jinyong-gpt2 / README.md
supermy's picture
Create README.md
1369277
|
raw
history blame
5.07 kB

language: zh datasets: jinyong inference: parameters: max_length: 108 num_return_sequences: 1 do_sample: True widget:

  • text: "杨过朗声说道:今番良晤,豪兴不浅,他日江湖相逢,再当杯酒言欢。咱们就此别过。 -" example_title: "神雕侠侣"
    • text: "乱世之际,人不如狗。 -" example_title: "射雕英雄传"

飞雪连天射白鹿,笑书神侠倚碧鸳

Model description

AI生成金庸小说,给出开头续写。

How to use

使用 pipeline 调用模型:

>>> # 调用微调后的模型
>>> senc="这些雪花落下来,多么白,多么好看.过几天太阳出来,每一片 雪花都变得无影无踪.到得明年冬天,又有许很多多雪花,只不过已不是 今年这些雪花罢了。"
>>> model_id="jinyong-gpt2-finetuning"
>>> from transformers import AutoTokenizer, GPT2LMHeadModel, TextGenerationPipeline

>>> tokenizer = AutoTokenizer.from_pretrained(model_id) 
>>> model = GPT2LMHeadModel.from_pretrained(model_id)
>>> text_generator = TextGenerationPipeline(model, tokenizer)   
>>> text_generator.model.config.pad_token_id = text_generator.model.config.eos_token_id
>>> text_generator( senc,max_length=108, do_sample=True)
[{'generated_text': '这些雪花落下来,多么白,多么好看.过几天太阳出来,每一片 雪花都变得无影无踪.到得明年冬天,又有许很多多雪花,只不过已不是 今年这些雪花罢了。 反正 老天爷 有眼 , 不知 哪里 是甚么 风 险 ?” 正 说到此处 , 突然 听得 谢逊 啸声 渐近 , 忍不住 张口 惊呼 , 一齐 向他 扑去 , 只听 谢逊 一声 怒吼 , 跟着 左手 用力 拍 出一掌 , 以 掌力 化开 。 众人 吃了一惊 , 同时 从 海 道 中 跃出 , 双双 倒退 。 张翠山和殷素素 对望一眼 , 均想 以 这两 大高手 之力 如何 抵挡 , 以 今日 之力 如何 攻敌 之'}]
>>> 

Here is how to use this model to get the features of a given text in PyTorch:

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("supermy/jinyong-gpt2")

model = AutoModelForCausalLM.from_pretrained("supermy/jinyong-gpt2")

Training data

此数据集基于金庸的【飞雪连天射白鹿,笑书神侠倚碧鸳】小说集训练。

统计信息


Training procedure

基于模型:GPT2 训练环境:英伟达16G显卡

bpe分词:"vocab_size"=30000

[INFO|trainer.py:1608] 2022-12-02 19:52:59,024 >> ***** Running training *****
[INFO|trainer.py:1609] 2022-12-02 19:52:59,024 >>   Num examples = 9443
[INFO|trainer.py:1610] 2022-12-02 19:52:59,024 >>   Num Epochs = 108
[INFO|trainer.py:1611] 2022-12-02 19:52:59,024 >>   Instantaneous batch size per device = 12
[INFO|trainer.py:1612] 2022-12-02 19:52:59,024 >>   Total train batch size (w. parallel, distributed & accumulation) = 12
[INFO|trainer.py:1613] 2022-12-02 19:52:59,024 >>   Gradient Accumulation steps = 1
[INFO|trainer.py:1614] 2022-12-02 19:52:59,024 >>   Total optimization steps = 84996
[INFO|trainer.py:1616] 2022-12-02 19:52:59,025 >>   Number of trainable parameters = 124439808

{'loss': 8.0431, 'learning_rate': 4.970998635229893e-05, 'epoch': 0.64}
{'loss': 7.4867, 'learning_rate': 4.94158548637583e-05, 'epoch': 1.27}
{'loss': 7.322, 'learning_rate': 4.912172337521766e-05, 'epoch': 1.91}
......
......
......
{'loss': 3.8686, 'learning_rate': 9.035719327968376e-07, 'epoch': 106.1}
{'loss': 3.8685, 'learning_rate': 6.094404442562004e-07, 'epoch': 106.73}
{'loss': 3.8678, 'learning_rate': 3.1530895571556306e-07, 'epoch': 107.37}

{'train_runtime': 71919.9835, 'train_samples_per_second': 14.18, 'train_steps_per_second': 1.182, 'train_loss': 4.661963973798675, 'epoch': 108.0}
***** train metrics *****
  epoch                    =       108.0
  train_loss               =       4.662
  train_runtime            = 19:58:39.98
  train_samples            =        9443
  train_samples_per_second =       14.18
  train_steps_per_second   =       1.182
12/03/2022 15:51:42 - INFO - __main__ - *** Evaluate ***
[INFO|trainer.py:2929] 2022-12-03 15:51:42,270 >> ***** Running Evaluation *****
[INFO|trainer.py:2931] 2022-12-03 15:51:42,270 >>   Num examples = 283
[INFO|trainer.py:2934] 2022-12-03 15:51:42,270 >>   Batch size = 12
100%|██████████| 24/24 [00:07<00:00,  3.17it/s]
[INFO|modelcard.py:449] 2022-12-03 15:51:52,077 >> Dropping the following result as it does not have all the necessary fields:
{'task': {'name': 'Causal Language Modeling', 'type': 'text-generation'}, 'metrics': [{'name': 'Accuracy', 'type': 'accuracy', 'value': 0.2100502721055507}]}
***** eval metrics *****
  epoch                   =      108.0
  eval_accuracy           =     0.2101
  eval_loss               =      6.889
  eval_runtime            = 0:00:07.90
  eval_samples            =        283
  eval_samples_per_second =      35.79
  eval_steps_per_second   =      3.035
  perplexity              =   981.4321