Self supervised SwinUnetr Pretrained weight usage

[justinzhang528]

Dear MONAI Community,

I have a question about the usage of swinUnetr pretrained weight, if i want to use the pretrained weight for my down stream task, is any constrain about input image such as image size, channel or dimension...

Thank in advance!

[KumoLiu]

Hi @justinzhang528, I think you can find more details in this tutorial about how to use this pre-trained weight.

model = SwinUNETR(
    img_size=(96, 96, 96),
    in_channels=1,
    out_channels=14,
    feature_size=48,
    drop_rate=0.0,
    attn_drop_rate=0.0,
    dropout_path_rate=0.0,
    use_checkpoint=True,
)

Hope it can help you, thanks!

[justinzhang528]

@KumoLiu I will take a look, thanks！

[tangy5]

If you would like to use the pretrained position embedding, you will need to follow the 96x96x96 patch size as input.
If you'd like to load part of the pre-trained model, for example, only the transformer blocks, it should be good to variant image size inputs. But you will need to customize the weights loading logics.

Thank you.

[justinzhang528]

I got it, thanks!