optimizer:
  name: adam
  learning_rate: 0.001

scheduler:
  name: cosine_annealing
  t_max: 100

labels:
  sigma: 3

trainer:
  max_epochs: 10
  check_val_every_n_epoch: 5

train_loader:
  batch_size: 8
  shuffle: true

  clipping_strategy:
    name: random_subclip
    duration: 0.256

  augmentations:
    enabled: true
    audio:
      - name: mix_audio
        probability: 0.2
        min_weight: 0.3
        max_weight: 0.7
      - name: add_echo
        probability: 0.2
        max_delay: 0.005
        min_weight: 0.0
        max_weight: 1.0
    spectrogram:
      - name: scale_volume
        probability: 0.2
        min_scaling: 0.0
        max_scaling: 2.0
      - name: warp
        probability: 0.2
        delta: 0.04
      - name: mask_time
        probability: 0.2
        max_perc: 0.05
        max_masks: 3
      - name: mask_freq
        probability: 0.2
        max_perc: 0.10
        max_masks: 3

val_loader:
  clipping_strategy:
    name: whole_audio_padded
    chunk_size: 0.256

loss:
  detection:
    weight: 1.0
    focal:
      beta: 4
      alpha: 2
  classification:
    weight: 2.0
    focal:
      beta: 4
      alpha: 2
  size:
    weight: 0.1

validation:
  tasks:
    - name: sound_event_detection
      metrics:
        - name: average_precision
    - name: sound_event_classification
      metrics:
        - name: average_precision