Using swin transformer with arbitrary sizes

[shengyi4]

Hi,

My customer dataset has arbitrary sizes from 500700-46083456(train and test). I used default upernet_swin_base_patch4_window7_512x512_160k_ade20k_pretrain_224x224_22K.py to train and test my data, but I only got 1.56 mIoU on it. I wonder if there is any set up I should config for my datasets?

Thanks！

[xiexinch]

Hi @shengyi4,
I think you could try to crop the picture yourself with roughly the same size or check to see if there is a problem with the training configuration.

[shengyi4]

Thanks for your reply! I really appreciate it. I just followed upernet_swin_tiny_patch4_window7_512x512_160k_ade20k_pretrain_224x224_1K.py without changing the training part, (just changed dataset, number of class). Will it cause any problems? Thanks! On Aug 31, 2022, at 3:22 AM, 谢昕辰 ***@***.***> wrote:  Hi @shengyi4<https://urldefense.com/v3/__https://github.com/shengyi4__;!!DZ3fjg!759vLkve2mvMZviLwVSb2Dwzt3UjuiXxOQE1QIdYLncricSvHHs04ZtU3q6z--IUKAbLB66dhsuv_KzA9HP_z4ChMTkH$>, I think you could try to crop the picture yourself with roughly the same size or check to see if there is a problem with the training configuration. — Reply to this email directly, view it on GitHub<https://urldefense.com/v3/__https://github.com/open-mmlab/mmsegmentation/issues/1979*issuecomment-1232622321__;Iw!!DZ3fjg!759vLkve2mvMZviLwVSb2Dwzt3UjuiXxOQE1QIdYLncricSvHHs04ZtU3q6z--IUKAbLB66dhsuv_KzA9HP_z1u6OjxT$>, or unsubscribe<https://urldefense.com/v3/__https://github.com/notifications/unsubscribe-auth/ALIDGYXJZYJY555R4VEN3FDV34I5PANCNFSM572LPH2A__;!!DZ3fjg!759vLkve2mvMZviLwVSb2Dwzt3UjuiXxOQE1QIdYLncricSvHHs04ZtU3q6z--IUKAbLB66dhsuv_KzA9HP_z2LnhbQu$>. You are receiving this because you were mentioned.Message ID: ***@***.***>