diff --git a/PaddleNLP/examples/lexical_analysis/README.md b/PaddleNLP/examples/lexical_analysis/README.md index 6a7c36cb97df72813ab3bb6ec2014c423654503f..cb651e7ee2794306a7edc66b33e473d9479f783e 100644 --- a/PaddleNLP/examples/lexical_analysis/README.md +++ b/PaddleNLP/examples/lexical_analysis/README.md @@ -20,7 +20,7 @@ - paddlepaddle >= 2.0.0rc1，安装方式请参考 [快速安装](https://www.paddlepaddle.org.cn/install/quick)。 -- paddlenlp >= 2.0.0b, 安装方式：`pip install paddlenlp>=2.0.0b` +- paddlenlp >= 2.0.0b2, 安装方式：`pip install paddlenlp\>=2.0.0b2` ### 2.2 数据准备 diff --git a/PaddleNLP/examples/lexical_analysis/eval.py b/PaddleNLP/examples/lexical_analysis/eval.py index 50742de7c2a0cce15032e6c11c289c463bc90e09..d8b78edae34fa88782004d12c8d69a87930c5ea9 100644 --- a/PaddleNLP/examples/lexical_analysis/eval.py +++ b/PaddleNLP/examples/lexical_analysis/eval.py @@ -69,8 +69,7 @@ def evaluate(args): test_dataset.num_labels) model = paddle.Model(network) chunk_evaluator = ChunkEvaluator( - int(math.ceil((test_dataset.num_labels + 1) / 2.0)), - "IOB") # + 1 for SOS and EOS + label_list=test_dataset.label_vocab.keys(), suffix=True) model.prepare(None, None, chunk_evaluator) # Load the model and start predicting diff --git a/PaddleNLP/examples/named_entity_recognition/express_ner/README.md b/PaddleNLP/examples/named_entity_recognition/express_ner/README.md index f0b61033e3f4028296ffad067fca5257e5a657fd..5e65b915ad807102d9d73de5e668d1b9e0d2bbb9 100644 --- a/PaddleNLP/examples/named_entity_recognition/express_ner/README.md +++ b/PaddleNLP/examples/named_entity_recognition/express_ner/README.md @@ -12,7 +12,7 @@ - paddlepaddle >= 2.0.0rc1，安装方式请参考 [快速安装](https://www.paddlepaddle.org.cn/install/quick)。 -- paddlenlp >= 2.0.0b, 安装方式：`pip install paddlenlp>=2.0.0b` +- paddlenlp >= 2.0.0b2, 安装方式：`pip install paddlenlp\>=2.0.0b2` ### 2.2 数据准备 @@ -20,10 +20,10 @@ 数据集已经保存在data目录中，示例如下 ``` -16620200077宣荣嗣甘肃省白银市会宁县河畔镇十字街金海超市西行50米 T-BT-IT-IT-IT-IT-IT-IT-IT-IT-IT-IP-BP-IP-IA1-BA1-IA1-IA2-BA2-IA2-IA3-BA3-IA3-IA4-BA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-IA4-I -13552664307姜骏炜云南省德宏傣族景颇族自治州盈江县平原镇蜜回路下段 T-BT-IT-IT-IT-IT-IT-IT-IT-IT-IT-IP-BP-IP-IA1-BA1-IA1-IA2-BA2-IA2-IA2-IA2-IA2-IA2-IA2-IA2-IA2-IA3-BA3-IA3-IA4-BA4-IA4-IA4-IA4-IA4-IA4-IA4-I +1^B6^B6^B2^B0^B2^B0^B0^B0^B7^B7^B宣^B荣^B嗣^B甘^B肃^B省^B白^B银^B市^B会^B宁^B县^B河^B畔^B镇^B十^B字^B街^B金^B海^B超^B市^B西^B行^B5^B0^B米 T-B^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BP-B^BP-I^BP-I^BA1-B^BA1-I^BA1-I^BA2-B^BA2-I^BA2-I^BA3-B^BA3-I^BA3-I^BA4-B^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I +1^B3^B5^B5^B2^B6^B6^B4^B3^B0^B7^B姜^B骏^B炜^B云^B南^B省^B德^B宏^B傣^B族^B景^B颇^B族^B自^B治^B州^B盈^B江^B县^B平^B原^B镇^B蜜^B回^B路^B下^B段 T-B^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BT-I^BP-B^BP-I^BP-I^BA1-B^BA1-I^BA1-I^BA2-B^BA2-I^BA2-I^BA2-I^BA2-I^BA2-I^BA2-I^BA2-I^BA2-I^BA2-I^BA3-B^BA3-I^BA3-I^BA4-B^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I^BA4-I ``` -数据集中以特殊字符"\t"分隔文本、标签，以特殊字符"\002"分隔每个字。标签的定义如下： +数据集中以特殊字符"\t"分隔文本、标签，以特殊字符"\002"(示例中显示为"^B")分隔每个字。