Commands¶

datasets create¶

You will need a CDS account to download ERA5 data with anemoi. By default, a sample dataset will be downloaded which should be small enough to fit on your personal computer. The full datasets are available on Isambard.

uv run cryocast datasets create

Anemoi tracks which date groups have been downloaded, so an interrupted download can be resumed by simply rerunning this command.

To create the synthetic dataset, use:

uv run cryocast datasets create --config-name synthetic

datasets inspect¶

uv run cryocast datasets inspect

Prints basic properties of each dataset. With the --verbose option it will also print statistical summaries of the variables.

datasets plot¶

uv run cryocast datasets plot --dataset samp-sicsouth-osisaf-25p0km-2020-2024-24h-v1 --timestep 0

Creates one static PNG per variable for the selected timestep of a configured downloaded dataset. Plots are written to ${base_path}/data/input_plots in a subdirectory named after the dataset. Running without --dataset will plot every configured dataset. This is useful for inspecting raw inputs without running model training or evaluation.

Use --timestep to select another dataset index and the normal --config-name to choose the dataset configuration.

Pass --video to create an animation over consecutive timesteps instead of a single static plot:

uv run cryocast datasets plot --dataset samp-sicsouth-osisaf-25p0km-2020-2024-24h-v1 --video --timestep 0 --n-steps 30

This animates --n-steps consecutive timesteps starting at --timestep, writing one video file per variable to the same input_plots directory.

train¶

Standard (non-synthetic) runs use Weights & Biases. Generate an API key, then authenticate:

export WANDB_API_KEY=<your_api_key>
wandb login

Trains the model end-to-end:

uv run cryocast train

PYTORCH_ENABLE_MPS_FALLBACK=1 uv run cryocast train

Multistage training¶

For EncodeProcessDecode models, pass --multistage to train each component separately before finetuning.

uv run cryocast train --multistage

Checkpoints are saved to ${BASE_DIR}/training/wandb/run-<date>-<id>/checkpoints/<name>.ckpt, where BASE_DIR is the base path defined in your config.

See Train in stages for a full walkthrough.

Weights & Biases logging¶

To disable logging to W&B, set either reporting.loggers.wandb.offline=true or the WANDB_MODE=offline environment variable

uv run cryocast train reporting.loggers.wandb.offline=true
WANDB_MODE=offline uv run cryocast train

Run data such as metrics and figures will still be written locally, but will not be uploaded.

Synthetic experiments do not use W&B. Use the synthetic configuration, which saves metrics and plotting artefacts locally under ${BASE_DIR}/report:

uv run cryocast train --config-name synthetic

sweep initialise¶

uv run cryocast sweep initialise --sweep-yaml example.sweep.yaml --config-name baseline/02_cnn_unet_cnn

Creates a W&B sweep and initialises a local Optuna study directory; hyperparameters are sampled per trial at runtime. See Run a hyperparameter sweep for the full workflow.

sweep trial¶

uv run cryocast sweep trial --sweep-path <path to sweep directory created above>

Runs a single hyperparameter trial as part of a W&B sweep. See Run a hyperparameter sweep for the full workflow.

sweep summarise¶

uv run cryocast sweep summarise --sweep-path <path to sweep directory created above>

Reads the local Optuna study and reports the number of completed trials, plus the value and hyperparameters for the best trial. This works without a W&B connection. See Run a hyperparameter sweep for the full workflow.

evaluate¶

uv run cryocast evaluate --checkpoint PATH_TO_A_CHECKPOINT

Pass --save-predictions to save denormalised model predictions for the configured test period in NetCDF format, written to predictions.nc in the run's directory:

uv run cryocast evaluate --checkpoint PATH_TO_A_CHECKPOINT --save-predictions

Use data.split.test to control which dates are exported. NetCDF export currently requires single-process evaluation.

Visualisations¶

To plot static images or animations of the raw input data, add the following to your local config:

evaluate:
  callbacks:
    plotting:
      make_input_plots: true

Output directories, styling, and animation parameters can be altered by changing config.evaluate.callbacks.plotting.plot_spec. Any of these can be overridden at the command line.

checkpoint upgrade¶

uv run cryocast checkpoint upgrade PATH_TO_A_CHECKPOINT [PATH_TO_ANOTHER_CHECKPOINT ...]

Upgrades checkpoints saved by older versions of CryoCast, or by its predecessor icenet-mp, so that they can be loaded safely. The run's files/model_config.yaml is also upgraded, replacing references to icenet_mp and converting the legacy predict settings into variables and window settings. Each file that needs upgrading is backed up to <name>.bak before being rewritten. Files that are already up to date are left unchanged, so it is safe to upgrade a checkpoint more than once, and an existing backup is never overwritten. If an upgraded checkpoint would still contain objects that cannot be loaded safely, the upgrade stops with an error naming them and no files are changed.

uv run cryocast checkpoint upgrade --trust PATH_TO_A_TRUSTED_CHECKPOINT

feature-importance¶

uv run cryocast feature-importance

Fits a single Random Forest to predict the configured training target from the spatial and temporal mean of each input variable, then prints the variables ranked by the forest's built-in feature importance. This is an exploratory signal for which inputs the data supports, not a model-quality metric. The target group's own variables are included as features and typically dominate the top of the ranking, since the target's previous step is usually its best predictor.