PLMTuningCompetition icon indicating copy to clipboard operation
PLMTuningCompetition copied to clipboard

擂台赛3-大规模预训练调优比赛的示例代码与baseline实现

Baseline

Prepare your environment

The implementation of Black-Box Tuning is quite simple, you can check our code and easily implement it in your own environment. Or you can create a new environment to run our implementation, which is based on pycma, Transformers and FastNLP. Optionally, you can use fitlog to monitor experimental results. You can uncomment the fitlog-related lines in our code to use it.

conda create --name bbt python=3.8
conda activate bbt
pip install transformers==4.1.1
pip install datasets
pip install fastNLP
pip install cma
pip install sklearn
pip install cryptography

Using BBT

Now you can run Black-Box Tuning with the following code:

python bbt.py --seed 42 --task_name 'sst2'

Then the best prompt will be saved to ./results. Run test.py to get the predictions on the test set.

Test script

输入接口

def test_api(
    sentence_fn: Callable[[str], str], 
    test_data_path: str,
    embedding_and_attention_mask_fn: Callable[[Tensor, Tensor], Tuple[Tensor, Tensor]]=None, 
    hidden_states_and_attention_mask_fn:
    	Callable[[int, Tensor, Tensor], Tuple[Tensor, Tensor]]=None,
    task_name: str='SST-2', 
    batch_size: int=32, 
    device: Optional[str, torch.device]='cuda': 
) -> Tuple[Union[Tensor, None], Tensor, Union[Tensor, None]]:
    ...

==这个函数是一个生成器==

Sentence_fn

对test set中的句子添加(自然语言组成的)离散模板的函数,以待预测的句子作为参数,返回处理后的句子。见下例:

pre_str = tokenizer.decode(list(range(1000, 1050))) + ' . '
middle_str = ' ? <mask> , '

def sentence_fn(test_data):
    return pre_str + test_data + middle_str + test_data
  • 注:sentence_fn的参数并非真实的数据,只是一个占位符,因此不能直接对数据的长度做操作。在测试阶段,我们统一把所有长于256 token的句子切割到256个token。
test_data_path

赛方提供的加密测试数据路径

embedding_and_attention_mask_fn

适用于仅在embedding层之后加入的soft prompt,通过操作embedding layer得到的表征和对应的attention_mask并按相同顺序返回,其出现在RobertaEmbedding和RobertaEncoder层之间。见下例:

 def embedding_and_attention_mask_fn(embedding, attention_mask):
        prepad = torch.zeros(size=(1, 1024), device=device)
        pospad = torch.zeros(size=(embedding.size(1) - 51, 1024), device=device)
        return embedding + torch.cat([prepad, best_prompt, pospad]), attention_mask
  • Shape: embedding bsz, seq_len, hidden_size, attention_mask bsz, seq_len
hidden_states_and_attention_mask_fn

适用于在embedding层之后和除最后一层encoder layer得到的结果中均加入的soft prompt(每层的prompt内容可以不同)见下例:

 def hidden_states_and_attention_mask_fn(i, embedding, attention_mask):
        prepad = torch.zeros(size=(1, 1024), device=device)
        pospad = torch.zeros(size=(embedding.size(1) - 51, 1024), device=device)
        return embedding + torch.cat([prepad, best_prompt[i], pospad]), attention_mask
  • i: 是在[0, num_layer)中的整数,代表prompt出现的层数,其中i=0对应embedding层之后
  • Shape: embedding bsz, seq_len, hidden_size, attention_mask bsz, seq_len

该函数在一次forward中被调用num_layer

==以上两个函数可以均为None,或其中一个为None,对应的函数将不起作用,但两参数都非None是意义不明的行为,因此会报错==

task_name

['SNLI', 'SST-2', 'MRPC', 'AGNews', 'Yelp', 'TREC']中的一个,需要与测试文件路径匹配

batch_size

测试采用的batch_size

device

测试时使用的模型和中间生成Tensor的device

生成值

Logits:一个batch中,模型预测的mask_pos位置对应的lm_logit,shape为batch_size, vocab_size

  • 最后一个batch可能不满
  • 如果sentence_fn中没有,则默认使用的是不基于MLM的学习方法,这个返回值将为None

Hidden_states: 一个batch中,每一层的所有hidden states,shape为batch_size, num_layer + 1, seq_len, hidden_size

Mask_pos: 一个batch中,每一句句子的mask位置,与Logits相同,如无mask token则返回None

More details...

For more technical details, check out our paper

If you meet any problem, you can seek our help in the Wechat group.