Hyper Parameter - choosehappy/QuickAnnotator GitHub Wiki

Hyper-Parameter Explanation

QA is shipped mostly fully configured and the few hyper-parameters of interest are easily modifiable via the configuration file. These hyper-parameters can be set in the config.ini file

  • [common]
  • [flask]
  • [cuda]
  • [sqlalchemy]
  • [pooling]
  • [train_ae]
  • [train_tl]
  • [make_patches]
  • [make_embed]
  • [get_prediction]
  • [frontend]
  • [superpixel]

When annotating different primitives, the user needs to focus on hyper-parameters, edgeweight, approxcellsize, and compactness.

  • edgeweight: The edgeweight is hyper-parameter in [train_tl]. Setting a higher edgeweight encourages the DL model to focus the loss function on incorrectly classified boundary pixels; increasing this weight is beneficial when clear boundaries are hard to distinguish.
  • approxcellsize: The approxcellsize is hyper-parameter in [superpixel]. This is set to the approximate width of the desired superpixel, and works well when set to the approximate width of the object of interest.
  • compactness: The compactness is hyper-parameter in [superpixel]. The nonnegative compactness value determines the regularity of the superpixel boundary, wherein higher compactness encourages superpixels to retain their initial square shape, while lower compactness allows for greater boundary irregularity.