Top-level options

Inference

The following options control the inference process:

checkpoint:

The only compulsory option is checkpoint, which specifies the checkpoint file. It can be a path to a local file, or a huggingface config.

checkpoint: /path/to/checkpoint.ckpt

checkpoint:
   huggingface:
      repo_id: "ecmwf/aifs-single"
      filename: "aifs_single_v0.2.1.ckpt"

device:

The device option specifies the device to use for inference. The default is cuda.

precision:

The precision option specifies the precision to use for inference. The default is taken from the checkpoint.

date:

The date option specifies the reference date of the forecast (the starting date). Is the date is not specified, the default will depend on the selected input. For example, grib will use the date found in the file, while mars will use yesterday’s date.

Setting the date in the configuration file can be useful for debugging. A more common use case is to set the date in the command line, as shown in the anemoi-inference run documentation.

anemoi-inference run config.yaml date=2020-01-01T00:00:00

lead_time:

The lead_time option specifies the forecast lead time in hours. Default is 240 hours.

write_initial_state:

The write_initial_state option specifies whether to write the initial state to the output (i.e. “step zero” of the forecast). Default is true.

allow_nans:

The allow_nans option specifies whether to allow NaNs in the input and output. It set to null (default), the value is set internally to true if any of the input fields contain NaNs, otherwise it is set to false. You can override this behaviour by setting it to true or false in the configuration file.

Inputs and outputs

The entries for the inputs and outputs are specified in the inputs and inputs sections of the documentation.

Debugging

verbosity:

The verbosity option specifies the verbosity level of the output. It is set to 0 by default.

report_errors:

The report_errors option specifies whether to produce a longer error report when the code of the model cannot be loaded. The aim od that report is to troubleshoot versioning issues (git branches, python modules, etc.). It is set to false by default.

Miscellaneous

use_grib_paramid:

The use_grib_paramid option specifies whether to use the eccodes paramId instead of parameter names when appropriate. It is set to false by default.

env:

The env option specifies a dictionary of environment variables to set before running the inference. This can be useful to set PyTorch or OpenMP environment variables. Note that the environment variables may be set too late in some cases.