PaddlePaddle
**非常感谢提供了predictor代码，可在运行时有问题请教下：**
非常感谢提供了predictor代码，可在运行时有问题请教下：
1、 采用 cblue 训练模型
python train_spo.py --batch_size 12 --max_seq_length 300 --learning_rate 6e-5 --epochs 3
对模型进行预训练，在Macbook上，用CPU跑，
但 spo_loss 从很大值降到 100多，spo fi: 始终为0，这是正常的吗？
global step 2300, epoch: 1, batch: 2300, loss: 223.04558, ent_loss: 114.37089, spo_loss: 108.67469, speed: 0.71 steps/s
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 897/897 [46:15<00:00, 3.09s/it]
eval loss: 249.20973, entity f1: 0.00000, spo f1: 0.00000
[2022-07-27 13:26:15,749] [ INFO] - tokenizer config file saved in ./checkpoint/model_2300/tokenizer_config.json
[2022-07-27 13:26:15,750] [ INFO] - Special tokens file saved in ./checkpoint/model_2300/special_tokens_map.json
2、我用中间模型输出 静态图做预测，跑的结果出错， 如下：
` python infer_spo.py --device cpu --dataset CMeIE --model_path_prefix ../../cblue/export_CMeIE/inference
[2022-07-27 14:27:25,541] [ INFO] - model_path_prefix : ../../cblue/export_CMeIE/inference
[2022-07-27 14:27:25,541] [ INFO] - model_name_or_path : ernie-health-chinese
[2022-07-27 14:27:25,541] [ INFO] - dataset : CMeIE
[2022-07-27 14:27:25,541] [ INFO] - data_file : None
[2022-07-27 14:27:25,541] [ INFO] - max_seq_length : 300
[2022-07-27 14:27:25,541] [ INFO] - use_fp16 : False
[2022-07-27 14:27:25,542] [ INFO] - num_threads : 4
[2022-07-27 14:27:25,542] [ INFO] - batch_size : 20
[2022-07-27 14:27:25,542] [ INFO] - device : cpu
[2022-07-27 14:27:25,542] [ INFO] - device_id : 0
[2022-07-27 14:27:25,542] [ WARNING] - Can't find the faster_tokenizer package, please ensure install faster_tokenizer correctly. You can install faster_tokenizer by `pip install faster_tokenizer`(Currently only work for linux platform).
[2022-07-27 14:27:25,542] [ INFO] - We are using <class 'paddlenlp.transformers.electra.tokenizer.ElectraTokenizer'> to load 'ernie-health-chinese'.
[2022-07-27 14:27:25,542] [ INFO] - Already cached /Users/lizzysong/.paddlenlp/models/ernie-health-chinese/vocab.txt
[2022-07-27 14:27:25,558] [ INFO] - tokenizer config file saved in /Users/lizzysong/.paddlenlp/models/ernie-health-chinese/tokenizer_config.json
[2022-07-27 14:27:25,558] [ INFO] - Special tokens file saved in /Users/lizzysong/.paddlenlp/models/ernie-health-chinese/special_tokens_map.json
[2022-07-27 14:27:25,558] [ INFO] - >>> [InferBackend] Creating Engine ...
[Paddle2ONNX] Start to parse PaddlePaddle model...
[Paddle2ONNX] Model file path: ../../cblue/export_CMeIE/inference.pdmodel
[Paddle2ONNX] Paramters file path: ../../cblue/export_CMeIE/inference.pdiparams
[Paddle2ONNX] Start to parsing Paddle model...
[Paddle2ONNX] Use opset_version = 13 for ONNX export.
[Paddle2ONNX] PaddlePaddle model is exported as ONNX format now.
[2022-07-27 14:27:31,926] [ INFO] - >>> [InferBackend] Use CPU to inference ...
[2022-07-27 14:27:33,617] [ INFO] - >>> [InferBackend] Engine Created ...
