关于预训练模型大小

[dazaier]

您好！
我下载了预训练模型，有个问题想要请教:为什么不同的数据集的预训练模型会大小不一样呀，我的理解当网络结构确定，模型大小不应该是固定的吗？

[yuell-science]

Thanks for raising the interesting question on different model sizes per dataset! In our model architecture, we are inheriting the multi-path design in this paper: https://openaccess.thecvf.com/content_CVPR_2020/papers/Sundermeyer_Multi-Path_Learning_for_Object_Pose_Estimation_Across_Domains_CVPR_2020_paper.pdf. Basically it uses separate decoders per object. Indeed, this design is also adopted in several follow-up works. As different datasets have different number of objects, the model size will vary. Let me know if you have further questions.

[dazaier]

Thanks for your quick response!
Since each object has its own decoder, as the number of object categories in the dataset increases, the size of the model also increases. I observed that for the LMO dataset with 15 classes, the model size is 1.5GB. However, for the YCBV dataset with 21 classes, the model size increases to 2.4GB. Is my understanding correct?

[yuell-science]

It is not necessarily proportional since the model comprises encoder + decoder + classification and regression heads. Only the decoder size is proportional to the number of objects.

[yuell-science]

Closing as no activity happens for a week.