/pgrad-thesis

Repository for organising content for my MSc thesis

Primary LanguageJupyter Notebook

pgrad-thesis

Impromptu List of Things to Re-create Environment

conda create -n pgrad-thesis -c conda-forge -c rapidsai-nightly python=3.8 cudatoolkit=11.2 cucim tensorflow-gpu pandas matplotlib

(for rasterio) conda config --add channels conda-forge conda config --set channel_priority strict

conda install -y rasterio tqdm

Updates

Jan-10

  • Original data (Maxar) is blurry and unlabelled, infeasible for MSc without labelling/ground

Experiment Managment/Tracking

General Notes

  • Dice coefficient (and focal loss) used to evaluate performance, NOT accuracy
  • Vanilla segmentation was never the right choice, need some sort of differencing network (need to read)

Yes, I know there are better ways to do this

  • Get a single training loop done with BW images
  • Add validation step to training loop
  • Add SEGNET and RGB data pipeline
  • [ ]Recheck SegNET architecture (not seeing skip connections in image)
  • Use cuCIM.skimage.transform.resize instead of cv2.resize
  • Fix loss function/model output, error where cannot calculate Dice loss, fix model output instead to select only channel 1
  • Store differenced images separately, then using ImageGenerator to reference differenced folder
  • Implement differencing network
  • Run training for Siamese network overnight