Skip to content

Latest commit

 

History

History
206 lines (154 loc) · 11.2 KB

README_en.md

File metadata and controls

206 lines (154 loc) · 11.2 KB

ChatGLM-6B-PT

This repository implements tuning of the ChatGLM-6B model based on P-Tuning v2. P-Tuning v2 reduces the amount of parameters that need to be optimized to 0.1% of the full fine-tuning, and then through model quantization, Gradient Checkpoint and other methods, it only needs a minimum of 7GB of video memory to run.

The following uses the ADGEN (advertising generation) dataset as an example to introduce how to use the code.

Software dependencies

Running p-tuning requires version 4.27.1 of transformers. In addition to the dependencies of ChatGLM-6B, the following dependencies are required

pip install rouge_chinese nltk jieba datasets

Instructions

Download the dataset

The task of the ADGEN dataset is to generate an advertisement word (summary) based on the input (content).

{
    "content": "类型#上衣*版型#宽松*版型#显瘦*图案#线条*衣样式#衬衫*衣袖型#泡泡袖*衣款式#抽绳",
    "summary": "这件衬衫的款式非常的宽松,利落的线条可以很好的隐藏身材上的小缺点,穿在身上有着很好的显瘦效果。领口装饰了一个可爱的抽绳,漂亮的绳结展现出了十足的个性,配合时尚的泡泡袖型,尽显女性甜美可爱的气息。"
}

From Google Drive or Tsinghua Cloud download the processed ADGEN dataset, and put the decompressed AdvertiseGen directory into this directory.

Training

P-Tuning v2

Run the following commands for training:

bash train.sh

PRE_SEQ_LEN and LR in train.sh are soft prompt length and training learning rate respectively, which can be adjusted to achieve the best results. The P-Tuning-v2 method will freeze all model parameters, and the quantization level of the original model can be adjusted by adjusting quantization_bit. If this option is not added, it will be loaded with FP16 precision.

Under the default configuration of per_device_train_batch_size=1, gradient_accumulation_steps=16, the model parameters of INT4 are frozen, and a training iteration will perform 16 cumulative forward and backward propagations with a batch size of 1, which is equivalent to the total batch size of 16, and only 6.7G GPU memory is required at this time with quantization_bit=4. If you want to improve the training efficiency under the same batch size, you can increase the value of per_device_train_batch_size while keeping the product of the two unchanged, but it will also bring more GPU memory consumption, please adjust it according to the actual situation.

If you want to load the model locally, you can change THUDM/chatglm-6b in train.sh to your local model path.

Finetune

To finetune the full parameters, you need to install Deepspeed, and then run the following command:

bash ds_train_finetune.sh

Inference

During P-tuning v2 training, the model only saves the parameters of the PrefixEncoder part, so the original ChatGLM-6B model and the weight of the PrefixEncoder need to be loaded at the same time during inference, and the arguments need to be specified in evaluate.sh:

--model_name_or_path THUDM/chatglm-6b
--ptuning_checkpoint $CHECKPOINT_PATH

It is still compatible with the old version of Checkpoint saved with full parameters, just set model_name_or_path as before:

--model_name_or_path $CHECKPOINT_PATH

The evaluation indicators are Chinese Rouge score and BLEU-4. The generated results are saved in ./output/adgen-chatglm-6b-pt-8-1e-2/generated_predictions.txt.

Example

Example 1

  • Input: 类型#上衣*材质#牛仔布*颜色#白色*风格#简约*图案#刺绣*衣样式#外套*衣款式#破洞
  • Label: 简约而不简单的牛仔外套,白色的衣身十分百搭。衣身多处有做旧破洞设计,打破单调乏味,增加一丝造型看点。衣身后背处有趣味刺绣装饰,丰富层次感,彰显别样时尚。
  • Output[before tuning]: 这件上衣的材质是牛仔布,颜色是白色,风格是简约,图案是刺绣,衣样式是外套,衣款式是破洞。
  • Output[after tuning]: 这是一款简约的牛仔外套,破洞设计,将牛仔布破洞,带来一种随意与个性。破洞的牛仔外套,展现出时尚气息,带来一种休闲感。同时,刺绣图案,让整件外套更加立体。

Example 2

  • Input: 类型#裙*版型#显瘦*风格#文艺*风格#简约*图案#印花*图案#撞色*裙下摆#压褶*裙长#连衣裙*裙领型#圆领
  • Label: 文艺个性的印花连衣裙,藏青色底蕴,低调又大气,撞色太阳花分布整个裙身,绚丽而美好,带来时尚减龄的气质。基础款的舒适圆领,简约不失大方,勾勒精致脸庞。领后是一粒包布扣固定,穿脱十分方便。前片立体的打褶设计,搭配后片压褶的做工,增添层次和空间感,显瘦又有型。
  • Output[before tuning]: 类型#裙版型#显瘦风格#文艺风格#简约图案#印花图案#撞色裙下摆#压褶裙长#连衣裙裙领型#圆领 1. 连衣裙:简约风格,裙长为膝盖以上,裙领型为圆领。2. 裙下摆:压褶设计,使裙摆呈现出流畅的褶皱效果。3. 裙领型:裙领型为圆领,使穿上连衣裙后更加有型。4. 版型:采用显瘦设计,让连衣裙看起来更加苗条。5. 风格:文艺风格,让连衣裙更加有内涵和品味。6. 图案:印花设计,在连衣裙上印有独特的图案。7. 撞色:采用撞色设计,让连衣裙在色彩上更加鲜明、富有层次感。
  • Output[after tuning]: 这是一款文艺范的连衣裙,以印花为元素,采用简约的印花,既能够突出文艺气质,又能够展现简约风。在印花的同时又有领子和裙摆的压褶设计,更加凸显文艺气质。简约而不会过于单调,搭配出街,穿着十分舒适。

