yaq007/autofocus-layer

Autofocus Layer for Semantic Segmentation

autofocus-layer's Issues

question about the autofocus implementation

In https://github.com/yaq007/Autofocus-Layer/blob/master/models.py Line 180 why do you use x.detach?

this paper was accepted in MICCAI 2018, it's worng about citation in your README.

Questions about the setting

Thank you for your sharing.

I have a question about the model setting. I want to train another new model for BRATS2015 dataset. Are the default setting in the code as same as that reported in your MICCAI2018 paper? For example, when I use your default setting, I find the learning rate decays since the first epoch and the batch size is 10.

Look forward to your reply. Thank you.

is it possible to working with CPU ?

Hello
is it possible to working with CPU?

Thanks

problem

Hello, may I ask if this Autofocus Layer can directly replace Dilated convolution? How is the effect, thank you

Questions about the code

Hi @yaq007 , thanks for sharing your work. I have a couple questions about the code: in Dataset.py line: 35~36, each sample contains 4 images and just 1 label. I'm wondering why is that? Also, according to line 167, it looks like that only one patch was cropped for each image (not sure if that would be sufficient for other datasets). When I tried the model on my own dataset, I found that the loss was decreasing pretty slowly and even stops decreasing after certain epochs. Tuning the learning rate didn't help. I'm wondering where I might did wrong.

Some infos:

1. I've tried both the normalized dataset and unnormalized version (if normalized, vols = (vols - vols.mean())/vols.std()). Is there a specific reason that each BRATS image has to be normalized individually?
2. I did not provide masks for the images and have removed the codes about the masks. Not sure how this will affect the performance.

Any help is greatly appreciated!!

Data problem

After downloading the BRATS2015 data, HGG and LGG will go through the data pre-processing through the program you provided. What is the difference between HGG and LGG, is there a data tag in HGG that is one more than the data tag in HGG? But the sample data that you gave through the pre-processing, I looked exactly the same. This is my question. Please answer your questions.

Code for pre-processing?

Dear Yao,
Thank you for your contribution.
Can you please share the code for the pre-processing used to create the small BRATS 2015 subset that you shared?

Thanks,

DATa loader

test_full.py working but results in exception

Running python test_full.py works, but gives the following exception as score and pred_seg are not defined in the test() method:

Traceback (most recent call last):
File "test_full.py", line 228, in
main(args)
File "test_full.py", line 139, in main
test(test_loader, model, args)
File "test_full.py", line 112, in test
return score, pred_seg
NameError: name 'score' is not defined