some doubts about Master Net design
LianShuaiLong opened this issue · comments
hi,first of all, thanks for your insteresting work;
however, i have some doubts about Master Net design,it is said you have designed 20 networks to select best Master Net, I want to know how you designed these networks or in other word i want to know the details of these 20 networks, i am looking forward to your reply
All Master Nets are listed in the appendix (p.12). Authors probably designed them using their own understanding of what should work and what shouldn't. I don't think there is much to say about their design choices.
Hi LianShuaiLong,
Thank you for your feedback! And thank you @bonlime for kind reply!
As @bonlime explained, we manually design these networks based on our understanding of what makes a good network. In our design, we intensionally explore different design strategies, such as using all DW-Blocks, or using DW-Blocks in low-level and XX-Blocks in high-level, to see which one is better. As we expected, any GPU efficient networks follow our design space. To design the 20 networks, we did our best. Usually a less carefully designed network would achieve top-1 accuracy around 75%~76%. So these 20 networks are all good enough structures.
Hi LianShuaiLong,
Thank you for your feedback! And thank you @bonlime for kind reply!As @bonlime explained, we manually design these networks based on our understanding of what makes a good network. In our design, we intensionally explore different design strategies, such as using all DW-Blocks, or using DW-Blocks in low-level and XX-Blocks in high-level, to see which one is better. As we expected, any GPU efficient networks follow our design space. To design the 20 networks, we did our best. Usually a less carefully designed network would achieve top-1 accuracy around 75%~76%. So these 20 networks are all good enough structures.
thanks for you kind reply