DataGenerator

Generation of shuffled and augmented images with coordinates of anterior and posterior points for every batch.

Parameters and output

Parameters:

  • imgs (DataFrame): First coulmn: image ID, second column: path to image

  • segs (DataFrame): First coulmn: image ID, second column: path to segmentation

  • df_coordinates (DataFrame): z: image ID, ap: anterior point, pp: posterior point

  • batch_size (int): Batch size

  • target_height (int, optional): Target height of images, default = 512 (2D-versions) or 256 (3D-versions)

  • target_width (int, optional): Target width of images, default = 256 (2D-versions) or 128 (3D-versions)

  • augment (bool, optional): If augmentation should be done, default = True

  • shuffle (bool, optional): Shuffling of images, default = True

  • random (int, optional): Set seed, default=42

  • radius (except GlottisNetV1) (int, optional): Radius of circle of anterior and posterior points in prediction maps, default = 15 (and GlottisNetV2a-e),

7.5 (GlottisNetV2 Channels, GlottisNetV2 3DConv, GlottisNetV2 LSTM)

Returns:
  • GlottisNetV1:
    • Shape: 4D-tensors (batch, width, height, channels)

      • Augmented and shuffled input images for training of neural network

      • Tuple of segmentation map and 4 coordinates [x (anterior), y (anerior), x(posterior), y(anterior)] for each image in batch

  • GlottisNetV2 (a, b, d, e):
    • Shape: 4D-tensors (batch, width, height, channels)

      • Augmented and shuffled input images for training of neural network

      • Tuple of segmentation map, prediction maps (anterior and posterior point stored in 2 channels)

  • GlottisNetV2c
    • Shape: 4D-tensors (batch, width, height, channels)

      • Augmented and shuffled input images for training of neural network

      • Tuple of segmentation map, prediction for anterior point, prediction map for posterior point

  • GlottisNetV2_3DConv and GlottisNetV2_LSTM
    • Shape: 5D-tensors (batch, frames, width, height, channels)

      • Augmented and shuffled input images for training of neural network

      • Tuple of segmentation map and prediction maps (2 channels)

  • GlottisNetV2_Channels
    • Shape: 4D-tensors (batch, width, height, channels)

      • Augmented and shuffled input images (n_frames channels) for training of neural network

      • Tuple of segmentation map and prediction maps (2*n_frames channels)

Helper Functions in DataGenerator

__len__():

Calculating number of batches per epoch.

_on_epoch_end():

Shuffling of training data after each epoch.

_get_augmenter():

Augmentation of input images.

_create_prediction_map(ap_coord, img_shape, radius):

Create prediction maps with circle of specific radius at the coordinates of anterior and posterior points.