Configuration¶

Your local config file¶

Create a file in cryocast/config named <chosen-name>.local.yaml. Local config files should inherit from base.yaml and override only what you need:

defaults:
  - base
  - _self_

base_path: /local/path/to/my/data

Run any command with your config using:

uv run cryocast <command> --config-name <your local config>.local

This uses the default model setup (rescaling encoder, small UNet, rescaling decoder), which is sufficient for quick tests but not for larger training runs.

Overriding model parameters¶

To switch to a different named model config or override specific parameters:

defaults:
  - base
  - override /model: cnn_unet_cnn
  - _self_

model:
  processor:
    start_out_channels: 32

base_path: /local/path/to/my/data

You can also override individual options at the command line without a config file:

uv run cryocast <command> ++base_path=/local/path/to/my/data

HPC systems¶

For shared HPC systems (Baskerville, Isambard-AI, or JASMIN), add the matching platform override, which sets the pre-downloaded data path and the right GPU accelerator:

uv run cryocast <command> --config-name <your local config>.local platform=isambardai data=full_north  # or platform=baskerville, or platform=jasmin

Datasets¶

Selecting a dataset¶

The default dataset group is controlled by the data key, which defaults to sample in base.yaml (i.e. data/sample.yaml). To understand how dataset properties are encoded in dataset names, see data/datasets/naming_convention.txt. To define a custom set of datasets, create data/my_datasets.local.yaml:

defaults:
  - datasets:
    - samp_sicsouth_osisaf_25p0km_2017_2019_24h_v2
    - samp_weathersouth_era5_0p5_2017_2019_24h_v2
  - split: sample_dataset
  - _self_

Then reference it from your main config:

defaults:
  - <the base config file you are using>
  - override /data: my_datasets.local
  - _self_

And run with:

uv run cryocast train --config-name my_local_config

Selecting input and target variables¶

The set of variables used as input and target are controlled by the variables config key. For each input dataset group (as defined by the group_as key in datasets) you should define the set of variables from that dataset that you want to use. The same logic applies to the target variables, as shown in this example

variables:
  input:
    era5:
      - 2t
      - msl
    sic-osisaf:
      - ice_conc
  target:
    sic-osisaf:
      - ice_conc

An empty list of target variables (e.g. sic-osisaf: []) selects every input variable in that dataset group.

Generating Argo float missing dates¶

Some dates have no Argo float data. When specifying a new Argo float dataset for the first time it is necessary to generate a list of missing dates for a dataset. This can be done as follows:

1. Add ignore_missing_dates: true to the relevant dataset file.
2. Delete any previously downloaded version of the dataset.
3. Run:

uv run cryocast datasets create --config-name <config that requires this dataset>

This downloads the full dataset, skipping exceptions from missing dates, and prints the missing dates at the end of each data group.