Update README.md
Browse files
README.md
CHANGED
@@ -12,7 +12,7 @@ widget:
|
|
12 |
|
13 |
The model is used to generate Chinese ancient poems. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-chinese-poem][poem].
|
14 |
|
15 |
-
Since the parameter skip_special_tokens is used in the pipelines.py, special tokens such as [SEP], [UNK] will be deleted,
|
16 |
|
17 |
## How to use
|
18 |
|
@@ -26,7 +26,7 @@ When the parameter skip_special_tokens is True:
|
|
26 |
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-poem")
|
27 |
>>> text_generator = TextGenerationPipeline(model, tokenizer)
|
28 |
>>> text_generator("[CLS]梅 山 如 积 翠 ,", max_length=50, do_sample=True)
|
29 |
-
|
30 |
```
|
31 |
|
32 |
When the parameter skip_special_tokens is False:
|
@@ -37,7 +37,7 @@ When the parameter skip_special_tokens is False:
|
|
37 |
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-poem")
|
38 |
>>> text_generator = TextGenerationPipeline(model, tokenizer)
|
39 |
>>> text_generator("[CLS]梅 山 如 积 翠 ,", max_length=100, do_sample=True)
|
40 |
-
|
41 |
```
|
42 |
|
43 |
## Training data
|
@@ -46,39 +46,44 @@ Training data contains 800,000 Chinese ancient poems which are collected by [chi
|
|
46 |
|
47 |
## Training procedure
|
48 |
|
49 |
-
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud
|
50 |
|
51 |
```
|
52 |
python3 preprocess.py --corpus_path corpora/poem.txt \
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
```
|
57 |
|
58 |
```
|
59 |
python3 pretrain.py --dataset_path poem_dataset.pt \
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
```
|
71 |
|
72 |
Finally, we convert the pre-trained model into Huggingface's format:
|
73 |
```
|
74 |
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path poem_gpt2_base_model.bin-200000 \
|
75 |
-
|
76 |
-
|
77 |
```
|
78 |
|
79 |
### BibTeX entry and citation info
|
80 |
|
81 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
@article{zhao2019uer,
|
83 |
title={UER: An Open-Source Toolkit for Pre-training Models},
|
84 |
author={Zhao, Zhe and Chen, Hui and Zhang, Jinbin and Zhao, Xin and Liu, Tao and Lu, Wei and Chen, Xi and Deng, Haotang and Ju, Qi and Du, Xiaoyong},
|
|
|
12 |
|
13 |
The model is used to generate Chinese ancient poems. You can download the model either from the [GPT2-Chinese Github page](https://github.com/Morizeyao/GPT2-Chinese), or via HuggingFace from the link [gpt2-chinese-poem][poem].
|
14 |
|
15 |
+
Since the parameter skip_special_tokens is used in the pipelines.py, special tokens such as [SEP], [UNK] will be deleted, the output results of Hosted inference API (right) may not be properly displayed.
|
16 |
|
17 |
## How to use
|
18 |
|
|
|
26 |
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-poem")
|
27 |
>>> text_generator = TextGenerationPipeline(model, tokenizer)
|
28 |
>>> text_generator("[CLS]梅 山 如 积 翠 ,", max_length=50, do_sample=True)
|
29 |
+
[{'generated_text': '[CLS]梅 山 如 积 翠 , 丛 竹 隠 疏 花 。 水 影 落 寒 濑 , 竹 声 随 暮 鸦 。 茅 茨 数 间 屋 , 烟 火 两 三 家 。 安 得 携 琴 酒 , 相 逢 烟 雨 赊 。 向 湖 边 过 , 偏 怜 雪 里 看 。 浮 峦 如 画 出 , 远 树 与 天 连 。 月 上 僧 房 静 , 风 回 萤 火 寒 。 幽 情 何 可 写 , 赖 有 子 期 弹 。 棠 真'}]
|
30 |
```
|
31 |
|
32 |
When the parameter skip_special_tokens is False:
|
|
|
37 |
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-poem")
|
38 |
>>> text_generator = TextGenerationPipeline(model, tokenizer)
|
39 |
>>> text_generator("[CLS]梅 山 如 积 翠 ,", max_length=100, do_sample=True)
|
40 |
+
[{'generated_text': '[CLS]梅 山 如 积 翠 , 秀 出 何 其 雄 。 矫 矫 云 间 质 , 映 日 生 玲 珑 。 根 大 乱 石 结 , 枝 高 青 云 蒙 。 常 因 风 露 晚 , 隠 映 瑶 台 中 。 忽 闻 山 石 裂 , 万 里 吹 天 风 。 又 觉 此 身 高 , 迥 出 凡 境 空 。 清 影 落 潭 水 , 暗 香 来 逈 峰 。 却 寻 白 太 白 , 月 影 摇 江 东 。 [SEP] 而 非'}]
|
41 |
```
|
42 |
|
43 |
## Training data
|
|
|
46 |
|
47 |
## Training procedure
|
48 |
|
49 |
+
The model is pre-trained by [UER-py](https://github.com/dbiir/UER-py/) on [Tencent Cloud](https://cloud.tencent.com/). We pre-train 200,000 steps with a sequence length of 128. We use extended vocabulary to handle out-of-vocabulary words. The Chinese character that occurs greater than or equal to 100 in poem corpus is added to the vocabulary.
|
50 |
|
51 |
```
|
52 |
python3 preprocess.py --corpus_path corpora/poem.txt \
|
53 |
+
--vocab_path models/poem_zh_vocab.txt \
|
54 |
+
--dataset_path poem_dataset.pt --processes_num 16 \
|
55 |
+
--seq_length 128 --target lm
|
56 |
```
|
57 |
|
58 |
```
|
59 |
python3 pretrain.py --dataset_path poem_dataset.pt \
|
60 |
+
--vocab_path models/poem_zh_vocab.txt \
|
61 |
+
--config_path models/gpt2/config.json \
|
62 |
+
--output_model_path models/poem_gpt2_model.bin \
|
63 |
+
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
|
64 |
+
--total_steps 200000 --save_checkpoint_steps 50000 --report_steps 1000 \
|
65 |
+
--learning_rate 5e-4 --batch_size 64 \
|
66 |
+
--embedding word_pos --remove_embedding_layernorm \
|
67 |
+
--encoder transformer --mask causal --layernorm_positioning pre \
|
68 |
+
--target lm --tie_weight
|
|
|
69 |
```
|
70 |
|
71 |
Finally, we convert the pre-trained model into Huggingface's format:
|
72 |
```
|
73 |
python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path poem_gpt2_base_model.bin-200000 \
|
74 |
+
--output_model_path pytorch_model.bin \
|
75 |
+
--layers_num 12
|
76 |
```
|
77 |
|
78 |
### BibTeX entry and citation info
|
79 |
|
80 |
```
|
81 |
+
@article{radford2019language,
|
82 |
+
title={Language Models are Unsupervised Multitask Learners},
|
83 |
+
author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
|
84 |
+
year={2019}
|
85 |
+
}
|
86 |
+
|
87 |
@article{zhao2019uer,
|
88 |
title={UER: An Open-Source Toolkit for Pre-training Models},
|
89 |
author={Zhao, Zhe and Chen, Hui and Zhang, Jinbin and Zhao, Xin and Liu, Tao and Lu, Wei and Chen, Xi and Deng, Haotang and Ju, Qi and Du, Xiaoyong},
|