文章目錄

• • 1. 導(dǎo)包
  • 2. 數(shù)據(jù)處理
  • 3. 輔助函數(shù)
  • • 3.1 評(píng)估函數(shù)
    • 3.2 預(yù)測(cè)函數(shù)
    • 3.3 預(yù)測(cè)結(jié)果解碼
  • 4. 訓(xùn)練

填寫快遞單據(jù)可以直接把所有信息直接粘貼進(jìn)客戶端，客戶端自動(dòng)識(shí)別 省市、人名、電話等信息，分類填入，然后打印出來粘貼。無須人工填寫，加快了作業(yè)效率。

learn from : https://aistudio.baidu.com/aistudio/projectdetail/1329361

通過使用預(yù)訓(xùn)練模型+finetune，訓(xùn)練一個(gè)快遞信息抽取模型。

1. 導(dǎo)包

# 快遞單信息抽取 from functools import partial # 打包函數(shù)，并給定默認(rèn)參數(shù) import paddle from paddlenlp.datasets import MapDataset # 自定義數(shù)據(jù)集 from paddlenlp.data import Stack, Tuple, Pad # batch化工具函數(shù) from paddlenlp.transformers import ErnieTokenizer, ErnieForTokenClassification from paddlenlp.metrics import ChunkEvaluator # 指標(biāo)計(jì)算 from paddle.utils.download import get_path_from_url

2. 數(shù)據(jù)處理

URL = "https://paddlenlp.bj.bcebos.com/paddlenlp/datasets/waybill.tar.gz" get_path_from_url(URL, "./") epochs = 10 batch_size = 16def load_dict(dict_path): # 讀取字典vocab = {}i = 0for line in open(dict_path, 'r', encoding='utf-8'):key = line.strip('\n')vocab[key] = ii += 1return vocab# 展示下數(shù)據(jù)格式 with open("./data/test.txt", 'r', encoding='utf-8') as f:i = 0for line in f:print(line)i += 1if i > 5:break# text_a label # # 黑龍江省雙鴨山市尖山區(qū)八馬路與東平行路交叉口北40米韋業(yè)濤18600009172 # A1-BA1-IA1-IA1-IA2-BA2-IA2-IA2-IA3-BA3-IA3-IA4-BA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IP-BP-IP-IT-BT-IT-IT-IT-IT-IT-IT-IT-IT-IT-I # A1 表示省，-B 開始， -I 內(nèi)部， P 人名， T 電話

• 數(shù)據(jù)轉(zhuǎn)換函數(shù)，把文字轉(zhuǎn)成數(shù)字 ids 類型

def convert_example(example, tokenizer, label_vocab):tokens, labels = exampletokenized_input = tokenizer(tokens, return_length=True, is_split_into_words=True)# Token '[CLS]' and '[SEP]' will get label 'O'labels = ['O'] + labels + ['O'] # 大寫的字母 O（歐）tokenized_input['labels'] = [label_vocab[x] for x in labels]return tokenized_input['input_ids'], tokenized_input['token_type_ids'], \tokenized_input['seq_len'], tokenized_input['labels']# 轉(zhuǎn)成數(shù)字list

• 加載數(shù)據(jù)集

def load_dataset(datafiles):def read(data_path):with open(data_path, 'r', encoding='utf-8') as fp:next(fp) # Skip headerfor line in fp.readlines():words, labels = line.strip('\n').split('\t')words = words.split('\002')# ['1', '6', '6', '2', '0', '2', '0', '0', '0', '7', '7',# '宣', '榮', '嗣',# '甘', '肅', '省',# '白', '銀', '市',# '會(huì)', '寧', '縣',# '河', '畔', '鎮(zhèn)', '十', '字', '街', '金', '海', '超', '市', '西', '行', '5', '0', '米']labels = labels.split('\002')# ['T-B', 'T-I', 'T-I', 'T-I', 'T-I', 'T-I', 'T-I', 'T-I', 'T-I', 'T-I', 'T-I',# 'P-B', 'P-I', 'P-I',# 'A1-B', 'A1-I', 'A1-I',# 'A2-B', 'A2-I', 'A2-I',# 'A3-B', 'A3-I', 'A3-I',# 'A4-B', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I', 'A4-I']yield words, labelsif isinstance(datafiles, str):return MapDataset(list(read(datafiles)))elif isinstance(datafiles, list) or isinstance(datafiles, tuple):return [MapDataset(list(read(datafile))) for datafile in datafiles]# Create dataset, tokenizer and dataloader. train_ds, dev_ds, test_ds = load_dataset(datafiles=('./data/train.txt', './data/dev.txt', './data/test.txt'))

