neptune.yaml

project: shared/showroom

name: mapping_challenge_baseline
tags: [solution_5]


parameters:
# Data Paths
  data_dir: data/raw
  meta_dir: data/meta
  masks_overlayed_prefix: masks_overlayed
  experiment_dir: data/experiments/mapping_challenge_baseline

# Execution
  overwrite: 0
  num_workers: 4
  num_threads: 1000
  load_in_memory: 0
  pin_memory: 1
  evaluation_data_sample: 1000
  border_width: 0
  small_annotations_size: 14
  loader_mode: resize
  stream_mode: 0

# General parameters
  image_h: 256
  image_w: 256
  image_channels: 3

# U-Net parameters (multi-output)
  nr_unet_outputs: 1
  channels_per_output: 2
  n_filters: 16
  conv_kernel: 3
  pool_kernel: 3
  pool_stride: 2
  repeat_blocks: 4
  encoder: ResNet101

# U-Net loss weights (multi-output)
  bce_mask: 1.0
  dice_mask: 0.2

# Training schedule
  epochs_nr: 100
  batch_size_train: 20
  batch_size_inference: 20
  lr: 0.0005
  momentum: 0.9
  gamma: 1.0
  patience: 30
  lr_factor: 0.3
  lr_patience: 30
  w0: 50
  sigma: 10
  dice_smooth: 1
  dice_activation: 'softmax'
  validate_with_map: 1

# Regularization
  use_batch_norm: 1
  l2_reg_conv: 0.0001
  l2_reg_dense: 0.0
  dropout_conv: 0.1
  dropout_dense: 0.0

# Postprocessing
  erode_selem_size: 0
  dilate_selem_size: 0
  tta_aggregation_method: gmean
  nms__iou_threshold: 0.5

# Inference padding
  crop_image_h: 300
  crop_image_w: 300
  h_pad: 10
  w_pad: 10
  pad_method: 'replicate'

#Neptune monitor
  unet_outputs_to_plot: '["multichannel_map",]'

#Scoring model
  scoring_model: 'lgbm'
  scoring_model__num_training_examples: 10000

#LightGBM
  lgbm__learning_rate: 0.01
  lgbm__num_leaves: 500
  lgbm__min_data: 100
  lgbm__max_depth: 20
  lgbm__number_of_trees: 50000
  lgbm__early_stopping: 10
  lgbm__train_size: 0.7
  lgbm__target: 'iou'

#Random Forest
  rf__n_estimators: 500
  rf__criterion: "mse"
  rf__max_depth: 20
  rf__min_samples_split: 100
  rf__min_samples_leaf: 100
  rf__max_features: 'auto'
  rf__max_leaf_nodes: 500
  rf__n_jobs: 10
  rf__verbose: 0