search

xiaolai-sqlai / mobilenetv3

mobilenetv3 with pytorch，provide pre-train model
MIT License
1.6k stars 340 forks source link

Can anyone really reproduce the result with the code? #19

Open triangleCZH opened 4 years ago

triangleCZH commented 4 years ago

To be honest I think the previous 75+ acc was a misreport because his previous version has the attention part with much more params.

cxxgtxy commented 4 years ago

@triangleCZH I find 3 models which is better than v3large based on NAS, please refer to https://github.com/xiaomi-automl/MoGA

nullptr-leo commented 4 years ago

@triangleCZH I find 3 models which is better than v3large based on NAS, please refer to https://github.com/xiaomi-automl/MoGA

@cxxgtxy MoGA is created for GPU. How about the performance on CPU?

cxxgtxy commented 4 years ago

@leo-nullptr Based on the paper, MoGA's peformance on mobile CPU is highly on par with MobileNetV3 too.

nullptr-leo commented 4 years ago

@cxxgtxy Thanks. I will read it later. Are you the author of MoGA?

cxxgtxy commented 4 years ago

@leo-nullptr Welcome. I recommend because I have tested it.

triangleCZH commented 4 years ago

@cxxgtxy Thank you. I will check the repo. Do they offer training scripts by the way?

cxxgtxy commented 4 years ago

@triangleCZH I think it's not difficult, because they train their models use google's tricks.