8000 Add attention maps · Issue #5 · mlcommons/GaNDLF · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Add attention maps #5

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
sarthakpati opened this issue Mar 4, 2021 · 5 comments
Closed

Add attention maps #5

sarthakpati opened this issue Mar 4, 2021 · 5 comments
Assignees
@Karol-G
Labels
enhancement New feature or request

Comments

@sarthakpati
Copy link
Collaborator
sarthakpati commented Mar 4, 2021 •
edited
Loading

Is your feature request related to a problem? Please describe.
Interpretability in DL training is critical and would be very nice to have in GaNDLF.

Describe the solution you'd like
Integrate M3d-Cam into the training/inference process.

Describe alternatives you've considered
Open to other suggestions.

Additional context
Also requested by @gastouna
@sarthakpati sarthakpati added the enhancement New feature or request label Mar 4, 2021
sarthakpati referenced this issue in sarthakpati/GaNDLF Mar 24, 2021
@Geeks-Sid
Merge pull request #5 from sarthakpati/refactor
4a29646
@Karol-G
Copy link
Collaborator
Karol-G commented Apr 1, 2021 •
edited
Loading

Hi,

I successfully integraded medcam into the validation section of the training already.
Currently there are probably some issues with assembling the attention patches with the torchio.inference.GridAggregator, but I am not entirely sure. These issues could also simply originate from the fact that the model is untrained.
Do you have a fully trained 2D segmentation dataset, model and config yaml at hand that you could send me?

Best
Karol

@sarthakpati
Copy link
Collaborator Author

I successfully integraded medcam into the validation section of the training already.

Awesome news!!!

These issues could also simply originate from the fact that the model is untrained.

Could you elaborate? I think the torchio.inference.GridAggregator does not care if a model is completely trained or not, as long as it is fed patch locations [ref].

Do you have a fully trained 2D segmentation dataset, model and config yaml at hand that you could send me?

I do not have any such example on hand, perhaps @Geeks-Sid does?

@Karol-G
Copy link
Collaborator
Karol-G commented Apr 1, 2021

Here is an example of the attention superimposed over the original image input:

attention_map

The model is of course not able to learn from just a few images what it should pay attention to, so it is more or less random. However, you can clearly see borders and different normalizations. My hope is that this will go away when the model is trained. I integrated medcam into the nnUNet for example as well which is also patch-based and had no such problems (but I only tested on the finished models). But it could also be the case that there is some problem with normalization or interpolation.

@sarthakpati
Copy link
Collaborator Author

Ah okay. That actually looks great! I'll ping @Geeks-Sid if he has a fully trained model.

@Karol-G Karol-G mentioned this issue May 16, 2021
Merged
8 tasks
@sarthakpati
Copy link
Collaborator Author

Fixed by #51

orhunguley pushed a commit to orhunguley/GaNDLF that referenced this issue Oct 21, 2021
@sarthakpati
Merge pull request mlcommons#5 from orhunguley/upstream-master
525658d 
Upstream master
sarthakpati pushed a commit that referenced this issue Mar 8, 2022
@Geeks-Sid
Merge pull request #5 from sarthakpati/sid_anonymizer
037acdf 
Updates for anonymizer
Geeks-Sid pushed a commit that referenced this issue Mar 21, 2022
Merge pull request #5 from sarthakpati/openvino_integration
68af85c 
Sync with upstream
sarthakpati pushed a commit that referenced this issue Mar 26, 2024
@scap3yvt
Merge pull request #5 from scap3yvt/828-feature-add-the-ability-to-sp…
8cfdfed 
…lit-csvs-for-trainingvalidationtesting-as-a-separate-script

828 feature add the ability to split csvs for trainingvalidationtesting as a separate script
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@Karol-G Karol-G

Labels
enhancement New feature or request
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
2 participants
@Karol-G @sarthakpati
0