标签的定义如下： | 标签 | 定义 | 标签 | 定义 | | -------- | -------- |-------- | -------- | diff --git a/PaddleNLP/examples/named_entity_recognition/msra_ner/README.md b/PaddleNLP/examples/named_entity_recognition/msra_ner/README.md index 3f570683a07a14896de3f99c8397ead1196e52fa..0fc7d6926addac421f91f08fe97d486344448e78 100644 --- a/PaddleNLP/examples/named_entity_recognition/msra_ner/README.md +++ b/PaddleNLP/examples/named_entity_recognition/msra_ner/README.md @@ -5,11 +5,11 @@ MSRA-NER 数据集由微软亚研院发布，其目标是识别文本中具有特定意义的实体，主要包括人名、地名、机构名等。示例如下： ``` -海钓比赛地点在厦门与金门之间的海域。 OOOOOOOB-LOCI-LOCOB-LOCI-LOCOOOOOO -这座依山傍水的博物馆由国内一流的设计师主持设计，整个建筑群精美而恢宏。 OOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO +不\002久\002前\002，\002中\002国\002共\002产\002党\002召\002开\002了\002举\002世\002瞩\002目\002的\002第\002十\002五\002次\002全\002国\002代\002表\002大\002会\002。 O\002O\002O\002O\002B-ORG\002I-ORG\002I-ORG\002I-ORG\002I-ORG\002O\002O\002O\002O\002O\002O\002O\002O\002B-ORG\002I-ORG\002I-ORG\002I-ORG\002I-ORG\002I-ORG\002I-ORG\002I-ORG\002I-ORG\002I-ORG\002O +这\002次\002代\002表\002大\002会\002是\002在\002中\002国\002改\002革\002开\002放\002和\002社\002会\002主\002义\002现\002代\002化\002建\002设\002发\002展\002的\002关\002键\002时\002刻\002召\002开\002的\002历\002史\002性\002会\002议\002。 O\002O\002O\002O\002O\002O\002O\002O\002B-LOC\002I-LOC\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O\002O ``` -数据集中以特殊字符"\t"分隔文本、标签，以特殊字符"\002"分隔每个字。 +PaddleNLP集成的数据集MSRA-NER数据集对文件格式做了调整：每一行文本、标签以特殊字符"\t"进行分隔，每个字之间以特殊字符"\002"分隔。 ## 2. 快速开始 @@ -19,7 +19,7 @@ MSRA-NER 数据集由微软亚研院发布，其目标是识别文本中具有 - paddlepaddle >= 2.0.0rc1，安装方式请参考 [快速安装](https://www.paddlepaddle.org.cn/install/quick)。 -- paddlenlp >= 2.0.0b, 安装方式：`pip install paddlenlp>=2.0.0b` +- paddlenlp >= 2.0.0b2, 安装方式：`pip install paddlenlp\>=2.0.0b2` ### 2.2 启动MSRA-NER任务 @@ -52,22 +52,21 @@ python -u ./run_msra_ner.py \ 训练过程将按照 `logging_steps` 和 `save_steps` 的设置打印如下日志： ``` -global step 1496, epoch: 2, batch: 192, loss: 0.010747, speed: 4.77 step/s -global step 1497, epoch: 2, batch: 193, loss: 0.004837, speed: 4.46 step/s -global step 1498, epoch: 2, batch: 194, loss: 0.011281, speed: 4.24 step/s -global step 1499, epoch: 2, batch: 195, loss: 0.005711, speed: 4.73 step/s -global step 1500, epoch: 2, batch: 196, loss: 0.003150, speed: 4.52 step/s -eval loss: 0.010307, precision: 0.884222, recall: 0.903190, f1: 0.893605 +global step 3996, epoch: 2, batch: 1184, loss: 0.008593, speed: 4.15 step/s +global step 3997, epoch: 2, batch: 1185, loss: 0.008453, speed: 4.17 step/s +global step 3998, epoch: 2, batch: 1186, loss: 0.002294, speed: 4.19 step/s +global step 3999, epoch: 2, batch: 1187, loss: 0.005351, speed: 4.16 step/s +global step 4000, epoch: 2, batch: 1188, loss: 0.004734, speed: 4.18 step/s +eval loss: 0.006829, precision: 0.908957, recall: 0.926683, f1: 0.917734 ``` 使用以上命令进行单卡 Fine-tuning ，在验证集上有如下结果： Metric | Result | ------------------------------|-------------| -precision | 0.884222 | -recall | 0.903190 | -f1 | 0.893605 | +precision | 0.908957 | +recall | 0.926683 | +f1 | 0.917734 | ## 参考 -[Microsoft Research Asia Chinese Word-Segmentation Data Set](https://www.microsoft.com/en-us/download/details.aspx?id=52531) [The third international Chinese language processing bakeoff: Word