configuration files

Configuration File

The core of our system is the configuration file which is passed to training/testing script:

python run.py --config path/to/config/file

Here, all needed parameters are specified such as the location of the dataset, the model definition, the batch generator for generating blocks as well as the optimizer and training parameters.

Datasets, models, batch generators and optimizers are referenced by their module name and are dynamically loaded. This makes adding a new model or batch generator very simple if you derive it from the predefined base classes.

Example Configuration File

modus: TRAIN_VAL # {TRAIN_VAL, TEST}
dataset:
  name: general_dataset # name of the module in the package datasets
  num_classes: 13
  data_path: dataset/stanford_indoor/ # path to the datasets with the ensured structure
  test_sets: ['Area_5'] # specify which areas are only for validation
  downsample_prefix: sample_0.03 # folder name of downsampled point cloud version
  colors: True
  laser: False
model:
  name: multi_scale_cu_model # name of the module in the package models
batch_generator:
  name: multi_scale_batch_generator # name of the module in the package batch_generators
  params: # parameters for batch generator which are directly passed to it
    batch_size: 24
    num_points: 4096
    grid_spacing: 0.5
    metric: chebyshev
    augmentation: True
    radii: [0.25, 0.5, 1.0]
optimizer:
  name: exponential_decay_adam # name of the module in the package optimizers
  params: # parameters for the optimizer which are directly passed to it
    initial_lr: 0.001
    decay_step: 300000
    decay_rate: 0.5
train: # training parameters
  epochs: 150 # maximal number of epochs to run