I'm curious why swish and sigmoid were used here instead of hard-swish and hard-sigmoid as per MobileNet-V3? Is there a specific mobile hardware target in mind where sigmoid isn't a concern?
Introducing MixNet: AutoML + a new mixed depthwise conv (MDConv). SOTA results for mobile: 78.9% ImageNet top-1 accuracy under typical mobile settings (<600M FLOPS). Paper: https://arxiv.org/abs/1907.09595 Code & models: https://github.com/tensorflow/tpu/tree/master/models/official/mnasnet/mixnet …pic.twitter.com/qP3XbpQ7Zb
-
-
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Why didn't you compare to your own Mobilenetv3?
- Još 1 odgovor
Novi razgovor -
-
-
Cool work. The reason that larger kernels hurt could be that the ideal kernel shape is a circle, not a square. Approximating a circular shape at 5x5 or above will also give a bit of a speedup if nothing else.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Looks super interesting, looking forward to reading it in depth.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.