how to pretrain resnet26d using ra2 on imagenet1k

[weikaiyun]

Thanks for your great work! I searched in the timm resnet model file. There are no resnet26d.ra2_in1k, resnet10d.ra2_in1k, resnet14d.ra2_in1k in timm, now I want to pretrain them on my own. How can I get the specific training scripts for this. I would like the pretrained weights having the same property with other weights like resnet34d.ra2_in1k. Are there any references?

[rwightman]

I found some old hparams for related 'ra2' models, put them in a gist ... not sure the yaml files load cleanly in latest timm but probably do.

https://gist.github.com/rwightman/07839a82d0f50e42840168bc43df70b3

[rwightman]

One of those hparams is the resnet26d w/ ECA attention, I didn't actually release those weights because the 'T' version was slightly better and same speed. But probably a good match for training the 26d w/o channel attention.

[rwightman]

The 'ra4' hparams are also interesting, I tried it on 18 and 34 and some v2 variants, but they are a looong grind.

https://gist.github.com/rwightman/f6705cb65c03daeebca8aa129b1b94ad#file-resnet18_34_ra4_e3600-yaml