evaluation result

Finetune P-tuning v2 LoRA
BLEU-4 8.01 8.10 7.62
Rouge-1 31.23 31.12 30.60
Rouge-2 7.36 7.11 6.96
Rouge-l 25.08 24.97 24.80
Training Loss 3.00 3.74 3.32

Experiment Settings

max_source_length=64
max_target_length=64
max_steps=3000
P-tuning v2
pre_seq_len=128
learning_rate=2e-2
quantization_bit=4
per_device_train_batch_size=16
gradient_accumulation_steps=1
Finetune
learning_rate=1e-4
fp16
num_gpus=4
per_device_train_batch_size=4
gradient_accumulation_steps=1
LoRA

The implementation uses simple_thu_chatglm6b

learning_rate=5e-4
per_device_train_batch_size=16
gradient_accumulation_steps=1

Model Deployment

First load the tokenizer:

from transformers import AutoConfig, AutoModel, AutoTokenizer

# Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
  1. If a new Checkpoint needs to be loaded (only contains the PrefixEncoder parameter):
config = AutoConfig.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True, pre_seq_len=128)
model = AutoModel.from_pretrained("THUDM/chatglm-6b", config=config, trust_remote_code=True)
prefix_state_dict = torch.load(os.path.join(CHECKPOINT_PATH, "pytorch_model.bin"))
new_prefix_state_dict = {}
for k, v in prefix_state_dict.items():
     if k.startswith("transformer.prefix_encoder."):
         new_prefix_state_dict[k[len("transformer.prefix_encoder."):]] = v
model.transformer.prefix_encoder.load_state_dict(new_prefix_state_dict)

Note that you may need to change pre_seq_len to the actual value of your training. If you load model from local, you need to change THUDM/chatglm-6b to the local model path (not the checkpoint path).

  1. If you need to load the old checkpoint (including both ChatGLM-6B and PrefixEncoder parameters), or perform full parameter fine-tuning, then directly load the entire checkpoint:
model = AutoModel.from_pretrained(CHECKPOINT_PATH, trust_remote_code=True)

Then it can be quantified according to the needs, or it can be used directly:

# Comment out the following line if you don't use quantization
model = model. quantize(4)
model = model.half().cuda()
model.transformer.prefix_encoder.float()
model = model.eval()

response, history = model.chat(tokenizer, "Hello", history=[])

[23/04/19] You can also directly run web demo which supports loading P-Tuning v2 checkpoint

bash web_demo.sh

It may be necessary to modify the content of web_demo.sh to match your actual checkpoint situation.

Use your own dataset

Modify train_file, validation_file and test_file in train.sh and evaluate.sh to your own JSON format dataset paths, and change prompt_column and response_column to the keys in the JSON file corresponding to input text and output text. You may also need to increase max_source_length and max_target_length to match the maximum input and output lengths in your own dataset.

Dialog Dataset

If you need to use multiple rounds of dialogue data to train the model, you can provide chat history. For example, the following is the training data for a three-round dialogue:

{"prompt": "长城h3风扇不转。继电器好的。保险丝好的传感器新的风扇也新的这是为什么。就是继电器缺一个信号线", "response": "用电脑能读数据流吗?水温多少", "history": []}
{"prompt": "95", "response": "上下水管温差怎么样啊?空气是不是都排干净了呢?", "history": [["长城h3风扇不转。继电器好的。保险丝好的传感器新的风扇也新的这是为什么。就是继电器缺一个信号线", "用电脑能读数据流吗?水温多少"]]}
{"prompt": "是的。上下水管都好的", "response": "那就要检查线路了,一般风扇继电器是由电脑控制吸合的,如果电路存在断路,或者电脑坏了的话会出现继电器不吸合的情况!", "history": [["长城h3风扇不转。继电器好的。保险丝好的传感器新的风扇也新的这是为什么。就是继电器缺一个信号线", "用电脑能读数据流吗?水温多少"], ["95", "上下水管温差怎么样啊?空气是不是都排干净了呢?"]]}

During training, you need to specify --history_column as the key of the chat history in the data (history in this example), and the chat history will be stitched automatically. Note that content exceeding the input length max_source_length will be truncated.

You can refer to the following instructions:

bash train_chat.sh

Citation

@inproceedings{liu2022p,
   title={P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks},
   author={Liu, Xiao and Ji, Kaixuan and Fu, Yicheng and Tam, Weng and Du, Zhengxiao and Yang, Zhilin and Tang, Jie},
   booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)},
   pages={61--68},
   year={2022}
}