MNasNet

[pool1892]

Hi, thank you for your work and for open sourcing your code!

As you don't list mail addresses in the paper I abuse the Github issue tracker to ask a question about the paper:
Have you compared your work to MNasNet (https://arxiv.org/abs/1807.11626)?
MNasNet is stronger than Mobilenet v2 across different FLOP settings and achieves 75% TOP-1 imagenet accuracy with a comparable compute envelope as you used for Figure 3c. Is there a reason why you excluded MNasNet from your comparisons?

Thanks,
Christoph

[sacmehta]

MNasNet is a more on an architectural space search rather than light-weight model design and is complementary to our work.

MNasNet uses mobileNet blocks and we have shown that EESP modules are better in learning representations than MobileNet modules. If a similar search is done using EESP module, I believe that will give better performance.

P.S: Searching over 8K design choices, MNAsNet delivers slightly better performance to our work (73 top-1 with 3.6 M params and 270 M FLOps) while our unit delivers 72.1 under similar budget.