• batch化數(shù)據(jù)

label_vocab = load_dict('./data/tag.dic') # {'P-B': 0, 'P-I': 1, 'T-B': 2, 'T-I': 3, 'A1-B': 4, 'A1-I': 5, # 'A2-B': 6, 'A2-I': 7, 'A3-B': 8, 'A3-I': 9, 'A4-B': 10, 'A4-I': 11, 'O': 12}tokenizer = ErnieTokenizer.from_pretrained('ernie-1.0')trans_func = partial(convert_example, tokenizer=tokenizer, label_vocab=label_vocab)train_ds.map(trans_func) dev_ds.map(trans_func) test_ds.map(trans_func) print(train_ds[0]) # ([1, 208, 515, 515, 249, 540, 249, 540, 540, 540, 589, 589, 803, 838, 2914, 1222, 1734, 244, 368, 797, 99, 32, 863, 308, 457, 2778, 484, 167, 436, 930, 192, 233, 634, 99, 213, 40, 317, 540, 256, 2], # [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], # 40, # [12, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 0, 1, 1, 4, 5, 5, 6, 7, 7, 8, 9, 9, 10, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 12])ignore_label = -1# batch化函數(shù) batchify_fn = lambda samples, fn=Tuple(Pad(axis=0, pad_val=tokenizer.pad_token_id), # input_idsPad(axis=0, pad_val=tokenizer.pad_token_type_id), # token_type_idsStack(dtype='int64'), # seq_lenPad(axis=0, pad_val=ignore_label, dtype='int64') # labels ): fn(samples)# data_loader train_loader = paddle.io.DataLoader(dataset=train_ds,batch_size=batch_size,return_list=True,collate_fn=batchify_fn) dev_loader = paddle.io.DataLoader(dataset=dev_ds,batch_size=batch_size,return_list=True,collate_fn=batchify_fn) test_loader = paddle.io.DataLoader(dataset=test_ds,batch_size=batch_size,return_list=True,collate_fn=batchify_fn)

3. 輔助函數(shù)

3.1 評(píng)估函數(shù)

@paddle.no_grad() def evaluate(model, metric, data_loader):model.eval()metric.reset()for input_ids, seg_ids, lens, labels in data_loader:logits = model(input_ids, seg_ids)preds = paddle.argmax(logits, axis=-1)n_infer, n_label, n_correct = metric.compute(None, lens, preds, labels)metric.update(n_infer.numpy(), n_label.numpy(), n_correct.numpy())precision, recall, f1_score = metric.accumulate()print("eval precision: %f - recall: %f - f1: %f" %(precision, recall, f1_score))model.train()

3.2 預(yù)測(cè)函數(shù)

def predict(model, data_loader, ds, label_vocab):pred_list = []len_list = []for input_ids, seg_ids, lens, labels in data_loader:logits = model(input_ids, seg_ids)pred = paddle.argmax(logits, axis=-1)pred_list.append(pred.numpy())len_list.append(lens.numpy())preds = parse_decodes(ds, pred_list, len_list, label_vocab)return preds

3.3 預(yù)測(cè)結(jié)果解碼

def parse_decodes(ds, decodes, lens, label_vocab):decodes = [x for batch in decodes for x in batch]lens = [x for batch in lens for x in batch]id_label = dict(zip(label_vocab.values(), label_vocab.keys()))outputs = []for idx, end in enumerate(lens):sent = ds.data[idx][0][:end]tags = [id_label[x] for x in decodes[idx][1:end]]sent_out = []tags_out = []words = ""for s, t in zip(sent, tags):if t.endswith('-B') or t == 'O':if len(words):sent_out.append(words)tags_out.append(t.split('-')[0])words = selse:words += sif len(sent_out) < len(tags_out):sent_out.append(words)outputs.append(''.join([str((s, t)) for s, t in zip(sent_out, tags_out)]))return outputs

4. 訓(xùn)練

# 加載預(yù)訓(xùn)練模型 model = ErnieForTokenClassification.from_pretrained("ernie-1.0", num_classes=len(label_vocab)) # 指標(biāo) metric = ChunkEvaluator(label_list=label_vocab.keys(), suffix=True) # 損失函數(shù) loss_fn = paddle.nn.loss.CrossEntropyLoss(ignore_index=ignore_label) # 優(yōu)化器 optimizer = paddle.optimizer.AdamW(learning_rate=2e-5, parameters=model.parameters())# 訓(xùn)練 step = 0 for epoch in range(epochs):for idx, (input_ids, token_type_ids, length, labels) in enumerate(train_loader):logits = model(input_ids, token_type_ids)loss = paddle.mean(loss_fn(logits, labels))loss.backward()optimizer.step()optimizer.clear_grad()step += 1print("epoch:%d - step:%d - loss: %f" % (epoch, step, loss))# 每個(gè) epoch 評(píng)估一次evaluate(model, metric, dev_loader)# 保存模型參數(shù)paddle.save(model.state_dict(),'./ernie_result/model_%d.pdparams' % step)# 訓(xùn)練完成，加載模型參數(shù) state_dict = paddle.load("./ernie_result/model_450.pdparams") model.load_dict(state_dict)# 預(yù)測(cè) preds = predict(model, test_loader, test_ds, label_vocab) file_path = "ernie_results.txt" with open(file_path, "w", encoding="utf8") as fout:fout.write("\n".join(preds)) # 打印預(yù)測(cè)結(jié)果 print("The results have been saved in the file: %s, some examples are shown below: "% file_path) print("\n".join(preds[:10]))

