euroc_imudb.yaml

# 3 stages:
#  1. poc: validate the idea
#  2. dev: try to get good results
#  3. paper: very organized experiments for good data for the paper
#  4. prod: for production
stage: 'dev'
datasets_name: 'euroc_imudb'
experiment_name: 'imudb'
experiment_management_log: 'experiments/experiment_management_log.csv'
logs_dir: "logs"
note: "bug fix use raw and pred for nsp, weight denoise loss with 10"

data:
  shuffle:
    train: true
    val: true
    test: true
  batch_size: 1024

  # where are raw data ?
  data_dir: "/root/EuRoC/Raw"
  # where record preloaded data ?
  caches_dir: "/root/EuRoC/Preload"
  # set train, val and test sequence
  train_seqs: [
      'MH_01_easy',
      'MH_05_difficult',
      'V2_01_easy',
      'V2_03_difficult',
      'MH_03_medium',
      'V1_02_medium',
  ]
  val_seqs: [
      'MH_02_easy',
      'MH_04_difficult',
      'V2_02_medium',
      'V1_03_difficult',
      'V1_01_easy',
  ]
  test_seqs: [
      'MH_02_easy',
      'MH_04_difficult',
      'V2_02_medium',
      'V1_03_difficult',
      'V1_01_easy',
  ]
  mask:
    mask_ratio: 0.15
    mask_alpha: 6
    max_gram: 10
    mask_prob: 0.8
    replace_prob: 0.0

model:
  inputs:
    imu_sequence_number: 30 #120 # 200 Hz, namely, X /200 second
  outputs:
    imu_sequence_number: 30 #future IMU predict sequence number # This is not being used yet # 10 # 200 Hz, namely X /200 second
  hyper_params:
    starting_learning_rate: 0.001
    denoise_loss_weights: 10
    nsp_loss_weights: 1
    mlm_loss_weights: 1
    T_0: 5000
    T_mult: 10
    eta_min: 1e-8

    nn_arch: 'limu_bert' # This is not used in the experiment. Just for reference
    # Below are copied from https://github.com/dapowan/LIMU-BERT-Public/blob/master/config/limu_bert.json#L4-L9
    feature_num: 6
    hidden: 72
    hidden_ff: 144
    n_layers: 4
    n_heads: 4
    seq_len: 120
    emb_norm: true
  onnx:
    inputs: ['x_imu']
    outputs: ['hat_imu']


sensors:
  cam0:
    frequency: 20 # hz
    model: 'MT9V034'
    info: 'global-shutter, monochrome' # This is purely to provide the info of the sensor
  imu:
    frequency: 200 # hz
    model: 'ADIS16448'
    info: 'MEMS based, intrinsic calibrated'
  synhronization: 'hardware-synchronized'
  ground_truth:
    frequency: 20 # Hz
    pose:
      methods: 'vicon'
      frequency: 100 # hz
    position:
      methods: 'Leica MS50'
      frequency: 20 # hz
      error_bound: 0.001 #m