search

tensorflow / tpu

Reference models and tools for Cloud TPUs.
https://cloud.google.com/tpu/
Apache License 2.0
5.2k stars 1.77k forks source link

Command to reproduce CondConv's Result #684

Open ildoonet opened 4 years ago

ildoonet commented 4 years ago

I use below arguments to train condconv-b0 model, which performs 21.7 top1 error reported in the original paper. But I got poor performance such as 22.67 top 1 error. 

--model_name efficientnet-condconv-b0-8e --augment_name autoaugment --mixup_alpha 0.2

  1. Can you provide the specific arguments to reproduce?

  2. Can you provide your training logs, such as tensorboard 'top_1_accuracy' graph?

Thanks.

ildoonet commented 4 years ago

image

Dark Blue is for B0 and Sky Blue is for B0 with CondConv. I ran this with the above command.

ildoonet commented 4 years ago

Sorry, Dark blue is for b1. But, still, B0 with CondConv can not achieve the performance as the original paper claims.

Xingrun-Xing commented 3 years ago

Sorry, Dark blue is for b1. But, still, B0 with CondConv can not achieve the performance as the original paper claims.

Hi, I have met the same problem in CIFAR10. CondConv seems not to work. Do you solve this problem now, thanks!