Volumetric probability map based PatchSampler

[surajpaib]

Discussed with @cnmy-ro where he spoke about his volumetric probability map based sampling. This takes in a probability map the same shape as the volume (if provided) with sampling probabilities for each voxel.

I think this is a very useful feature when you don't want to sample randomly from all voxels.
Usecases

1. Sampling around specific tumors/OARs
2. Sampling within masks
3. Sampling for modalities like PET

[ibro45]

That's cool! How do you provide the probability map when running an experiment?

[ibro45]

PatchSampler same, stochastic focal, random, probabilty

[ibro45]

let's rename the stochastic focal to poisition based patch sampler

[ibro45]

once this is done, i'll implement a Progressively growing patch sampler, that increases the patch size that you train with at the specified checkpoints. this will help with faster training, relevant only for 3D training imo

[surajpaib]

Count me in to help! I think this is a very important thing to have.

…

On Wed, Sep 15, 2021 at 1:41 PM Ibrahim Hadzic ***@***.***> wrote: once this is done, i'll implement a Progressively growing patch sampler, that increases the patch size that you train with at the specified checkpoints. this will help with faster training, relevant only for 3D training imo — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <https://mailtrack.io/trace/link/e60e3f18665e648f8740cee0198300f81252aab2?url=https%3A%2F%2Fgithub.com%2Fganslate-team%2Fganslate%2Fissues%2F89%23issuecomment-919943767&userId=1781080&signature=cf5fcc08bfd2e5b8>, or unsubscribe <https://mailtrack.io/trace/link/f55aeb13b5f6e0deb58b2fd434023f40eb88bcb5?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FACP3TXGTYITV3HCO45HYVN3UCCA73ANCNFSM44FGFI5Q&userId=1781080&signature=a0dba0482346d64f> .

-- Regards, *Suraj Pai* Contact no:+918105050486 Email id: ***@***.***