config_template.yaml

global:
  savedir: /space/partner/nrcan/geobase/work/transfer/work/deep_learning/lidar/CMM_2018/convpoint_tests/results
  rootdir: /space/partner/nrcan/geobase/work/transfer/work/deep_learning/lidar/CMM_2018/convpoint_tests/prepared
  mlruns_dir: /space/partner/nrcan/geobase/work/transfer/work/deep_learning/mlflow/mlruns
  exp_name: ConvPoint     # MLFlow experiment name

training:
  model: SegBig           # Currently the only model available.
  mode: 2                 # Class mode. See class_mode function for more available options.
  nepochs: 50
  drop: 0
  lr: 1e-3
  gamma: 0.1
  step_size: 10
  batchsize: 10
  npoints: 8168           # Number of points to be sampled in the block
  blocksize: 25           # Size of the infinite vertical column, to be processed
  tolerance: [5, 25]      # Tolerance range (in %) of the difference between number of points expected (npoints) and total in block size. Outer tolerance, a new block size is calculated.
  trn_iter: 500           # Number of iterations during training.
  val_iter: 200           # Number of iterations during validation.
  num_workers: 8
  features: xyzni         # Features to process. xyzni means xyz + number of returns + intensity.
  local_features: True    # Bool to use or not the local features of local density and bloc size. They are computed for every bloc.
  finetune:               # If provided, path to a state dict to be used as pretraining weights.

test:
  test: True
  test_step: 5
  test_labels: True       # Labels available for test dataset
  test_model:             # If provided, path to a folder containing a state_dict.pth, to run test only
  test_folder:            # If provided, path to a folder containing one or more .las files to test.
  savepts: False          # Bool to write predictions or not.