Traceback (most recent call last):
File "/Users/lizzysong/PaddleNLP/model_zoo/ernie-health/deploy/predictor/infer_spo.py", line 69, in <module>
predictor.predict(input_data)
File "/Users/lizzysong/PaddleNLP/model_zoo/ernie-health/deploy/predictor/predictor.py", line 320, in predict
infer_result = self.infer_batch(encoded_inputs)
File "/Users/lizzysong/PaddleNLP/model_zoo/ernie-health/deploy/predictor/predictor.py", line 151, in infer_batch
results = self._infer(input_dict)
File "/Users/lizzysong/PaddleNLP/model_zoo/ernie-health/deploy/predictor/predictor.py", line 140, in _infer
infer_data = self.inference_backend.infer(input_dict)
File "/Users/lizzysong/PaddleNLP/model_zoo/ernie-health/deploy/predictor/predictor.py", line 116, in infer
result = self.predictor.run(None, input_dict)
File "/Users/lizzysong/opt/anaconda3/lib/python3.9/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 188, in run
raise ValueError("Model requires {} inputs. Input Feed contains {}".format(num_required_inputs, num_inputs))
ValueError: Model requires 4 inputs. Input Feed contains 3
Originally posted by @charlieliu9999 in https://github.com/PaddlePaddle/PaddleNLP/issues/2827#issuecomment-1196326683
-
CMeIE数据集上收敛速度是比较慢，我这边的经验是4卡训练5000~6000steps之后F1值才会从0逐渐增加，训练大概100 epochs才能达到README表格中给出的F1值。 - 预测时用的
export_model.py是拉的最新代码么，这里的静态图导出的实现也有更新，可以用最新代码导出一版再试一试。
更新了 export_model.py 可以跑出来了, 模型训练不够，结果不行
[2022-07-28 16:15:31,769] [ INFO] - >>> [InferBackend] Use CPU to inference ... [2022-07-28 16:15:33,716] [ INFO] - >>> [InferBackend] Engine Created ... [2022-07-28 16:15:34,248] [ INFO] - input data: 骶髂关节炎是明确诊断JAS的关键条件。若有肋椎关节病变会使胸部扩张度减小。 [2022-07-28 16:15:34,248] [ INFO] - detected entities and relations: [2022-07-28 16:15:34,248] [ INFO] - ----------------------------- [2022-07-28 16:15:34,249] [ INFO] - input data: 稳定型缺血性心脏疾病@肥胖与缺乏活动也导致高血压增多。 [2022-07-28 16:15:34,249] [ INFO] - detected entities and relations: [2022-07-28 16:15:34,249] [ INFO] - * entity: 稳定型缺血性心脏疾病, position: (0, 9) [2022-07-28 16:15:34,249] [ INFO] - ----------------------------- [2022-07-28 16:15:34,249] [ INFO] - input data: 呼吸内科 ,071岁,M,因反复咳嗽30年，气促3年，再发伴发热10余天。入院。胸廓桶状胸，肋间隙增宽，语颤减弱，叩诊过清音 [2022-07-28 16:15:34,249] [ INFO] - detected entities and relations: [2022-07-28 16:15:34,249] [ INFO] - * entity: 呼吸内科 ,071岁,M,因反复咳嗽30年，气促3年，再发伴发热10余天。入院。胸廓桶状胸, position: (0, 44) [2022-07-28 16:15:34,250] [ INFO] - ----------------------------- [2022-07-28 16:15:34,250] [ INFO] - input data: 反复咳嗽、咳痰、活动后气促10年，再发加重3天 [2022-07-28 16:15:34,250] [ INFO] - detected entities and relations: [2022-07-28 16:15:34,250] [ INFO] - ----------------------------- [2022-07-28 16:15:34,250] [ INFO] - input data: 右侧腰痛2年余 [2022-07-28 16:15:34,250] [ INFO] - detected entities and relations: [2022-07-28 16:15:34,250] [ INFO] - -----------------------------
再请教个问题，如果要提取医学SPO实体关系，建议用 ernie-health 训练的模型，还是在 uie 模型上微调 ？ 如果 uie , 采用 uie-tiny 还是 uie-medic 哪个效果更好？
具体效果取决于实际数据。
- ernie-health对训练数据量要求比较高，比如CMeIE有1.4万条左右，如果实际数据和CMeIE中抽取的schema类似，可以尝试混合训练或者微调好的模型来看看效果。
- uie在SPO设计上相对灵活，可以直接用taskflow调用
uie-base和uie-medical分别试下效果。如果效果不能达到预期，建议在uie-base上进行微调，此时也需要一定量的数据标注。
uie 模型，定义是用小样本微调，用大数据量是否能有效果？多少数据量是合适的？ 我用150多标注数据（一个科室数据），训练 uie-tiny ，有一定效果，实体识别较好，关系抽取差些。 在 uie 模型中，标注数据是否要涵盖所有实体和关系？对于未学习过的实体能否推理出来
This issue is stale because it has been open for 60 days with no activity. 当前issue 60天内无活动，被标记为stale。
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}
This issue was closed because it has been inactive for 14 days since being marked as stale. 当前issue 被标记为stale已有14天，即将关闭。