creatis-myriad / ascent Goto Github PK

Dear author, thanks for your excellent work! When I run ascent_preprocess_and_plan dataset=CAMUS_challenge, an error occurs below. The cropped data folder can be produced, but the data folder data_and_propertites under preprocessed data folder is empty. I use your code camus.py to generate the raw data folder nnUNet format dataset. Thanks for your help previously.

I have trained the model with processed ASCENT dataset, with the command ascent_train experiment=camus_challenge_2d logger=tensorboard.

However when I try to perform evaluation and run inference, there is a problem of size mismatch.
size mismatch for net.upsamples.3.transp_conv.weight: copying a param with shape torch.Size([480, 480, 2, 2]) from checkpoint, the shape in current model is torch.Size([480, 256, 2, 2]). size mismatch for net.upsamples.3.conv_block.conv1.conv.weight: copying a param with shape torch.Size([480, 960, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 512, 3, 3]). size mismatch for net.upsamples.3.conv_block.conv1.conv.bias: copying a param with shape torch.Size([480]) from checkpoint, the shape in current model is torch.Size([256]). size mismatch for net.upsamples.3.conv_block.conv2.conv.weight: copying a param with shape torch.Size([480, 480, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 256, 3, 3]). size mismatch for net.upsamples.3.conv_block.conv2.conv.bias: copying a param with shape torch.Size([480]) from checkpoint, the shape in current model is torch.Size([256]). size mismatch for net.upsamples.4.transp_conv.weight: copying a param with shape torch.Size([480, 256, 2, 2]) from checkpoint, the shape in current model is torch.Size([256, 128, 2, 2]). size mismatch for net.upsamples.4.conv_block.conv1.conv.weight: copying a param with shape torch.Size([256, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([128, 256, 3, 3]). size mismatch for net.upsamples.4.conv_block.conv1.conv.bias: copying a param with shape torch.Size([256]) from checkpoint, the shape in current model is torch.Size([128]). size mismatch for net.upsamples.6.conv_block.conv2.conv.weight: copying a param with shape torch.Size([64, 64, 3, 3]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).

I have already specify in my run inference command and predict.yaml that the model is camus_challenge_2d. I wonder how to make the shape of the current model to be 64,128,256,480?

Currently, the data augmentations are hard-coded in nnunet_datamodule. It is better to move them to a hydra config file to provide more flexibility.

Refers #265 and here.

Thank you so much for sharing the code! I am trying to run the model on my own data but encountered some problems.
I have the .nii.gz file shown below, that was transformed from a png file, cropped, rotated, and has the shape of (670, 490, 1).

Patient_Frame028_0000.nii.gz
Original png image: Patient_Frame028.png

When I feed it into the camus_challenge_2d model to run inference, there is problem in resampling the image. My image's is tagged as anisotropic so separate z resampling is used.
The shape[axis] = 1 and new_shape[axis] = 0, so the the line if not shape[axis] == new_shape[axis]: is true and resize is run. However, after running this resize function on line 107 in preprocessing.py, the resized image array becomes completely empty.
May I know how can I further process the data before feeding it into the model?

Hey, thanks for sharing this repository. I was wondering: how do I reproduce your results from the leaderboard of CAMUS?

Which commands do I have to run? Is it simply:

ascent_train experiment=camus_2d logger=tensorboard

as outlined in the readme file?

EDIT: some afterthoughts

Do I leave the image dimensions, and pixel aspect ratios as is, when converting the CAMUS data to the desired format?

Also, is this implementation any different from the original nnUnet? As in, can I expect different results if I train on the same data with the original nnUnet implementation?

How can I specify a specific GPU among multiple GPUs to train the model? There is no relevant description in the documentation. Looking forward to your reply!

Printings done by Joblib.Parallel processes are not in order. Some output strings are stuck together. The printing is even messier on Windows devices.