segmentation and named entity recognition](https://faculty.washington.edu/levow/papers/sighan06.pdf) diff --git a/PaddleNLP/examples/named_entity_recognition/msra_ner/run_msra_ner.py b/PaddleNLP/examples/named_entity_recognition/msra_ner/run_msra_ner.py index 943f6aee8f914eef0b0ab21c2193689765ffa0df..a3bf4bf7ff4adaf67b064cad6e73936b8487d8f5 100644 --- a/PaddleNLP/examples/named_entity_recognition/msra_ner/run_msra_ner.py +++ b/PaddleNLP/examples/named_entity_recognition/msra_ner/run_msra_ner.py @@ -245,8 +245,8 @@ def do_train(args): if paddle.distributed.get_world_size() > 1: paddle.distributed.init_parallel_env() - train_dataset, dev_dataset = ppnlp.datasets.MSRA_NER.get_datasets( - ["train", "dev"]) + train_dataset, test_dataset = ppnlp.datasets.MSRA_NER.get_datasets( + ["train", "test"]) tokenizer = BertTokenizer.from_pretrained(args.model_name_or_path) label_list = train_dataset.get_labels() @@ -276,11 +276,11 @@ def do_train(args): num_workers=0, return_list=True) - dev_dataset = dev_dataset.apply(trans_func, lazy=True) + test_dataset = test_dataset.apply(trans_func, lazy=True) dev_batch_sampler = paddle.io.BatchSampler( - dev_dataset, batch_size=args.batch_size, shuffle=False, drop_last=True) - dev_data_loader = DataLoader( - dataset=dev_dataset, + test_dataset, batch_size=args.batch_size, shuffle=False, drop_last=True) + test_data_loader = DataLoader( + dataset=test_dataset, batch_sampler=dev_batch_sampler, collate_fn=batchify_fn, num_workers=0, @@ -336,7 +336,7 @@ def do_train(args): lr_scheduler.step() optimizer.clear_gradients() if global_step % args.save_steps == 0: - evaluate(model, loss_fct, metric, dev_data_loader, label_num) + evaluate(model, loss_fct, metric, test_data_loader, label_num) if (not args.n_gpu > 1) or paddle.distributed.get_rank() == 0: paddle.save(model.state_dict(), os.path.join(args.output_dir, diff --git a/PaddleNLP/examples/text_generation/ernie-gen/README.md b/PaddleNLP/examples/text_generation/ernie-gen/README.md index f95c487b528ba5c5a0f2d3f51894f3d3642389c3..14984e8cba1e7173b1b8103860c6ec66bc865772 100644 --- a/PaddleNLP/examples/text_generation/ernie-gen/README.md +++ b/PaddleNLP/examples/text_generation/ernie-gen/README.md @@ -14,7 +14,7 @@ ERNIE-GEN 是面向生成任务的预训练-微调框架，首次在预训练阶 - paddlepaddle >= 2.0.0rc1，安装方式请参考 [快速安装](https://www.paddlepaddle.org.cn/install/quick)。 -- paddlenlp >= 2.0.0b, 安装方式：`pip install paddlenlp>=2.0.0b` +- paddlenlp >= 2.0.0b, 安装方式：`pip install paddlenlp\>=2.0.0b` - tqdm，安装方式：`pip install tqdm` diff --git a/PaddleNLP/paddlenlp/datasets/__init__.py b/PaddleNLP/paddlenlp/datasets/__init__.py index a899b1d0fd1db3000e96cd9f39ab3cc7e8a26c99..5f63dc0516e7caa95b2f025d5224a69800718503 100644 --- a/PaddleNLP/paddlenlp/datasets/__init__.py +++ b/PaddleNLP/paddlenlp/datasets/__init__.py @@ -17,6 +17,7 @@ from .dataset import * from .glue import * from .lcqmc import * from .msra_ner import * +from .peoples_daily_ner import * from .ptb import * from .squad import * from .translation import * diff --git a/PaddleNLP/paddlenlp/datasets/msra_ner.py b/PaddleNLP/paddlenlp/datasets/msra_ner.py index