JiaangL / RandomQuantization

The code of paper Random Entity Quantization for Parameter-Efficient Compositional Knowledge Graph Representation, EMNLP 2023 (Oral)
4 stars 0 forks source link

About Parameter Efficient #1

Closed lainanhui closed 1 month ago

lainanhui commented 1 month ago

作者你好,我想请问你是否尝试过计算codeword_weight所占据的参数量呢,虽然它是不需要更新梯度的,但是它所占据的参数量非常大,比entity_embedding所占据的参数量都大。 是因为codeword_weight即使不设置成parameter,也不会影响训练过程,设置成parameter只是为了代码更加清晰吗?

JiaangL commented 1 month ago

你好,我们只统计了可学习参数量,其中不包含codeword_weght。EARL使用的codeword_weight由知识图谱的拓扑性质决定,在模型训练时维持不变。 codeword_weight可以被设置成parameter类型,只要确保其不接收梯度进行更新。

JiaangL commented 1 month ago

The reason why codeword weight takes up more parameters than entity embedding is that all entities need codeword weight, but only a small number of entities require independent embedding. I'm closing this issue for now, but feel free to follow up if you have any questions.