cnn_experiment.yaml

description: Parameters used to train a CNN for the BFF problem

# -------------------------------------------------------------------------------------------------
# *** Settings for data preprocessing ***
# IMPORTANT - avoid changing these data_parameters (other than the file paths), unless you really
# understand the consequences in the code. Most parameters came from the ray-tracing simulations
# themselves
data_parameters:
    input_file: '../bff_data/final_table'
    preprocessed_file: '../bff_data/preprocessed_dataset.pkl'
    max_time: 6                 #in microseconds
    sample_freq: 20             #in MHz
    beamformings: 32
    original_tx_power: 30       #in dBm
    original_rx_gain: 0         #in dBi
    power_offset: 170
    power_scale: 0.01
    pos_grid: [400.0, 400.0]    #in meters
    pos_shift: [183.0, 176.0]   #in meters
    keep_timeslots: [1, 82]
    run_sanity_checks: True


# -------------------------------------------------------------------------------------------------
# *** Settings for overall experiment control ***
experiment_settings:
    detection_threshold: -100   #in dBm  [depends on the sample frequency, thermal noise]
    tx_power: 45                #in dBm
    rx_gain: 10                 #in dBi
    noise_std: 6                #in dB  [Gaussian noise applied over dB values -> "log-normal"]
    scaler_type: "binarizer"
    model_type: "cnn"
    tests_per_position: 30
    predictions_file: "model_predictions.pkl"


# -------------------------------------------------------------------------------------------------
# *** Settings for model hyperparameters ***
ml_parameters:
    # IO settings
    input_name: "input"
    input_shape: [32, 81, 1]    # BF, usable timeslots, 1 "data channel"
    output_name: "location"
    output_type: "regression"
    output_shape: [2]
    validation_metric: "euclidean_distance"
    model_folder: "../experiments_results/cnn/"

    # basic model hyperparams
    optimizer_type: "ADAM"
    batch_size: 64
    batch_size_inference: 256
    dropout: 0.01
    max_epochs: 1000
    early_stopping: 50          # Stops training after these epochs without improvement on val set
    fc_layers: 12
    fc_neurons: 256
    learning_rate: 0.0001
    learning_rate_decay: 0.995  # Multiplicative term at the end of each epoch
    target_gpu: "0"
    mc_dropout: 0               # Number of samples for MC Dropout-based uncertainty estimation. 0 to deactivate.

    # architecture specific hyperparams
    conv_layers: 1
    conv_filters: [8]
    conv_filter_size: [[3, 3]]
    conv_maxpool: [[2, 1]]