Configuration options

model_arch : nn.Module to be used for the attention net F and the feature extraction of the scale net G. Must return a dictionary with 'attention' and 'fg_feats' as keys, which represent, respectively, the attention output of F, and the feature vector from G
scale_arch : nn.Module to be used to classify fg_feats obtained from G
freeze_model : Whether to freeze the attention and scale networks or not
ndf : Width parameter for model_arch
nc : Number of channels in input images
attention_sparsity : The desired sparsity. Specified as a real number in [0, 1]
attention_sparsity_r : Parameter determining the compression in the sigmoid

batch_size : Training batch size to use
lr : Desired learning rate
momentum : Desired momentum for optimiser
weight_decay : L2 regularisation weight hyperparameter for optimiser
n_epochs : Number of training epochs
losses : A list specifying which losses to use for training. See variable loss_definitions in train.py for possible losses
predictions : A list specifying what predictions to make
checkpoint_special : A list or an int. A list specifies which epochs to keep specific checkpoints. An int specifies a checkpoint every so many epochs
lr_decay : Decay factor for learning rate
lr_decay_every : LR is decayed at ...
equivariance_scale : Whether to enforce equivariance to scale for attention network F. Preferably False
equivariance_aug : Whether to enforce equivariance to rigid transforms for attention network F. See utils.py for possible values for this option
optimiser : Which optimiser to use. Allowed values are 'adam' and 'sgd'.
pixel_means : Pixel means for input images
pixel_stds : Pixel stds for input images
max_train_batches_per_epoch : Number of maximum training batches per epoch
max_val_batches_per_epoch : Number of maximum validation batches per epoch
use_colour_transform : Whether to use colour augmentation
use_image_transforms : Whether to use geometric augmentation

dset_name : Name of dataset
image_size : Sizes of input images. If an image is not this size, it is padded by placing the original in the top-left corner
patch_size : Sizes of patches to pick out from input images
workers : Number of dataset workers
dataroot : Path to dataset files
stain_normaliser_file : Relative path to stain normalisation target (relative to dataroot)
hed_decomp : Whether to use HED decomposition instead of RGB images
hed_channels : Which HED channels to use
seg_threshold : Threshold to reject input tiles. This value in [0, 1] specifies at least what fraction of an input image must be tissue in order for it to be used
levels : Magnification levels to use for training. OpenSlide's convention is followed, so 'max' denotes the maximum magnification, while '-1' and '-2' denote one and two lower levels of magnification, respectively.
splits_file : Splits file specifying train/val/test split in dataroot.
seg_cover_file: File in dataroot specifying fraction of tissue cover for extracted images. This file for MoNuSeg images extracted with the provided code is included in data_processing.

Provide feedback