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