gromdimon/brain-segment

Design of NN model for brain 3D images segmentation.

Hi there 👋

I am a hardworking peacefull human being. Building Software 👾

For more information about me - visit gromdimon.github.io

brain-segment's Issues

Create doc "github general workflow"

Is your feature request related to a problem? Please describe.
It'll be useful to have a doc, which specifies how to work with GH to make more fun, then pain.

Describe the solution you'd like
Create this doc

Describe alternatives you've considered
N/A

Additional context
N/A

Investigate `UNETR` NN Architecture

Is your feature request related to a problem? Please describe.
We should investigate UNETR Architecture as described in one of our doc.

Describe the solution you'd like
Investigate UNETR architecture:

• Investigate the architecure
• Create example notebook (Look into MONAI tutorials). Play aroung this notebook
• INvestigate possible oprimizations for the model
• Try to train model on some data subset and receive accuracy results
• Create results doc and summarize all the found results

Describe alternatives you've considered
N/A

Additional context
We first need a consisten data loader, so @gromdimon should create it soon.

Explore the optimal number of epochs

Check docs/optimizations.md

Add `W&B` tracking

Is your feature request related to a problem? Please describe.
We should use weights and biases for tracking results of model training.

Describe the solution you'd like

• Investigate Weights and Biases
• Add setup function in src/utils.py for setting up W&B tracking in jupyter notebook
• Define separate config file for all hyperparameters
• Research + include all necessary parameters to track (in the created config file)
• Setup tracking in notebooks/UNet.ipynb with help of the created helper function

Describe alternatives you've considered
N/A

Additional context
Useful links:
Example setup in python: https://docs.wandb.ai/quickstart
Tutorial: https://markjcarlebach.medium.com/getting-started-with-weights-and-biases-19450c33a4ed

Document UNETR characteristics

Is your feature request related to a problem? Please describe.
Add UNETR characteristics in related doc.

Describe the solution you'd like
N/A

Describe alternatives you've considered
N/A

Additional context
N/A

Create MVN for U-Net

Is your feature request related to a problem? Please describe.
Create a notebook for training UNet

Describe the solution you'd like
N/A

Describe alternatives you've considered
N/A

Additional context
N/A

Investigate `U-Net` architecture

Is your feature request related to a problem? Please describe.
We should investigate U-Net Architecture as described in one of our doc.

Describe the solution you'd like
Investigate U-Net architecture:

• Investigate the architecure
• Create example notebook (Look into MONAI tutorials). Play aroung this notebook
• INvestigate possible oprimizations for the model
• Try to train model on some data subset and receive accuracy results
• Create results doc and summarize all the found results

Describe alternatives you've considered
N/A

Additional context
We first need a consisten data loader, so @gromdimon should create it soon.

Adjust the model architecture

Check docs/optimizations.md

Explore regularization techniques

Check docs/optimizations.md

Explore activation functions

Check docs/optimizations.md

environment.yml does not work with Conda

Describe the bug
environment.yml is not functional as of right now on SSH

To Reproduce
Steps to reproduce the behavior:

1. Log into SSH
2. cd into brain-segm
3. Create a new Conda environment based on environment.yml

Expected behavior
The right packages are downloaded and the new Conda environment is created.

Actual behavior
Either it will report the following packages as not found:

ResolvePackageNotFound: 
  - torchaudio
  - ignite

Or, if we exclude the invalid packages, it will remain on

Solving environment: |

Until terminated.

Make `Readme` greate again!

Is your feature request related to a problem? Please describe.
Sometimes it's not just about the beauty of code, but a general beauty. So, we should make our Readme attractive enough.

Describe the solution you'd like
Update the Readme. Inspiration can be found e.g. in

• https://github.com/gromdimon/beLLM
• https://github.com/gromdimon/albi-prak
• https://github.com/gromdimon/LongCovid

Describe alternatives you've considered
Create a documentation and deploy it separately. But that's more difficult and not needed

Additional context
N/A

Create unified data loader

Is your feature request related to a problem? Please describe.
As adressed in issues #3, #4, #5 we need a consistent unified data loader to make the results comparable.

Describe the solution you'd like
Create a python file for data loading + data splitting.

Describe alternatives you've considered
N/A

Additional context
Look at MONAI tutorials

Investigate `ResNet` architecture

Is your feature request related to a problem? Please describe.
We should investigate ResNet Architecture as described in one of our doc.

Describe the solution you'd like
Investigate ResNet architecture:

• Investigate the architecure
• Create example notebook (Look into MONAI tutorials). Play aroung this notebook
• INvestigate possible oprimizations for the model
• Try to train model on some data subset and receive accuracy results
• Create results doc and summarize all the found results

Describe alternatives you've considered
N/A

Additional context
We first need a consisten data loader, so @gromdimon should create it soon.

Explore optimizers and learning rate schedulers

Check docs/optimizations.md

Explore loss functions

Check docs/optimizations.md

Explore data augmentation

Check docs/optimizations.md