e496e6e227bbec317cd186aaa75c9bdb87c7ee4b..ba522af19392fdf4d5cfe00b77b01b4350030964 100644 --- a/PaddleNLP/paddlenlp/datasets/msra_ner.py +++ b/PaddleNLP/paddlenlp/datasets/msra_ner.py @@ -27,7 +27,7 @@ __all__ = ['MSRA_NER'] class MSRA_NER(TSVDataset): - URL = "https://bj.bcebos.com/paddlehub-dataset/msra_ner.tar.gz" + URL = "https://paddlenlp.bj.bcebos.com/datasets/msra_ner.tar.gz" MD5 = None META_INFO = collections.namedtuple( 'META_INFO', ('file', 'md5', 'field_indices', 'num_discard_samples')) @@ -37,11 +37,6 @@ class MSRA_NER(TSVDataset): '67d3c93a37daba60ef43c03271f119d7', (0, 1), 1, ), - 'dev': META_INFO( - os.path.join('msra_ner', 'dev.tsv'), - 'ec772f3ba914bca5269f6e785bb3375d', - (0, 1), - 1, ), 'test': META_INFO( os.path.join('msra_ner', 'test.tsv'), '2f27ae68b5f61d6553ffa28bb577c8a7', diff --git a/PaddleNLP/paddlenlp/datasets/peoples_daily_ner.py b/PaddleNLP/paddlenlp/datasets/peoples_daily_ner.py new file mode 100644 index 0000000000000000000000000000000000000000..42c131479664d9d78056b4db6b9c24e5bcdf589d --- /dev/null +++ b/PaddleNLP/paddlenlp/datasets/peoples_daily_ner.py @@ -0,0 +1,77 @@ +# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import collections +import os +import warnings + +from paddle.io import Dataset +from paddle.dataset.common import md5file +from paddle.utils.download import get_path_from_url +from paddlenlp.utils.env import DATA_HOME + +from .dataset import TSVDataset + +__all__ = ['PeoplesDailyNER'] + + +class PeoplesDailyNER(TSVDataset): + URL = "https://paddlenlp.bj.bcebos.com/datasets/peoples_daily_ner.tar.gz" + MD5 = None + META_INFO = collections.namedtuple( + 'META_INFO', ('file', 'md5', 'field_indices', 'num_discard_samples')) + SPLITS = { + 'train': META_INFO( + os.path.join('peoples_daily_ner', 'train.tsv'), + '67d3c93a37daba60ef43c03271f119d7', + (0, 1), + 1, ), + 'dev': META_INFO( + os.path.join('peoples_daily_ner', 'dev.tsv'), + 'ec772f3ba914bca5269f6e785bb3375d', + (0, 1), + 1, ), + 'test': META_INFO( + os.path.join('peoples_daily_ner', 'test.tsv'), + '2f27ae68b5f61d6553ffa28bb577c8a7', + (0, 1), + 1, ), + } + + def __init__(self, mode='train', root=None, **kwargs): + default_root = os.path.join(DATA_HOME, 'peoples_daily_ner') + filename, data_hash, field_indices, num_discard_samples = self.SPLITS[ + mode] + fullname = os.path.join(default_root, + filename) if root is None else os.path.join( + os.path.expanduser(root), filename) + if not os.path.exists(fullname) or (data_hash and + not md5file(fullname) == data_hash): + if root is not None: # not specified, and no need to warn + warnings.warn( + 'md5 check failed for {}, download {} data to {}'.format( + filename, self.__class__.__name__, default_root)) + path = get_path_from_url(self.URL, default_root, self.MD5) + fullname = os.path.join(default_root, filename) + super(PeoplesDailyNER, self).__init__( + fullname, + field_indices=field_indices, + num_discard_samples=num_discard_samples, + **kwargs) + + def get_labels(self): + """ + Return labels of the GlueCoLA object. + """ + return ["B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC", "O"]