Skip to content

NectarGAN Toolbox - Dataset

NectarGAN Toolbox - Home

[!IMPORTANT] All data augmentations have a chance (defined by their percentage parameter) of being applied any time a dataset image is loaded, Not just the first time a given image is loaded. So a dataset image may be loaded with one set of augmentations one epoch, and a completely different set the next epoch.

Dataset Files

Setting Description
Dataset Root The root directory for your dataset files. This directory should have subdirectories called train, val, and, optionally, test. These subdirectories should contain your paired (A
Load Size The resolution (^2) to load each dataset image at. This is separate from Crop Size, which is the actual resolution of the images used during training, although they can be the same value if you don't want to use random cropping.
Crop Size The resolution (^2) of the images used during training. If this is smaller than Load Size, an Albumentations RandomCrop operation will be performed to reach this target size.
### Loading
Setting Description
:---: ---
Direction What direction to load the dataset (or, in simpler terms, which image is input and which is output where A is the left image, and B is the right image.)
Batch Size The batch size to use during training.
### Augmentations (Input)
Setting Description
:---: ---
Colorjitter The percentage chance to apply random colorjitter to any given input image.
Colorjitter Range The min and max allowed values of the colorjitter if it is applied.
Gaussian Noise The percentage chance to apply gaussian noise to any given input image.
Gaussian Range The min and max range of the gaussian noise.
Motion Blur The percentage chance to apply a pseudo-motion blur effect to any given input image.
Motion Blur Limit The maximum kernal size for the blur operation. The minimum allowed kernal size is 3 so this value should be >3.
Random Gamma The percentage chance to randomize the gamma of any given input image.
Gamma Limits The upper and lower bounds (in %) of the random gamma operation.
Grayscale The percentage chance to convert any image to grayscale
Grayscale Method What method to use when converting dataset images to grayscale.
Compression The percentage chance to apply simulated compression artifacts to any given dataset image.
Compression Type What type of compression artifacts to apply (e.g. JPEG, WEBP).
Compression Quality Min and max quality for the applied compression operation.
### Augmentations (Both)
Setting Description
:---: ---
Horizontal Flip The precentage chance to flip each image of any given dataset image along the horizontal axis.
Vertical Flip The percentage chance to flip each image of any given dataset image along the vertical axis.
90° Rotation The percentage chance to apply a random 90° stepped rotation to each image of any given dataset pair.
Elastic Xform The percentage chance to apply an elastic transformation to each image of any given dataset pair.
Elastic Alpha Scaling factor for the elastic distortion. Higher values will produce more deformation.
Elastic Sigma Smoothing factor for the distortions. Higher values will produce smoother results.
Optical Distort Percentage chance to apply an optical distortion effect to any given dataset image.
Distortion Limits Min and max range of the applied distortions.
Distortion Mode What type of optical distortion to apply (e.g. camera or fisheye distortion).
Coarse Dropout The percentage chance to apply coarse dropout (i.e. random rectangular cutouts) to any given dataset image.
Hole Count The min and max number of holes to add if coarse dropout is applied to an image.
Hole Height Min and max allowable height of each hole.
Hole Width Min and max allowable width of each hole.

See Also: | Toolbox - Home | Experiment | Training | Testing | Review | Utilities | Settings |