# GPU版本
conda create -n py37testmaas python=3.7
pip install cryptography==3.4.8 tensorflow-gpu==1.15.5 torch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0
pip install "modelscope[nlp]" -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html
# CPU版本
conda create -n py37testmaas python=3.7
pip install cryptography==3.4.8 tensorflow==1.15.5 torch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0
pip install "modelscope[nlp]" -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html
点击右上角快速体验按钮,选在CPU或GPU机器运行实例,创建notebook执行推理或训练代码即可
第一次启动时可能会花费一些时间创建索引
POI(Point Of Interest,兴趣点)搜索是地图类应用的核心功能,需要根据用户的query找到对应地图上的POI。一个POI的组成通常是其对应的地址描述以及经纬度描述。本任务需要在给定用户query以及候选POI列表的情况下,根据POI与query的相关性程度对POI进行排序。
任务的每一条输入包括用户query、用户位置以及候选POI列表,每个POI包括POI的地址描述以及POI位置。需要根据query与POI的相关性按照相关度从高到底为POI打分。
输入:用户query、用户位置、候选POI列表
输出:POI得分
注意本模型基于OpenStreetMap杭州POI进行训练,不保证在其他POI库以及其他地方query上的效果。
地址由于其丰富的表达以及与地图联动的多模态属性,一直是自动化处理的一个难题。达摩院联合高德发布了多任务多模态地址预训练底座MGeo模型。该模型基于地图-文本多模态架构,使用多任务预训练(MOMETAS)技术融合了注意力对抗预训练(ASA)、句子对预训练(MaSTS)、多模态预训练,训练得到适合于多类地址任务的预训练底座,为下游广泛的地址处理任务带来性能提升。
在ModelScope中我们开源的版本是基于开源地址数据以及开源地图OpenStreetMap训练的MGeo预训练底座以及其在GeoGLUE地理语义评测榜中多个任务的下游模型。
更多信息详见MGeo底座模型:https://modelscope.cn/models/damo/mgeo_backbone_chinese_base/summary
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
task = Tasks.text_ranking
model = 'damo/mgeo_geographic_textual_similarity_rerank_chinese_base'
# 多模态输入,包括需要排序的文本以及地理信息
multi_modal_inputs = {
"source_sentence": ['杭州余杭东方未来学校附近世纪华联商场(金家渡北苑店)'],
"first_sequence_gis": [[[13159, 13295, 13136, 13157, 13158, 13291, 13294, 74505, 74713, 75387, 75389, 75411], [3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4], [3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4], [[1254, 1474, 1255, 1476], [1253, 1473, 1256, 1476], [1247, 1473, 1255, 1480], [1252, 1475, 1253, 1476], [1253, 1475, 1253, 1476], [1252, 1471, 1254, 1475], [1254, 1473, 1256, 1475], [1238, 1427, 1339, 1490], [1238, 1427, 1339, 1490], [1252, 1474, 1255, 1476], [1252, 1474, 1255, 1476], [1249, 1472, 1255, 1479]], [[24, 23, 15, 23], [24, 28, 15, 18], [31, 24, 22, 22], [43, 13, 37, 13], [43, 6, 35, 6], [31, 32, 22, 14], [19, 30, 9, 16], [24, 30, 15, 16], [24, 30, 15, 16], [29, 24, 20, 22], [28, 25, 19, 21], [31, 26, 22, 20]], "120.08802231437534,30.343853313981505"]],
"sentences_to_compare": [
'良渚街道金家渡北苑42号世纪华联超市(金家渡北苑店)',
'金家渡路金家渡中苑南区70幢金家渡中苑70幢',
'金家渡路140-142号附近家家福足道(金家渡店)'
],
"second_sequence_gis": [
[[13083, 13081, 13084, 13085, 13131, 13134, 13136, 13147, 13148], [3, 3, 3, 3, 3, 3, 3, 3, 3], [3, 4, 4, 4, 4, 4, 4, 4, 4], [[1248, 1477, 1250, 1479], [1248, 1475, 1250, 1476], [1247, 1478, 1249, 1481], [1249, 1479, 1249, 1480], [1249, 1476, 1250, 1476], [1250, 1474, 1252, 1478], [1247, 1473, 1255, 1480], [1250, 1478, 1251, 1479], [1249, 1478, 1250, 1481]], [[30, 26, 21, 20], [32, 43, 23, 43], [33, 23, 23, 23], [31, 13, 22, 13], [25, 43, 16, 43], [20, 33, 10, 33], [26, 29, 17, 17], [18, 21, 8, 21], [26, 23, 17, 23]], "120.08075205680345,30.34697777462197"],
