config.yaml

# lightning.pytorch==2.1.1
seed_everything: 0
trainer:
  accelerator: gpu # we can also use auto or cpu
  strategy: auto
  devices: auto
  num_nodes: 1
  logger: True # will use tensorboardlogger

  callbacks:
    - class_path: RichProgressBar
    - class_path: LearningRateMonitor
      init_args:
        logging_interval: epoch
    - class_path: EarlyStopping 
      init_args:
        monitor: val/loss
        patience: 30

  max_epochs: 200
  check_val_every_n_epoch: 1
  log_every_n_steps: 1
  enable_checkpointing: true
  default_root_dir: ./../data/fine_tuning/granite_geospatial_uki_flood_detection_v2
data:
  class_path: GenericNonGeoSegmentationDataModule
  init_args:
    batch_size: 4
    num_workers: 1
    constant_scale: 0.0001
    dataset_bands: # what bands are in your data 
      - VV
      - VH
      - BLUE
      - GREEN
      - RED
      - NIR_NARROW
      - SWIR_1
      - SWIR_2
      - CLOUD
    output_bands: # which bands do you want to fine-tune 
      - BLUE
      - GREEN
      - RED
      - NIR_NARROW
      - SWIR_1
      - SWIR_2
      - VV
      - VH
      - CLOUD
    rgb_indices:
      - 4
      - 3
      - 2
    train_data_root: ./../data/regions/combined_uki_spain/images/
    train_label_data_root: ./../data/regions/combined_uki_spain/labels/
    val_data_root: ./../data/regions/combined_uki_spain/images/
    val_label_data_root: ./../data/regions/combined_uki_spain/labels/
    test_data_root: ./../data/regions/combined_uki_spain/images/
    test_label_data_root: ./../data/regions/combined_uki_spain/labels/
    train_split: ./../data/regions/combined_uki_spain/splits/flood_train_data.txt
    test_split: ./../data/regions/combined_uki_spain/splits/flood_test_data.txt
    val_split: ./../data/regions/combined_uki_spain/splits/flood_val_data.txt
    img_grep: "*_image.tif"
    label_grep: "*_label.tif"
    no_label_replace: -1
    no_data_replace: 0
    means:
      - 0.1290484133335582      # BLUE
      - 0.13423481405157794    # GREEN
      - 0.1328938801112928      # RED
      - 0.20036851044035797     # NIR_NARROW
      - 0.13804629743141042      # SWIR_1
      - 0.10409700513471637      # SWIR_2
      - -0.0018052691820029847   # VV
      - -0.0023712696527645486  # VH
      - 0.000024014472961425782 #CLOUD

    stds:
      - 0.25406999374272976
      - 0.22949378991348005
      - 0.21689414406289836
      - 0.22552362238920548
      - 0.1600542128720416
      - 0.12602917719190815
      - 0.0011294842635096356
      - 0.0008879269711519241
      - 0.00004271712050839232

    num_classes: 2

model:
  class_path: terratorch.tasks.SemanticSegmentationTask 
  init_args:
    model_args:
      decoder: FCNDecoder
      backbone_pretrained: true
      backbone: granite_geospatial_uki
      backbone_pretrained_cfg_overlay:
        file: ./../data/checkpoints/granite_geospatial_uki.pt
      backbone_pretrain_img_size: 512
      decoder_channels: 256
      # in_channels: 9
      backbone_bands:
        - BLUE
        - GREEN
        - RED
        - NIR_NARROW
        - SWIR_1
        - SWIR_2
        - VV
        - VH
        - CLOUD 
      # num_frames: 1
      num_classes: 2
      head_dropout: 0.1
      decoder_num_convs: 4
      head_channel_list:
        - 256
      necks:
        - name: SelectIndices
          indices:
            - -1
        - name: ReshapeTokensToImage
    loss: ce
    aux_heads:
      - name: aux_head
        decoder: FCNDecoder
        decoder_args:
          decoder_channels: 256
          decoder_in_index: -1
          decoder_num_convs: 2
          head_dropout: 0.1
    aux_loss:
      aux_head: 1.0
    ignore_index: -1
    class_weights:
      - 0.3
      - 0.7
    freeze_backbone: false
    freeze_decoder: false
    model_factory: EncoderDecoderFactory
    tiled_inference_parameters:
      h_crop: 512
      h_stride: 496
      w_crop: 512
      w_stride: 496
      average_patches: true
optimizer:
  class_path: torch.optim.AdamW
  init_args:
    lr: 6.e-5
    weight_decay: 0.05
lr_scheduler:
  class_path: ReduceLROnPlateau
  init_args:
    monitor: val/loss