【PaddleNLP No.21】Create paddle.inference infer example for ernie-3.0-tiny

[hanlintang]

PR types

Function optimization

PR changes

Models

Description

1. 为Ernie-3.0-tiny删除无法匹配PIR的fastdeploy部署方式，提供基于paddle.inference的推理示例。
2. 基于实现内容修改文档，更新更换推理方式的运行结果。

Modified:
slm/model_zoo/ernie-3.0-tiny/README.md
slm/model_zoo/ernie-3.0-tiny/deploy/README.md
slm/model_zoo/ernie-3.0-tiny/deploy/python/README.md
slm/model_zoo/ernie-3.0-tiny/deploy/python/infer_demo.py

其他问题

在训练Ernie-3.0-tiny时遇到了一些问题，可能需要后续进行修复：

1. 部分参数丢失

Traceback (most recent call last):
  File "run_train.py", line 286, in <module>
    main()
  File "run_train.py", line 151, in main
    model = JointErnie.from_pretrained(
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddlenlp/transformers/model_utils.py", line 2567, in from_pretrained
    model, missing_keys, unexpected_keys, mismatched_keys = cls._load_pretrained_model(
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddlenlp/transformers/model_utils.py", line 2254, in _load_pretrained_model
    raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")
RuntimeError: Error(s) in loading state_dict for JointErnie:
        Skip loading for ernie.embeddings.word_embeddings.weight. ernie.embeddings.word_embeddings.weight receives a shape [5965, 312], but the expected shape is [6000, 312].
        You may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method.

按照最后一句提示在指定位置加入参数设置忽略不匹配的size可以通过。

2. 评估部分报错

Traceback (most recent call last):
  File "run_train.py", line 288, in <module>
    main()
  File "run_train.py", line 230, in main
    trainer.train(resume_from_checkpoint=checkpoint)
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddlenlp/trainer/trainer.py", line 892, in train
    return self._inner_training_loop(
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddlenlp/trainer/trainer.py", line 1281, in _inner_training_loop
    self._maybe_log_save_evaluate(tr_loss, model, epoch, ignore_keys_for_eval, inputs=inputs)
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddlenlp/trainer/trainer.py", line 1578, in _maybe_log_save_evaluate
    metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddlenlp/trainer/trainer.py", line 3118, in evaluate
    output = self.evaluation_loop(
  File "/home/aistudio/.local/lib/python3.8/site-packages/paddlenlp/trainer/trainer.py", line 3340, in evaluation_loop
    metrics = self.compute_metrics(EvalPrediction(predictions=all_preds, label_ids=batch_labels))
  File "/home/aistudio/PaddleNLP/slm/model_zoo/ernie-3.0-tiny/utils.py", line 113, in compute_metrics
    intent_label, slot_label = p.label_ids
ValueError: too many values to unpack (expected 2)

问题出现在评估部分，需要后续修改；暂时跳过的方式是在训练参数中去掉以下部分可以只训练跳过dev评估：

--do_eval
--eval_steps 100
--load_best_model_at_end True
--metric_for_best_model eval_accuracy

Issue: #9763
@DrownFish19

[paddle-bot]

Thanks for your contribution!

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 48.67%. Comparing base (ce7b4cc) to head (474f42d).
Report is 17 commits behind head on develop.

❌ Your project status has failed because the head coverage (48.67%) is below the target coverage (58.00%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop   #10480      +/-   ##
===========================================
- Coverage    48.99%   48.67%   -0.32%     
===========================================
  Files          765      768       +3     
  Lines       125974   126915     +941     
===========================================
+ Hits         61720    61778      +58     
- Misses       64254    65137     +883     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[DrownFish19]

可以尝试输出intent_label, slot_label = p.label_ids 中的p.label_id，检查输出内容

[DrownFish19]

LGTM