[WIP] 519 Add Radiogenomic GAN #531
Conversation
gagandaroach
force-pushed
the
519-radiogenomic-gan
branch
2 times, most recently
from
c6d9829 to
d920dd3
Compare
gagandaroach
force-pushed
the
519-radiogenomic-gan
branch
from
d920dd3 to
0861b5a
Compare
gagandaroach
force-pushed
the
519-radiogenomic-gan
branch
3 times, most recently
from
77e9786 to
2d926be
Compare
gagandaroach
force-pushed
the
519-radiogenomic-gan
branch
from
99588a6 to
4a3378e
Compare
gagandaroach
force-pushed
the
519-radiogenomic-gan
branch
3 times, most recently
from
2bdfad2 to
8786f02
Compare
gagandaroach
force-pushed
the
519-radiogenomic-gan
branch
2 times, most recently
from
ee384df to
59c7197
Compare
|
Hi @Nic-Ma and @wyli. I re-wrote the RGGAN model with MONAI blocks and wired it up to GanTrainer. Could you review the The pull request is still a WIP and not ready for merge. I first must redo the data input pipeline to exemplify using the TCIA DICOM data and not Ziyue's cleaned niifti files. Edit: Should I refactor the network class to handle non 128^2 imagery? |
* Based off research source
Update model average weights every iteration, load and restore weights around model checkpointing. Sample output: INFO:ignite.engine.engine.GanTrainer:Epoch[2] Complete. Time taken: 00:01:20 INFO:Reparameterization:Loaded average params at epoch: 2 INFO:ignite.engine.engine.GanTrainer:Saved checkpoint at epoch: 2 INFO:Reparameterization:Restored network params at epoch: 2
…e mask, and gene impact map * modularize method with signature params * add background bases to plot * template out argparser, unused currently
* remove cv2 erode call on G background reconstruction loss * remove reparameterization trick * rename loss func variables in run_training * remove d_inferer pointer in gan trainer. disc network is called in d loss func * add option to randomize latents in eval script
* use layer factories * patch update weights init to catch monai block * create rggan file for network
* RGGAN uses MONAI framework where possible * Rename files to prefix RG * Remove unused transforms, ReparameterizationTrick, util func
gagandaroach
force-pushed
the
519-radiogenomic-gan
branch
from
a9ffaaa to
f1032fc
Compare
|
Hi @gagandaroach , What's your progress here? Do you still plan to complete it and ask for review? |
Hi @Nic-Ma 👋 I put this down the last few weeks during the start of the academic year. I plan to complete it and ping for review. Right now I am working on the TCIA dataflow into the trainer. Regarding the CT seg dicoms, the ITK python library is missing a SingleBit datatype. If I can add it to the library, then the pydicom dependency is removed from this example's data preprocessor. Issue Link I will update the Todos in the PR body with the rest of the work I think I need. |
|
Sounds good. |
Parse NCIA data to grab only CT scan information. Prepare RNA sequencing data into vector PR not complete, TODO: load filenames using Dataloader abstraction and updated itk Move into new research folder update monai transforms to segment 3d data into 2d slices
|
/build |
|
Hi @gagandaroach thanks for your efforts but this PR has gotten outdated compared to the current state of the repo. I'm closing it now but if you wanted to come back to this work we can easily reopen, although I think most of this here would make sense in our research contributions repo. |
Fixes #519 Radiogenomic GAN
Description
In the research folder I am creating a Monai implementation and tutorial of DLMED's Radiogenomic GAN as detailed in paper. Network and clinical application details in issue description.
Status
Work in progress
TODO