Dataset

Describes how the dataset is parsed and prepared.

Options

• batch_size: batch size for each iteration
• min_value: minimum value (usually 0)
• max_value: maximum value (usually 255)
• clip_value: if true clip values at the end to minimum, maximum
• color_mode: color mode (rgb, rgba, grayscale)
• random_blur: if true apply blur in multiplicative noise
• subsample_size: if > 0 enable subsampling noise (2, 4, 8)
• random_invert: if true randomly invert result
• random_rotate: maximum radians for rotation augmentation
• random_up_down: if true randomly invert up down
• random_left_right: if true randomly invert left right
• input_shape: randomly crop this size from each input image
• additional_noise: select randomly additive noise with mean 0 and std deviation from this list
• multiplicative_noise: select randomly multiplicative noise with mean 1 and std deviation from this list
• inputs: list of dictionaries
  • directory: path to the images directory
  • dataset_shape: resize input dataset to this size (height, width)

Example

  "dataset": {
    "batch_size": 16,
    "min_value": 0,
    "max_value": 255,
    "clip_value": true,
    "random_blur": true,
    "subsample_size": -1,
    "round_values": true,
    "random_invert": false,
    "random_rotate": 0.314,
    "random_up_down": true,
    "color_mode": "grayscale",
    "random_left_right": true,
    "input_shape": [256, 256, 1],
    "multiplicative_noise": [],
    "additional_noise": [5, 10, 20, 40, 60, 90],
    "inputs": [
      {
        "dataset_shape": [256, 768],
        "directory": "/media/data1_4tb/datasets/KITTI/data/depth/raw_image_values/"
      }
    ]
  }