Implemetation of the U-NET Structure. Built based on https://www.youtube.com/watch?v=IHq1t7NxS8k and modified further.
The image shape must be (n, m, ny, nx)
where
n
: number of imagesm
: image depts (e.g. rgb=3, gray=1)ny
: vertical image size [pixel]nx
: horizontal image size [pixel]
The label shape must be (n, m, ny, nx)
where
n
: number of imagesm
: number of features, typically 1ny
: vertical image size [pixel]nx
: horizontal image size [pixel]
Besides the obvious parameters like device
, batch_size
etc,
you can adjust/vary the following parameters using the hyperparameters.yaml
:
- number of
features
for each level and the number of levels stride
andkernel_size
for down- and up-path and for thebottleneck
If your runs are managed using hydra
, the results can by nicely monitored
using tensorboard
.
There's an example pytho file using hydra in the example directory of this repo: example/using_hydra.py
.
For that, navigate to the respective directory and run tensorboar --logdir