train_config.yml

# configures the tile used
tile_id: '32TNS'

# configures the satellite used
satellite: 'sentinel2' # 'sentinel2' or 'landsat8'

# if you wish to run the pipeline for hand selected dates
# you can specify them here (using yml array syntax)
dates: [ ]
# selected dates (36 dates):
# dates: [ '20210111T102309', '20210116T102351', '20210121T102239', '20210126T102311', '20210131T102149',
#  '20210215T102121', '20210220T101939', '20210225T102021', '20210302T101839', '20210307T102021', '20210322T101649',
#  '20210401T101559', '20210406T102021', '20210416T102021', '20210526T102021', '20210531T101559', '20210610T101559',
#  '20210615T102021', '20210625T102021', '20210705T102031', '20210710T101559', '20210720T101559', '20210730T101559',
#  '20210814T102031', '20210819T101559', '20210824T102031', '20210903T102021', '20210908T101559', '20210913T102021',
#  '20210918T101639', '20211008T101829', '20211013T101951', '20211018T101939', '20211028T102039', '20211102T102201',
#   '20211112T102251' ]

# This section describes which masks should be used to train the model
# masks can be generated automatically based on existing algorithms
# or masks can be loaded from a directory
ground_truth_masks:

  # use the auto annotator to create the ground truth (1 or 0)
  # currently this computes the s2cloudless predictions which is very slow
  # then combines them with the ExoLab predictions and the water mask
  # automatic maks can only be created for S2 data not for L8
  auto_annotation: 0

# This section describes the training and validation dataset creation/loading
dataset:

  # forces a recreation of the dataset (1 or 0)
  # this deletes the dataset folder and recreates it using the current raw data
  # will be ignored if create_on_the_fly is enabled
  recreate_dataset: 0

  # limits the number of dates to be used for the dataset
  # if you wish to use all dates, just set this to 0
  limit_dates: 1


training:

  # enables the training (1 or 0)
  enable: 1

  # continues the training from the checkpoint "unet.pth" (1 or 0)
  continue_training: 0