p-tuning_NLU
Overview
这个小项目是受乐于分享的苏剑林大佬这篇p-tuning 文章启发，也实现了个使用P-tuning进行NLU分类的任务， 思路是一样的，prompt实现方式有不同，这里是将[unused*]的embeddings参数抽取出用于初始化prompt_embed后，再接一个lstm和mlp用于关联各prompt， 与最初p-tuning提出《GPT Understands, Too》 的实现一样，结果显示在few-shot上p-tuning非常接近finetune效果。
Dataset
数据是情感分类，下载地址百度网盘 提取码：osja
Evaluation
1. finetune
python few_shot_finetune.py
测试集效果：
epoch: 0 - acc: 0.897679 - best_test_acc: 0.8976788252013264
epoch: 1 - acc: 0.876362 - best_test_acc: 0.8976788252013264
epoch: 2 - acc: 0.884889 - best_test_acc: 0.8976788252013264
epoch: 3 - acc: 0.884415 - best_test_acc: 0.8976788252013264
epoch: 4 - acc: 0.884415 - best_test_acc: 0.8976788252013264
全量参数对小样本进行finetune，仅1个epoch就收敛了
2. p-tuning
python few_shot_ptuning.py
测试集效果：
epoch: 0 - acc: 0.546660 - best_test_acc: 0.5466603505447655
epoch: 1 - acc: 0.687826 - best_test_acc: 0.6878256750355282
epoch: 2 - acc: 0.737091 - best_test_acc: 0.7370914258645191
epoch: 3 - acc: 0.722406 - best_test_acc: 0.7370914258645191
epoch: 4 - acc: 0.776883 - best_test_acc: 0.7768829938417812
epoch: 5 - acc: 0.805306 - best_test_acc: 0.8053055423969683
epoch: 6 - acc: 0.833254 - best_test_acc: 0.8332543818095689
epoch: 7 - acc: 0.837991 - best_test_acc: 0.8379914732354334
epoch: 8 - acc: 0.854571 - best_test_acc: 0.8545712932259593
epoch: 9 - acc: 0.858361 - best_test_acc: 0.8583609663666508
epoch: 10 - acc: 0.856466 - best_test_acc: 0.8583609663666508
epoch: 11 - acc: 0.853150 - best_test_acc: 0.8583609663666508
epoch: 12 - acc: 0.868783 - best_test_acc: 0.8687825675035529
epoch: 13 - acc: 0.877309 - best_test_acc: 0.877309332070109
epoch: 14 - acc: 0.873993 - best_test_acc: 0.877309332070109
epoch: 15 - acc: 0.877783 - best_test_acc: 0.8777830412126955
epoch: 16 - acc: 0.882994 - best_test_acc: 0.8829938417811464
epoch: 17 - acc: 0.881573 - best_test_acc: 0.8829938417811464
epoch: 18 - acc: 0.889626 - best_test_acc: 0.8896257697773567
epoch: 19 - acc: 0.877783 - best_test_acc: 0.8896257697773567
仅prompt_embed和lstm及mlp去做p-tuning，20个epoch后接近收敛，acc=0.8896，略小于finetun的acc 0.8977
附上苏神结果对比：
163 Dec 31, 2022
4.3k Jan 01, 2023
2 Jan 14, 2022
1 Dec 17, 2021
141 Dec 27, 2022
1 Jan 27, 2022
4 Jul 08, 2022
819 Dec 30, 2022
13 Jul 30, 2022
30 Dec 02, 2022
30 Dec 16, 2022
26 Oct 17, 2022
11 Oct 31, 2022
1 Feb 15, 2022
13 Dec 23, 2022
73 Nov 28, 2022
4 Sep 28, 2022
41 Dec 30, 2022
40 Nov 07, 2022
1 Nov 18, 2021