config.yaml

dataloader:
  num_workers: 0
  pin_memory: true
dataset:
  cache_dir: .cache/classifier_data
  hf_dataset: maximuspowers/muat-mean-std
  input_mode: signature
  max_dimensions:
    max_layers: 13
    max_neurons_per_layer: 8
    max_sequence_length: 5
  neuron_profile:
    methods:
      mean: {}
      std: {}
  patterns:
  - palindrome
  - sorted_ascending
  - sorted_descending
  - alternating
  - contains_abc
  - starts_with
  - ends_with
  - no_repeats
  - has_majority
  - increasing_pairs
  - decreasing_pairs
  - vowel_consonant
  - first_last_match
  - mountain_pattern
  random_seed: 42
  test_split: 0.1
  train_split: 0.8
  val_split: 0.1
device:
  type: auto
evaluation:
  decision_threshold: 0.5
  metrics:
  - accuracy_exact_match
  - accuracy_hamming
  - precision_macro
  - recall_macro
  - f1_macro
  - f1_micro
  per_pattern_metrics: true
hub:
  enabled: true
  private: false
  push_frequency: epoch
  push_logs: true
  push_metrics: true
  push_model: true
  repo_id: maximuspowers/muat-mean-std-classifier
  token: <REDACTED>
logging:
  checkpoint:
    enabled: true
    mode: max
    monitor: val_f1_macro
    save_best_only: true
    save_dir: ./checkpoints/classifier_mean_std
  tensorboard:
    enabled: true
    log_dir: ./runs/classifier_mean_std
    log_interval: 10
  verbose: true
model:
  fusion:
    activation: relu
    dropout: 0.2
    hidden_dims:
    - 128
    - 64
  output:
    num_patterns: 14
  signature_encoder:
    activation: relu
    dropout: 0.2
    hidden_dims:
    - 512
    - 256
    - 256
    - 128
  use_batch_norm: true
  weight_encoder:
    activation: relu
    dropout: 0.2
training:
  batch_size: 16
  early_stopping:
    enabled: true
    mode: min
    monitor: val_loss
    patience: 50
  epochs: 1000
  learning_rate: 0.001
  loss: bce_with_logits
  lr_scheduler:
    enabled: true
    factor: 0.5
    min_lr: 1.0e-05
    patience: 20
    type: reduce_on_plateau
  optimizer: adam
  pos_weight: null
  weight_decay: 0.0001