search

luminxu / ViPNAS

The official repo for CVPR2021——ViPNAS: Efficient Video Pose Estimation via Neural Architecture Search.
MIT License
47 stars 11 forks source link

About search algorithm #15

Open helloyuning opened 1 year ago

helloyuning commented 1 year ago

Thanks for your grate work

I am new to this profession, so I have some basic questions about VIPNAS:

As stated in the paper, VIPNAS searches five dimensions to find the optimal structure. But there is no mention of what kind of policy to adjust the model. For example, the author said that the current frame and the next frame will use different structures, so what policies or conditions are the structure adjustments based on? In short, the model is adjusted based on what changes

Sorry for my basic question

Best

jin-s13 commented 1 year ago

"what policies or conditions are the structure adjustments based on? " ---Are you asking about what aspects of the model can be changed? ---> The five dimension search space includes network depth, width, kernel size, group number, and attentions.

The goal of network search is to find the best model with lowest flops.

helloyuning commented 1 year ago

Tanks for your reply!

helloyuning commented 1 year ago

Have you ever trained vipnas_res34? Do you have pre-trained nas_res34 network weights?