search

dmklee / equivision

ImageNet1k-pretrained SE(2) Equivariant Vision Models
MIT License
8 stars 0 forks source link

ResNet18 d8 #1

Open ph-code-repo opened 5 months ago

ph-code-repo commented 5 months ago

Hi there, Would it be possible for you to train a d8 ResNet18 network for me. I'd really like to be able to have the same number of rotations as the C8 network, but within the dihedral group. I doubt that I have the resources to be able to train it myself.

Kind Regards!

dmklee commented 5 months ago

Yes, I'd be happy to train that for you. I can create a d8resnet using a base_width=20 which comes out to around the same number of trainable parameters as the other models. Here are some stats for it:

Num Params: 11.9M Eval forward pass (batchsize=1): 25ms | 3.4 Gb Train forward pass (batchsize=32): 230ms | 8.4 Gb

I got these numbers with an RTX2080Ti. Does this sound like it would work for your situation? It will take a few days to train so I want to confirm it's right for your application.

ph-code-repo commented 5 months ago

That is great news, thank you! Just to check, is the base_width parameter the number of channels/kernels used in the first layer of ResNet, so that the number in the final layer of ResNet 18 would be 8 times this amount? In which case, I think this should be fine. Thank you for responding so quickly!

dmklee commented 4 months ago

It's training now, sorry for the delay. I'll upload the checkpoint by the end of the week.