NIH | National Cancer Institute | NCI Wiki  

Versions Compared

Old Version 1

changes.mady.by.user Unknown User (bearceben)

Saved on

compared with

New Version Current

changes.mady.by.user Unknown User (bearceben)

Saved on

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

1) Learn the basics of pyTorch in order to build sample submission that exercises the GPU.

2) Update form with edits from Ulli and Carolyn. 

Staging environment is up and working

2) Image template has been modified to include sample data for training (needs to be done as we are cutting internet to the participant's dockers)

3) Carolyn and I met to discuss creating first series of videos3) Started working on CodaLab V2 and got Miccai 2020 challenge loaded into it. There are still some issues as during a submission it never registers a pass.

Planned for next month

1) Implement Docker image submission with real inference implemented on GPU remote machineRun test phase and switch competition over to docker submission.

2) Make edits to PR website videos as needed.

3) Try to get issue with V2 sorted out) Help Carolyn with code submission and potentially docker submissions explanations.

4) Learn about integrating Kubernetes with CodaLab for larger scale challenges. Medici will be referenced as the organizing and developing agency. 

5) Integrated Okta into CodaLab for restriction to MGB users.

Comments

Things that went well:

  • Moving data to GPU VM took a day but it's all there and seems to work. I had to create an Azure disk and mount to the VM for this.
  • A full working example of docker submission between VMs is located here: http://miccai-staging-environment.centralus.cloudapp.azure.com/
  • Official secure participant docker is read on dockerhub for Organization "cpmcomps" and repo: "cpmcomps/miccai-2020:training_phase_dl_submission"
  • pyTorch is approachable and I'm getting the hang of it. Just trained my very first model. Essentially running this on the path\rad data is all we need.
  • Met with Eric Carmichael from CKCollab (https://www.linkedin.com/in/ericatckcollab/). He is the author of CodaLab V1 and V2 and runs a 6 person company at https://ckcollab.com/.

Things that could be improved:

  • https with queues
  • Okta integration would be nice and I'm working on that right now as well
  • My current understanding of pyTorch
  • Identifying proper Azure OSes that have GPU enabled software and identify which GPUs run on them.
MeetingDate
Bi-weekly Meeting #1

14 May  

Bi-weekly Meeting #2

28 May  

Milestones

DescriptionDate
Trained my first model in pyTorch

 

Got Miccai 2020 loaded into V2 of CodaLab

21 May  

Task

DescriptionResolutionStatusCreation DateClose Date
Create Docker imagestartedpending

 

--
Getting NIH badgestartedpending

 

--

...