/R_keras_Unet

Unet implementation in R flavor of Keras

Primary LanguageRMIT LicenseMIT

R_keras_Unet

Unet implementation in R flavor of Keras

Semantic segmentation on aerial imagery

Buildings

Alt text Alt text

Roads

Alt text Alt text

infinite training chip generator

on-the-fly image and mask augmentation

Alt text Alt text Alt text

TODO:

  • Need image loader to remove or sample from chips that have no target in them. They could be out of the study area or just lacking a thing. For rare classes, this could be many/most random tiles. might want to think of better way.

  • somehow split out training image vs val image

    • add stride/padding/offset and train/val split values to config
    • add function to make all possible overlapping tile UL coords
    • split all-possible into train/val based on config value
    • only pull infinite chip from train list of possible UL corner coords
    • has to check to see if all-possble coords exist, but DOES NOT KNOW if you changed dim, need to make it aware
      • ran into this bug before b/c no notice of using old data. add message when ussing old data
    • add option to config to reser the all-possble train/val splits
    • has to be able to also function as it does now where split% == 0 so train/val randomly from same of seperate study area images
    • Alternative to random select of train/val set (best in Kfolds CV) is to have full train/val images or either seperate geographies or strips/sections/blocks of single geography where train is NA where Val has data.

  • add functionality to record dice/jaccard by batch and chip coords to reconstruct error map

    • somehow write out coords form batch and then join to error metric of that batch.
    • did this with a custom callback, but only records training dice coef because validation is only at end of epoch
    • create second validation iterator to get coords of validation after model fit
      • post-model-fit validation loop for mapping moved into own script.
      • Then work on train/val all-possible coords lists to use instead...
      • Need to figure out mapping pixel coords to geographic coords

  • image normalization to dataset mean/variance (in config)

  • Test additional architectures

  • Test Time Augmenting, flipping, mirror, etc... (large area prediction)

  • Make it easy to test on different classes (still all binary class prediction)

Info/Sources