s3-new-train.toml

# RoboSat.pink Configuration


[dataset]
  # The slippy map dataset's base directory.
  aws_profile = "esip"
  image_bucket = "planet-snowcover-imagery"
  # regex defines each slippy-map directory, for buckets with many
  imagery_directory_regex = '2018042\d_.*_tiled'
  mask_bucket = "planet-snowcover-snow/ASO_3M_SD_USCAJW_20180423"
  mask_directory_regex = "ASO_3M_SD_USCAJW_20180423-MASK_02-COG$" # $ = end of string = only dirs (no .tif)
  train_percent = 0.7

[[classes]]
  name = "snow"
  color = "white"

# Indicate which dataset sub-directory and bands to take as input.
# You could so, add several channels blocks to compose your input Tensor. Orders are meaningful.
[[channels]]
  bands = [1, 2, 3, 4]
  mean  = [0.485, 0.456, 0.406, 1.0]
  std   = [0.229, 0.224, 0.225, 1.0]


[model]
  # Model name.
  name = "albunet"

  # Encoder model name.
  encoder = "resnet50"

  # Use, or not, ImageNet weights pretraining.
  pretrained = true

  # Loss function name.
  loss = "lovasz"

  # Batch size for training.
  batch_size = 7

  # tile side size in pixels.
  tile_size = 512

  # Total number of epochs to train for.
  epochs = 10

  # Learning rate for the optimizer.
  lr = 0.000025

  # Data augmentation, Flip or Rotate probability.
  data_augmentation = 0.75

  # Weight decay l2 penalty for the optimizer.
  decay = 0.0