ Convert ids to raw text in a single stage. """ results = [] for example, sentence_id, prob in zip(examples, sentence_ids, probs): if len(sentence_id) == 0: results.append([]) continue result_list = [] text = example["text"] prompt = example["prompt"] for i in range(len(sentence_id)): start, end = sentence_id[i] if start < 0 and end >= 0: continue if end < 0: start += len(prompt) + 1 end += len(prompt) + 1 result = {"text": prompt[start:end], "probability": prob[i]} result_list.append(result) else: result = {"text": text[start:end], "start": start, "end": end, "probability": prob[i]} result_list.append(result) results.append(result_list) return results @classmethod def auto_joiner(cls, short_results, short_inputs, input_mapping): concat_results = [] is_cls_task = False for short_result in short_results: if not short_result: continue elif "start" not in short_result[0].keys() and "end" not in short_result[0].keys(): is_cls_task = True break else: break for k, vs in input_mapping.items(): if is_cls_task: cls_options = {} for v in vs: if len(short_results[v]) == 0: continue if short_results[v][0]["text"] not in cls_options.keys(): cls_options[short_results[v][0]["text"]] = [1, short_results[v][0]["probability"]] else: cls_options[short_results[v][0]["text"]][0] += 1 cls_options[short_results[v][0]["text"]][1] += short_results[v][0]["probability"] if len(cls_options) != 0: cls_res, cls_info = max(cls_options.items(), key=lambda x: x[1]) concat_results.append([{"text": cls_res, "probability": cls_info[1] / cls_info[0]}]) else: concat_results.append([]) else: offset = 0 single_results = [] for v in vs: if v == 0: single_results = short_results[v] offset += len(short_inputs[v]) else: for i in range(len(short_results[v])): if "start" not in short_results[v][i] or "end" not in short_results[v][i]: continue short_results[v][i]["start"] += offset short_results[v][i]["end"] += offset offset += len(short_inputs[v]) single_results.extend(short_results[v]) concat_results.append(single_results) return concat_results