[[13291, 13159, 13295, 74713, 75387, 75389, 75411], [3, 3, 3, 4, 4, 4, 4], [3, 4, 4, 4, 4, 4, 4], [[1252, 1471, 1254, 1475], [1254, 1474, 1255, 1476], [1253, 1473, 1256, 1476], [1238, 1427, 1339, 1490], [1252, 1474, 1255, 1476], [1252, 1474, 1255, 1476], [1249, 1472, 1255, 1479]], [[28, 28, 19, 18], [22, 16, 12, 16], [23, 24, 13, 22], [24, 30, 15, 16], [27, 20, 18, 20], [27, 21, 18, 21], [30, 24, 21, 22]], "120.0872539617001,30.342783672056953"],
[[13291, 13290, 13294, 13295, 13298], [3, 3, 3, 3, 3], [3, 4, 4, 4, 4], [[1252, 1471, 1254, 1475], [1253, 1469, 1255, 1472], [1254, 1473, 1256, 1475], [1253, 1473, 1256, 1476], [1255, 1467, 1258, 1472]], [[32, 25, 23, 21], [26, 33, 17, 33], [21, 19, 11, 19], [25, 21, 16, 21], [21, 33, 11, 33]], "120.08839673752281,30.34156156893651"]
]
}
# 单模态输入,只包括需要排序的文本
single_modal_inputs = {
"source_sentence": ['杭州余杭东方未来学校附近世纪华联商场(金家渡北苑店)'],
"sentences_to_compare": [
'良渚街道金家渡北苑42号世纪华联超市(金家渡北苑店)',
'金家渡路金家渡中苑南区70幢金家渡中苑70幢',
'金家渡路140-142号附近家家福足道(金家渡店)'
]
}
# 模型可接受多模态输入
pipeline_ins = pipeline(
task=task, model=model)
print(pipeline_ins(input=multi_modal_inputs))
# 输出
# {'scores': [0.9997552633285522, 0.027718106284737587, 0.03500296175479889]}
# 模型可接受单模态输入
pipeline_ins = pipeline(
task=task, model=model)
print(pipeline_ins(input=single_modal_inputs))
# 输出
# {'scores': [0.9986912608146667, 0.0075200702995061874, 0.014017169363796711]}
当用户有自己标注好的数据希望基于MGeo底座进行训练或基于训练好的下游模型进行继续训练时,可使用自定义训练功能。
以GeoGLUE的Query-POI排序任务为例
import os
from modelscope.msdatasets import MsDataset
from modelscope.metainfo import Trainers, Preprocessors
from modelscope.utils.constant import ModelFile, Tasks
from modelscope.trainers import build_trainer
tmp_dir = 'tmp_dir'
def finetune(model_id,
train_dataset,
eval_dataset,
name=Trainers.nlp_text_ranking_trainer,
cfg_modify_fn=None,
**kwargs):
kwargs = dict(
model=model_id,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
work_dir=tmp_dir,
cfg_modify_fn=cfg_modify_fn,
**kwargs)
os.environ['LOCAL_RANK'] = '0'
trainer = build_trainer(name=name, default_args=kwargs)
trainer.train()
results_files = os.listdir(tmp_dir)
def cfg_modify_fn(cfg):
neg_sample = 19
cfg.task = 'text-ranking'
cfg['preprocessor'] = {'type': 'mgeo-ranking'}
cfg.train.optimizer.lr = 5e-5
cfg['dataset'] = {
'train': {
'type': 'mgeo',
'query_sequence': 'query',
'pos_sequence': 'positive_passages',
'neg_sequence': 'negative_passages',
'text_fileds': ['text', 'gis'],
'qid_field': 'query_id',
'neg_sample': neg_sample,
'sequence_length': 64
},
'val': {
'type': 'mgeo',
'query_sequence': 'query',
'pos_sequence': 'positive_passages',
'neg_sequence': 'negative_passages',
'text_fileds': ['text', 'gis'],
'qid_field': 'query_id'
},
}
cfg.evaluation.dataloader.batch_size_per_gpu = 16
cfg.train.dataloader.batch_size_per_gpu = 3
cfg.train.dataloader.workers_per_gpu = 16
cfg.evaluation.dataloader.workers_per_gpu = 16
# 数据量较大,限制训练步数为10,如需全量训练可删去下面两句
cfg.train.train_iters_per_epoch = 10
cfg.evaluation.val_iters_per_epoch = 10
cfg['evaluation']['metrics'] = "mrr@1"
cfg.train.max_epochs = 1
cfg.model['neg_sample'] = neg_sample
cfg.model['gis_num'] = 2
cfg.model['finetune_mode'] = 'multi-modal'
cfg.train.hooks = [{
