unet_train#

Training parameters including batch size, epochs, learning rate, weight decay, loss function (dice or bce), number of workers, whether to use the GPU

batch_size: 1
epochs: 1
learning_rate: 0.01
weight_decay: 0.0001
loss_fn: bce
num_workers: 1
use_gpu: true

Threshold/interpolation to apply to image For interpolation options are: log2, log10 or linear

img_interpolate: log2
img_threshold: 0

Transformations to apply during training/test

train_transforms:
  dtypeconv: null
  erasing: null
  h_flip: null
  perspective: 0.2
  rotation: 90
  v_flip: null
test_transforms:
  dtypeconv: null

Weights and bias parameters

wandb_dataset: c15_imgs
wandb_project: c15_img_unet_bce

The following is not generic, if you need to use this please raise an issue and tag @oubino

For two channel image, we visualise both channels then sum them together Need the name of the first and second channel in terms of the real concepts

channels:
- egfr
- ereg

Whether to sum the two channels

sum_chan: true