【LLM Review】Decrypt Ptuning - 2023W32

参考论文：P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks

1. 总结

ptuning 是介于prompt engineer和Full FT之间的一种方案，目的是希望用少量训练成本，优化Handcrafted prompt效果，接近Full FT效果，具体效果和model size和task有关，整体看有一定实践价值。

2. ptuning、prompt tuning和prefix tuning的区别

ptuning v1即prompt tuning，prompt virtual token 加入位置不固定且一般为一个token，只在Embedding labyer加 。v2类似prefix tuning ，但部分task去掉了reparameterized，还有一些其他trick，见table1。

• Full FT的问题：While fine-tuning obtains good performance, it is memory-consuming during training because gradients and optimizer states for all parameters must be stored. Moreover, keeping a copy of model parameters for each task during inference is inconvenient since pre-trained models are usually large. （pt的优势：降低训练内存，同时每个task只需存pt部分的trainable variables）
• Fixed Handcraft prompt的问题：discrete prompting (Shin et al., 2020; Gao et al., 2020) can lead to suboptimal performance in many cases compared to fine-tuning. （人类认为合理的prompt不一定对语言模型有效。 其次，手工制定的prompt的微小变化会导致显着的性能差异。）
  • 输入Template：[CLS] I like the Disney films very much. [SEP] It was [MASK]. [SEP]
  • Verbalizer（label word）：具体的分类任务，需要选择指定的标签词（label word）。例如情感分析中，我们期望Verbalizer可能是 V(positive)=great ，  V(negative)=terrible  （positive和negative是类标签）。同样，不同的任务有其相应的label word，但需要注意的是，Verbalizer的构建需要取决于对应的Pattern。因此如何构建Verbalizer是另一个研究挑战。
• Prompt Turning（continuous prompts）的问题：While prompt tuning improves over discrete prompt on many tasks, it still underperforms fine-tuning when the model size is not large, specifically less than 10 billion parameters.
• Prefix tuning - Deep prompt tuning：The main difference between prompt tuning and prefix tuning is that prompt tuning only fine-tunes the embedding layer, while prefix tuning fine-tunes all the layers corresponding to the virtual tokens1.
  • The most significant improvement originates from appling continuous prompts for every layer of the pretrained model, instead of the mere input layer.
• Performance：Vary depending on the specific task. For example, in Table2Text tasks, prompt tuning with a parameter size of only 0.1% can outperform fine-tuning, while in Xsum summary tasks (NLU), like hard sequence tagging tasks (QA), prompt tuning may have slightly worse performance compared to fine-tuning. It is important to consider the task requirements and model scale when choosing between these techniques.

3. ptuning v2 实现代码

class RobertaPrefixForTokenClassification(RobertaPreTrainedModel):

# deep prompt tuning
def get_prompt(self, batch_size):
	prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(self.roberta.device)
	past_key_values = self.prefix_encoder(prefix_tokens)
	past_key_values = past_key_values.view(
		batch_size,
		self.pre_seq_len,
		self.n_layer * 2, 
		self.n_head,
		self.n_embd
	)
	past_key_values = self.dropout(past_key_values)
	past_key_values = past_key_values.permute([2, 0, 3, 1, 4]).split(2)
	return past_key_values

# reparameterized encoder（MLP）
class PrefixEncoder(torch.nn.Module):
    r'''
    The torch.nn model to encode the prefix

    Input shape: (batch-size, prefix-length)

    Output shape: (batch-size, prefix-length, 2*layers*hidden)
    '''
    def __init__(self, config):
        super().__init__()
        self.prefix_projection = config.prefix_projection
        if self.prefix_projection:
            # Use a two-layer MLP to encode the prefix
            self.embedding = torch.nn.Embedding(config.pre_seq_len, config.hidden_size)
            self.trans = torch.nn.Sequential(
                torch.nn.Linear(config.hidden_size, config.prefix_hidden_size),
                torch.nn.Tanh(),
                torch.nn.Linear(config.prefix_hidden_size, config.num_hidden_layers * 2 * config.hidden_size)
            )
        else:
            self.embedding = torch.nn.Embedding(config.pre_seq_len, config.num_hidden_layers * 2 * config.hidden_size)

    def forward(self, prefix: torch.Tensor):
        if self.prefix_projection:
            prefix_tokens = self.embedding(prefix)
            past_key_values = self.trans(prefix_tokens)
        else:
            past_key_values = self.embedding(prefix)
        return past_key_values