tjklein/deepnat

Hi 👋, I'm Tassilo

Principial Research Scientist at SAP AI CTO Office

• I’m currently working on:

  • Representation learning in NLP with minimal supervision

    • Paper at ACL 2023 on self-supervised few-shot sentence representation learning.
      
    • Paper at ACL 2022 on self-supervised sentence representation learning.
      

  • Learning with minimal supervision (few-shot, self-supervised)

  • LLM alignment

  • Multi-modal LLMs (advanced reasoning with LLMs, knowledge-grounded LLMs)

• All of my projects are available at https://tjklein.github.io

Connect with me:

Languages and Tools:

 

deepnat's Issues

The deployment model

Conversion to other platform

Hi,
I am curious to see if I can make the code work for my research.
However I had problems installing Caffe.
Do you have the code in Keras, Tensorflow or anything? Or can you convert the .caffemodel files?

Thanks.
Kirsten

Any suggestions for using the pretrained weights on other platforms

I am currently working on KERAS but failed to convert your model on KERAS. I can make the model myself but I need the trained weights. I tried different github projects for converting it though no success. It's been weeks, but couldn't get a path to follow.

Any suggestions?

About dataset

In your paper, you said

We evaluate the segmentation on the dataset of the MICCAI Multi-Atlas Labeling challenge1 (Landman and Warfield, 2012), which consists of T1-weighted MRI scans from 30 subjects of OASIS (Marcus et al., 2007). Manual segmentations were provided by Neuromorphometrics, Inc.2 under academic subscription.

I accessed the Neuromorphometrics website and it required to purchase about 1.500usd for academic subscription. It is very huge money for student. Do you really pay the money? Does it include raw image or just label image only? I also access the first link (http://masiweb.vuse.vanderbilt.edu/workshop2012/index.php?title=Main_Page&action=history) but it has no raw image

Thanks

Laplacian coordinates

Hello! I found the article really interesting, and I'm trying to replicate the algorithm. In particular I was wondering about the laplacian eigenfunctions coordinates system you used. When building the adjacency matrix, is it enough to consider the first 6 nearest neighbours of each voxel?

Skull-Stripping Images

I quote your paper:
"We skull strip the images to focus the prediction on the brain mask; a brain scan from which the skull and other non-brain tissue like dura and eyes are removed."
@TJKlein Can you please help me figure out how did you do the skull stripping? I have downloaded the "MICCAI2012_CompleteRelease_v2" dataset but couldn't find skull stripped images.