sl_kitti.yaml

name: 'sl_kitti_sample_01_stride_05_beam_64_downsample_05'

tensors:
  placeholders:
    X:
      shape: [32, 256, 32]
    Y:
      shape: [ 1 ]
  hyper_params:
    base_depth: 32
    classification_loss_weight: 0.01
    cross_modal_recon_loss_weight: 1.0

train:
  stop_val_loss_decrease: -0.05
  first_stage_epochs: 15
  epochs: 30
  val_epoch: 1
  save_epoch: 1
  batch_size: 32
  start_learning_rate: 0.0001
  stage2_learning_rate: 0.00001
  lr_decay_step: 300000
  lr_decay_rate: 0.1
  pre_trained_weights: []
  continue_training: False
  optimizer_var_list: []
  devices:
    GPU: 1

machine:
  save_dir: './results/save/'
  log_dir: './results/log/'

data:
  num_parallel_reads: 16
  inputs:
    # Ensure, this key is the same the model.tensor_dict
    # Below are to be resized and sliced in pre-process to match the placeholders
    X:
      modality: 'image'
      data_type: 'float32'
      nhwc: True
      H: 32
      W: 256
      C: 32
      feature_names:
        - 'X'
    Y:
      modality: 'scalar'
      data_type: 'float32'
      shape: [ 1 ]
      # Feature listed below will be random chosen on the fly
      feature_names:
        - 'Y'
  # Below are configs for tfrecords files
  compression_type: '' # no compression
  suffix: 'tfrecord'
  tfrecords_train_dirs:
    - "Replace this with where you saved the training tfrecords"
  tfrecords_test_dirs:
    - "Replace this with where you saved the validation tfrecords"


inference:
  included_tensor_names:
  - 'sl_kitti_sample_01_stride_05_beam_64_downsample_05'
  - 'classifier'

  freeze:
    output_node_name: 'sl_kitti_sample_01_stride_05_beam_64_downsample_05/prediction_from_classifier'