'type': 'CheckpointHook',
'interval': 1
}, {
'type': 'TextLoggerHook',
'interval': 100
}, {
'type': 'IterTimerHook'
}, {
'type': 'EvaluationHook',
'by_epoch': True
}]
# lr_scheduler的配置
cfg.train.lr_scheduler = {
"type": "LinearLR",
"start_factor": 1.0,
"end_factor": 0.5,
"total_iters": int(len(train_ds) / cfg.train.dataloader.batch_size_per_gpu) * cfg.train.max_epochs,
"options": {
"warmup": {
"type": "LinearWarmup",
"warmup_iters": int(len(train_ds) / cfg.train.dataloader.batch_size_per_gpu)
},
"by_epoch": False
}
}
return cfg
# load dataset
train_dataset = MsDataset.load('GeoGLUE', subset_name='GeoTES-rerank', split='train', namespace='damo')
dev_dataset = MsDataset.load('GeoGLUE', subset_name='GeoTES-rerank', split='validation', namespace='damo')
train_ds = train_dataset['train']
dev_ds = dev_dataset['validation']
model_id = 'damo/mgeo_backbone_chinese_base'
finetune(
model_id=model_id,
train_dataset=train_ds,
eval_dataset=dev_ds,
cfg_modify_fn=cfg_modify_fn,
name=Trainers.mgeo_ranking_trainer)
output_dir = os.path.join(tmp_dir, ModelFile.TRAIN_OUTPUT_DIR)
print(f'model is saved to {output_dir}')
如果需要从本地load用户自定义数据,可以先将数据处理为如下格式,并保存为train.json和dev.json:
{"query_id": 0, "query": "丽华公寓(通惠中路)向北检验检疫科学研究所", "query_gis": "[[], [], [], [], [], \"120.59443087451544,30.315515932852602\"]", "idx": "0", "positive_passages": [{"text": "杭州中新街(惠港二路)76饶平县检验检疫局", "gis": "[[], [], [], [], [], \"120.20509044775532,30.076259797983873\"]"}], "negative_passages": [{"text": "杭州中新街", "gis": "[[], [], [], [], [], \"120.20509044775532,30.076259797983873\"]"}]}
例子中的gis信息为空缺状态,用户可补充上自己的gis信息,模型在有gis和无gis的场景下均可以进行推断,gis信息的加入可以提升模型效果。更新原始代码中的neg_sample数量为自定义训练集的负例个数。
然后替换原流程中的train_dataset和dev_dataset为:
local_train = 'train.json'
local_test = 'dev.json'
train_dataset = MsDataset.load('json', data_files={'train': [local_train]})
dev_dataset = MsDataset.load('json', data_files={'validation': [local_test]})
train_ds = train_dataset.to_hf_dataset()
dev_ds = dev_dataset.to_hf_dataset()
@article{DBLP:journals/corr/abs-2210-10293,
author = {Hongqiu Wu and
Ruixue Ding and
Hai Zhao and
Boli Chen and
Pengjun Xie and
Fei Huang and
Min Zhang},
title = {Forging Multiple Training Objectives for Pre-trained Language Models
via Meta-Learning},
journal = {CoRR},
volume = {abs/2210.10293},
year = {2022},
url = {https://doi.org/10.48550/arXiv.2210.10293},
doi = {10.48550/arXiv.2210.10293},
eprinttype = {arXiv},
eprint = {2210.10293},
timestamp = {Mon, 24 Oct 2022 18:10:06 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2210-10293.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@misc{2301.04283,
Author = {Ruixue Ding and Boli Chen and Pengjun Xie and Fei Huang and Xin Li and Qiang Zhang and Yao Xu},
Title = {A Multi-Modal Geographic Pre-Training Method},
Year = {2023},
Eprint = {arXiv:2301.04283},
}
@misc{2206.12608,
Author = {Hongqiu Wu and Ruixue Ding and Hai Zhao and Pengjun Xie and Fei Huang and Min Zhang},
Title = {Adversarial Self-Attention for Language Understanding},
Year = {2022},
Eprint = {arXiv:2206.12608},
}
如果您在使用中遇到任何困难,请钉钉搜索加入答疑群,群号:26735013283