訓(xùn)練過程：

epoch:0 - step:1 - loss: 2.788503 epoch:0 - step:2 - loss: 2.520449 epoch:0 - step:3 - loss: 2.365216 epoch:0 - step:4 - loss: 2.255839 epoch:0 - step:5 - loss: 2.108390 epoch:0 - step:6 - loss: 2.006438 ... epoch:0 - step:100 - loss: 0.045199 eval precision: 0.969141 - recall: 0.977292 - f1: 0.973199 epoch:1 - step:101 - loss: 0.026065 ... epoch:1 - step:200 - loss: 0.012335 eval precision: 0.984925 - recall: 0.989066 - f1: 0.986991 epoch:2 - step:201 - loss: 0.014337 ... epoch:2 - step:300 - loss: 0.004556 eval precision: 0.987427 - recall: 0.990749 - f1: 0.989085 epoch:3 - step:301 - loss: 0.003423 ... epoch:3 - step:400 - loss: 0.002968 eval precision: 0.987427 - recall: 0.990749 - f1: 0.989085 epoch:4 - step:401 - loss: 0.001868 ... epoch:4 - step:500 - loss: 0.016371 eval precision: 0.989933 - recall: 0.992431 - f1: 0.991180 epoch:5 - step:501 - loss: 0.006276 ... epoch:5 - step:530 - loss: 0.001634 ...

一些預(yù)測(cè)結(jié)果：

The results have been saved in the file: ernie_results.txt, some examples are shown below: ('黑龍江省', 'A1')('雙鴨山市', 'A2')('尖山區(qū)', 'A3')('八馬路與東平行路交叉口北40米', 'A4')('韋業(yè)濤', 'P')('18600009172', 'T') ('廣西壯族自治區(qū)', 'A1')('桂林市', 'A2')('雁山區(qū)', 'A3')('雁山鎮(zhèn)西龍村老年活動(dòng)中心', 'A4')('17610348888', 'T')('羊卓衛(wèi)', 'P') ('15652864561', 'T')('河南省', 'A1')('開封市', 'A2')('順河回族區(qū)', 'A3')('順河區(qū)公園路32號(hào)', 'A4')('趙本山', 'P') ('河北省', 'A1')('唐山市', 'A2')('玉田縣', 'A3')('無終大街159號(hào)', 'A4')('18614253058', 'T')('尚漢生', 'P') ('臺(tái)灣', 'A1')('臺(tái)中市', 'A2')('北區(qū)', 'A3')('北區(qū)錦新街18號(hào)', 'A4')('18511226708', 'T')('薊麗', 'P') ('廖梓琪', 'P')('18514743222', 'T')('湖北省', 'A1')('宜昌市', 'A2')('長(zhǎng)陽土家族自治縣', 'A3')('賀家坪鎮(zhèn)賀家坪村一組臨河1號(hào)', 'A4') ('江蘇省', 'A1')('南通市', 'A2')('海門市', 'A3')('孝威村孝威路88號(hào)', 'A4')('18611840623', 'T')('計(jì)星儀', 'P') ('17601674746', 'T')('趙春麗', 'P')('內(nèi)蒙古自治區(qū)', 'A1')('烏蘭察布市', 'A2')('涼城縣', 'A3')('新建街', 'A4') ('云南省', 'A1')('臨滄市', 'A2')('耿馬傣族佤族自治縣', 'A3')('鑫源路法院對(duì)面', 'A4')('許貞愛', 'P')('18510566685', 'T') ('四川省', 'A1')('成都市', 'A2')('雙流區(qū)', 'A3')('東升鎮(zhèn)北倉路196號(hào)', 'A4')('耿丕嶺', 'P')('18513466161', 'T') 創(chuàng)作挑戰(zhàn)賽新人創(chuàng)作獎(jiǎng)勵(lì)來咯，堅(jiān)持創(chuàng)作打卡瓜分現(xiàn)金大獎(jiǎng)

總結(jié)

以上是生活随笔為你收集整理的Paddle 使用预训练模型 实现快递单信息抽取的全部?jī)?nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔推薦給好友。