reproducing results of GENets

[shairoz-deci]

Thank you for sharing your code.
I've tried reproducing your results using the training hyperparameters from the paper but was unable to.
I used the popular timm package for training (https://github.com/rwightman/pytorch-image-models) and ran the following training script on 2 gpus for GENet normal (lr and batch_size reduced due to memory limitations):
./distributed_train.sh 2 /Imagenet/ -b 128 --sched cosine --epochs 480 --opt sgd -j 8 --weight-decay 4e-5 --warmup-epochs 5 --aa original-mstd0.5 --lr 0.05 --mixup 0.2 --smoothing 0.1 --reprob 0.2

the accuracy I got was 78.9 while the published checkpoint reached 79.9.
Can you please provide a training script or perhaps point out what is wrong with my training command?

Thanks in advanced

[MingLin-home]

Sorry for the late reply!

This project is obsoleted and the development is merged into our ICCV 2021 work Zen-NAS. The new Github repository is https://github.com/idstcv/ZenNAS.