swin_unetr_btcv_segmentation

[katy070222]

hi, I want to change this model and let it can work in 1 channel input & 2 channel output.(background&spleen).
That is,change swin_unetr_btcv_segmentation to work like spleen_ct_segmentation.

I change it like below.But,it didn't work. it would stop and keep"2023-02-03 04:05:35,605 - ignite.engine.engine.SupervisedTrainer - INFO - Engine run resuming from iteration 0, epoch 0 until 2 epochs",and have't any Error message.

I check the place which it jump in and don't jump out. The place is in "/home/u2562867/.local/lib/python3.8/site-packages/ignite/engine/engine.py",1064 line =" self.state.output = self._process_function(self, self.state.batch)",which is also in "/home/u2562867/.local/lib/python3.8/site-packages/monai/engines/trainer.py",226 line"engine.scaler.scale(engine.state.output[Keys.LOSS]).backward()"

Here is my json code.
https://gist.github.com/katy070222/20a694f66c2676e003dc5e30308cf34f

Please help me.

[KumoLiu]

Hi @katy070222, if you want to change the output channel, you could do something like this:

import torch
from monai.networks.nets import SwinUNETR

model = SwinUNETR(
    img_size=(96, 96, 96),
    in_channels=1,
    out_channels=2,
    feature_size=48,
    use_checkpoint=True,
)

img = torch.rand(1, 1,96,96,96)
out = model(img)

Hope it can help you, thanks!

[KumoLiu]

Hi @katy070222,
I can't reproduce your error. I modified the JSON file in the model zoo repo which make it can be used to run spleen data. Attach the JSON file for your reference. train.txt

Hope it can help you, thanks!

[KumoLiu]

Hi @katy070222, from the error message, I think your path is not set correctly. There are some possible reasons:

• You aren't under the correct path
• You didn't download the model

However, if you retrain the model, it doesn't need the weight.

You can download the model weight by using python -m monai.bundle download --name swin_unetr_btcv_segmentation --version xxx --bundle_dir .
Thanks!

[katy070222]

Hi ,@KumoLiu ,thank you for your reply.

yes. I just know this fault. Sorry,I would quickly modify it. I'm so sorry.

And,I got this wrong.

Because my input spatial_shape is "480, 640, 163" ,what should I do?

I have used "SpatialPadd' before the origin "random_transforms" ,and it work.
But I want to asking is it ok for my training?

[KumoLiu]

Hi @katy070222, if your input spatial shape is (480, 640, 163), it won't raise The size of the proposed random crop ROI is larger than the image size. this error, since your crop shape (96, 96, 96) is smaller than image shape.
Maybe you should check this Spacingd transform to set the proper pixdim.
Additionally, you can also set allow_smaller=True in RandCropByPosNegLabel, so that any smaller dimensions will be set to match the cropped size.
https://github.com/Project-MONAI/MONAI/blob/a77de4adff49037ffb24ea2cae217fac40302181/monai/transforms/croppad/array.py#L1002

Thanks!

[katy070222]

Hi ,@KumoLiu ,thank you for your reply.

I tried to do what you write,and it work!

Thank you very much for helping me a lot!
Without you ,maybe I would be trapped by those fault.

Thanks!