2021-paper-reproduction.md

Reproducing results from the Guillaumin (2021) paper

See README.md at repo root for further details. Here, we will provide just commands and commentary.

Options such as --out-dir will be omitted. (The script will prompt you for missing options. You can display help by adding --help to your invocation.)

1. Training data generation

python src/gz21_ocean_momentum/cli/data.py \
--config-file resources/cli-configs/data-paper.yaml

Unclear whether you may need --ntimes 4000.

2. Model training

Not tested due to issues with training.

Model hyperparameters adapted from Table A1.

python src/gz21_ocean_momentum/cli/train.py \
--config-file resources/cli-configs/train-paper.yaml \
--subdomains-file resources/cli-configs/train-subdomains-paper.yaml \
--train-split-end 0.8 --test-split-start 0.85

Add --in-train-data-dir <forcings generated above>.

3. Inference

The CLI inference script has no configuration other than model to predict on, and input low-resolution data to predict forcings of:

python src/gz21_ocean_momentum/cli/infer.py

Currently will not reproduce the same predictions as used in the paper. See #97 for further details.

For --model-state-dict-file, you may use a pretrained model instead of running the training described above. A low-resolution one is provided here: https://huggingface.co/M2LInES/gz21-ocean-momentum/blob/main/low-resolution/files/trained_model.pth

Similarly, instead of generating forcings as above, you may use pre-generated data for --input-data-dir. Low-resolution (~100 timepoints) CM2.6 data: https://huggingface.co/datasets/M2LInES/gz21-forcing-cm26/tree/main/forcing