task5.yaml

active_set: dev

sets:
  - set_id: dev # development dataset (folds)
    general:
      active_fold_list: [1,2,3,4]
    dataset:
      method: baseline_development
    feature_processing_chain:
      method: DCASE2018_Task5_baseline
    data_processing_chain:
      method: DCASE2018_Task5_baseline
    learner:
      method: DCASE2018_Task5_baseline

  - set_id: train # development dataset (full) to be used on eval set
    dataset:
      method: baseline_challenge_train
    feature_processing_chain:
      method: DCASE2018_Task5_baseline
    data_processing_chain:
      method: DCASE2018_Task5_baseline
    learner:
      method: DCASE2018_Task5_baseline

  - set_id: eval # challenge evaluation set
    dataset:
      method: baseline_challenge_eval # available from DCASE util v0.2.4 (release 30/06/2018)
    feature_processing_chain:
      method: DCASE2018_Task5_baseline
    data_processing_chain:
      method: DCASE2018_Task5_baseline
    learner:
      method: DCASE2018_Task5_baseline

defaults:
  flow:
    feature_extraction: true
    feature_normalization: true
    learning: true
    testing: true
    evaluation: true

  general:
    overwrite: false # Overwrite previously stored data
    eval_mode: true # challenge evaluation mode
    active_fold_list: [1,2,3,4]

  path:
    dataset: dataset/
    log: log/
    application:
      base: output/
      feature_extractor: features
      feature_normalizer: normalization
      learner: learner
      recognizer: recognizer

  dataset:
    method: baseline_development

  dataset_method_parameters:
    baseline_development:
      dataset: DCASE2018_Task5_DevelopmentSet
      evaluation_mode: folds
    baseline_challenge_train:
      dataset: DCASE2018_Task5_DevelopmentSet
      evaluation_mode: full
    baseline_challenge_eval:
      dataset: DCASE2018_Task5_EvaluationSet
      evaluation_mode: full

  feature_extractor:
    recipe: mel
    win_length_seconds: 0.04
    hop_length_seconds: 0.02
    fs: 16000
    channels: 4

  feature_extractor_method_parameters:
    mel:
      spectrogram_type: magnitude
      window_type: hamming_asymmetric
      n_mels: 40
      n_fft: 1024
      fmin: 50
      fmax: 8000
      htk: false
      normalize_mel_bands: false
      logarithmic: true

  feature_processing_chain:
    method: DCASE2018_Task5_baseline

  feature_processing_chain_method_parameters:
    DCASE2018_Task5_baseline:
      chain:
        - processor_name: AudioReadingProcessor # read multi-channel audio
        - processor_name: RepositoryFeatureExtractorProcessor # extract features based on 'feature_extractor'

  data_processing_chain_method_parameters:
    DCASE2018_Task5_baseline:
      chain:
        - processor_name: RepositoryFeatureReadingProcessor # read features into repository
        - processor_name: RepositoryNormalizationProcessor # normalize features
          init_parameters:
            enable: true
        - processor_name: RepositorySequencingProcessor # add dimension
          init_parameters: # no sequencing
            sequence_length: 501
            hop_length: 501
        - processor_name: RepositoryToMatrixProcessor # convert repo to matrix
          init_parameters:
            label: mel
            expanded_dimension: first
        - processor_name: DataShapingProcessor # rearrange dimensions to match keras model input
          init_parameters:
            axis_list:
              - channel_axis
              - sequence_axis
              - data_axis
              - time_axis

  learner_method_parameters:
    DCASE2018_Task5_baseline:
      random_seed: 0
      profile: deterministic
      backend: tensorflow
      validation_amount: 0.3

      model:
        constants:
          CLASS_COUNT: 9

          FEATURE_VECTOR_LENGTH: 40
          INPUT_SEQUENCE_LENGTH: 501

          CONVOLUTION_FILTER: 32
          CONVOLUTION_INIT: he_normal
          CONVOLUTION_BORDER_MODE: valid

          CONVOLUTION_DATA_FORMAT: channels_first

          CONVOLUTION_ACTIVATION: relu
          CONVOLUTION_DROPOUT: 0.2

          DENSE_UNITS: 64
          DENSE_DROPOUT: 0.2

        config:
          # CNN layer 1
          # ====================================
          # Convolution layer
          - class_name: Conv2D
            config:
              input_shape:
                - 1 # channel_axis
                - FEATURE_VECTOR_LENGTH # data_axis
                - INPUT_SEQUENCE_LENGTH # time_axis
              filters: CONVOLUTION_FILTER
              kernel_size:
                - FEATURE_VECTOR_LENGTH
                - 5
              strides:
                - 1
                - 1
              padding: CONVOLUTION_BORDER_MODE
              kernel_initializer: CONVOLUTION_INIT
              data_format: CONVOLUTION_DATA_FORMAT

          # Batch normalization
          - class_name: BatchNormalization
            config:
              axis: 1 # data_axis

          # Detection layer
          - class_name: Activation
            config:
              activation: CONVOLUTION_ACTIVATION

          # Pooling layer
          - class_name: MaxPooling2D
            config:
              pool_size:
                - 1
                - 5
              strides:
                - 1
                - 5
              data_format: CONVOLUTION_DATA_FORMAT

          # Drop out layer
          - class_name: Dropout
            config:
              rate: CONVOLUTION_DROPOUT

          # CNN layer 2
          # ====================================
          # Convolution layer
          - class_name: Conv2D
            config:
              filters: CONVOLUTION_FILTER * 2
              kernel_size:
                - 1
                - 3
              strides:
                - 1
                - 1
              padding: CONVOLUTION_BORDER_MODE
              kernel_initializer: CONVOLUTION_INIT
              data_format: CONVOLUTION_DATA_FORMAT

          # Batch normalization
          - class_name: BatchNormalization
            config:
              axis: 1 # data_axis

          # Detection layer
          - class_name: Activation
            config:
              activation: CONVOLUTION_ACTIVATION

          # Pooling layer
          - class_name: GlobalMaxPooling2D
            config:
              data_format: CONVOLUTION_DATA_FORMAT

          # Drop out layer
          - class_name: Dropout
            config:
              rate: CONVOLUTION_DROPOUT

          # ====================================
          # Fully connected layer
          - class_name: Dense
            config:
              units: DENSE_UNITS
              kernel_initializer: uniform
              activation: relu

          - class_name: Dropout
            config:
              rate: DENSE_DROPOUT

          # ====================================
          # Output layer
          - class_name: Dense
            config:
              units: CLASS_COUNT
              kernel_initializer: uniform
              activation: softmax

      compile:
        loss: sparse_categorical_crossentropy
        metrics:
          - sparse_categorical_accuracy

      optimizer:
        class_name: Adam
        config:
          lr: 0.0001

      fit:
        epochs: 500
        processing_interval: 10
        batch_size: 256
        shuffle: True
        undersampling: uniform

  recognizer:
    fusion:
      method: mean