xyltt / lpt Goto Github PK

Thank you for sharing your work! Could you provide the datasets you used in the experiments?

Thanks!

First, great work! A new approad towards the prompt study. Your's study make me confirmed my experiments on P-Tuning v2.

I have some problem about parameter calculation.

self.proj_down = nn.Linear(config.hidden_size, config.proj_down_size)
self.intermediate_act_fn = nn.ReLU()
self.proj_up = nn.Linear(config.proj_down_size, config.num_prompt_tokens * config.hidden_size)

self.prompt_generator = PromptGeneratorWithNaivePromptGenerator(self.config)

PromptGeneratorWithNaivePromptGenerator(
  (proj_down): Linear(in_features=768, out_features=128, bias=True)
  (intermediate_act_fn): ReLU()
  (proj_up): Linear(in_features=128, out_features=3840, bias=True)
  (dropout): Dropout(p=0.1, inplace=False)
)

On your code, you use the same weight for all prompt generation, regardless of layers. Such should be much less.
On your paper, the tunable parameter is 792K vs 985K, compared with P-Tuning v2. In this case, the weight shall be different for each layer.

简单用中文说，从代码上看NPG可调的参数量，可能比论文中少，不知道是不是我看漏了或者算错了。按照论文中计算，在添加Prompt的每一层，Prompt都由不同的weight生成。（对比P-Tuning v2毕竟每层都加）