dl_clustering.yaml

name: "dl_clustering"

feature_definition:
  use_norm_features: True

data_loader:
  batch_size: 512
  cluster_batch_size: 256
  # Define how each data batch is sampled from multiple datasets and/or people
  # option: "across_dataset", "within_dataset", "across_person", "within_person"
  generate_by: "across_dataset" 
  
  mixup: "across" # option: null, "across", "within"
  mixup_alpha: 0.2

model_params:
  arch: "1dCNN" # model architecture
  # input dimension
  # 2: a multi-channel time series data (for models like 1dCNN)
  # 3: a single-channel image data (for models like 2dCNN)
  input_dim: 2
  conv_shapes: [8,8,8] # a list of convolutional layers shapes for the siamese network
  embedding_size: 16 # vector length of the feature embedding network output
  flag_y_vector: False # False: use sparse label for negative mining during contrastive learning
  triplet_loss_margin: 0.2 # triplet loss margin

  clustering: # parameter for the unsupervised clustering model
    conv_shapes: [64, 32]
    embedding_size: 10
    n_clusters: 60
    loss_weight: 0.1
    tdistribution_shape_param: 1
    error_tolerance: 0.001

training_params:
  optimizer: "Adam" # option: SGD or Adam
  learning_rate: 0.001
  epochs: 1000
  steps_per_epoch: 10 # number of batches per epoch
  cos_annealing_step: 100
  cos_annealing_decay: 0.95
  best_epoch_strategy: "direct" # option: direct or on_test
  verbose: 0
  clustering_verbose: 2
  skip_training: False
  # only if train models for particular clusters, add the clusters idx.
  # This can be used on the cluster server (\eg SLURM-based) to accelerate the training
  focus_cluster_idx: [] 

# Parameters used based on the embeddings directions
metrics_params:
  n_neighbor: 3 # how many top close/far datapoints used for classification
  positive_weight_ratio: 1.5
  direction: "close" # option: both, close, far. Which side of datapoints to be used for classification
  flag_distance_weight: False # whether use distance as a weight (closer/further datapoints have higher weights)