Parameter Reference¶

This page describes every field in the files under conf/, the additional arguments supported by selectable datasets, models, and losses, and how main.py uses them. Defaults below come from the repository; Python constructor defaults are identified separately. For complete commands using the WPR dataset, see Workflow Examples. For internal function signatures, see the API.

How to configure parameters¶

Run commands from the repository root with the package installed in the active environment. Hydra composes conf/config.yaml and the selected groups. _target_ is the Python path of the class Hydra instantiates; normally change the group rather than this field. ${name} references another configuration field, not a Bash variable.

A multirun actually starts training jobs. Add the Zarr path and use separate checkpoint directories for each combination.

python main.py task=train training=default model=unet/direct \
  dataset.dataset.zarr_path=/prj/cptec/nowcasting/data/benchmark/benchmark_dataset_wpr.zarr \
  training.lr=0.0003 dataset.train_loader.batch_size=2 \
  --cfg job --resolve

Groups and global parameters¶

The current dataset returns one channel per frame. Increasing input_channels does not stack variables: that requires adapting the loader. The autoregressive UNet fixes 12/6 in its YAML; also override model.input_timesteps and model.output_timesteps when changing the global values.

Dataset: dataset.dataset¶

The double prefix is intentional: dataset contains the dataset object, loaders, and split overrides. _target_ points to ainpp_pb_latam.datasets.gsmap.AINPPPBLATAMDataset.

The following arguments exist in the constructor but not in this object's YAML. Add them with +dataset.dataset.<field>=... when needed.

The loader replaces NaNs with zero. Windows are index-based: it does not validate hourly timestamp continuity. The frame count must accommodate input+output length. The final patches may overlap at the boundary even when the stride equals the patch size.

Splits: dataset.overrides¶

Full validation example: dataset.overrides.validation.steps_per_epoch=null. With 500 samples and batch size 2, the loader has 250 batches. Reducing the stride increases the number of candidate windows; it does not change the 500-sample limit when that limit is set.

Loaders¶

These dictionaries are passed to DataLoader; additional arguments depend on the installed PyTorch API. Avoid adding shuffle to the validation loader: the CLI already passes shuffle=False explicitly for evaluation/historical inference.

Training: training¶

Adam is fixed in build_optimizer; there is no optimizer selector in the YAML.

Checkpoints and early stopping¶

Both profiles share these values.

Best model: best_model.pt. Periodic checkpoints: checkpoint_model_epoch_005.pt, etc. These are state_dicts without Adam state or an epoch counter. Disabling early stopping also prevents best-model saving through that mechanism. Sample images are written to samples/, a fixed path in the engine.

Models: model¶

All models must respect the (B,T,C,H,W) tensor contract. When changing the architecture, use the same arguments when loading its checkpoint.

Direct and autoregressive UNet¶

python main.py training=default model=unet/direct \
  'model.features=[32,64,128,256]' model.nonnegativity=softplus --cfg job --resolve

ConvLSTM¶

This constructor has no model.input_timesteps argument: context comes from the tensor.

AFNO¶

The temporal fields exist in the YAML; the others require +model.<field>=....

Example: model=afno/direct '+model.img_size=[320,320]' +model.depth=4.

ResNet50, InceptionV4, and Xception¶

All expose model.input_timesteps=${input_timesteps} and model.output_timesteps=${output_timesteps}. The additional argument pretrained has a Python default of true: it loads encoder weights and may require a download/cache. Use +model.pretrained=false to initialize without those weights. These constructors do not expose input_channels; the workflow assumes one channel per frame. Example: model=resnet50/direct +model.pretrained=false.

Losses: loss¶

Thresholds use the target units (mm/h for untransformed hourly GSMaP). They do not automatically convert units or normalize data.

Hybrid losses require one weight per component. Use an index to edit an item:

python main.py training=default loss=hybrid_mse_ssim \
  'loss.weights=[1.0,0.3]' loss.losses.0.alpha=3.0 \
  loss.losses.1.window_size=7 --cfg job --resolve

window_size is the SSIM spatial window (a positive odd number); in_channels must match the data. layer_ids selects VGG16 feature layers. Perceptual loss attempts to load pretrained weights and falls back to MSE if that fails. The Python default for layer_ids is [3,8,15,22], unlike the sota group.

BinaryFocalLoss expects logits; changing the loss of a network with ReLU outputs does not turn the pipeline into a suitable classification setup. CrossEntropyLoss is available through the API, without a YAML group: its weights=None argument accepts class weights, and it requires logits/integer classes, which are not directly compatible with the current regression targets. Standalone SSIM and Perceptual losses do not have their own YAML groups either.

Discriminator: discriminator¶

patchgan points to PatchDiscriminator3D. These fields only take effect when a workflow instantiates and uses that discriminator; main.py does not.

Inference: inference¶

Use dataset.val_loader.batch_size to control memory/batches for historical inference. Single-sample prediction uses a fixed CLI timestamp (20260316_1200); there is no date-selection argument. Current files do not preserve real patch timestamps/coordinates. The historical store contains predictions with axes (sample,horizon,channel,height,width), without mosaic reconstruction.

Evaluation: evaluation¶

For hourly data, record 'evaluation.lead_times_min=[60,120,180,240,300,360]' in the experiment configuration, but interpret T+1 as the first predicted frame: label conversion to minutes has not been implemented. For a quick check, use +dataset.overrides.test.steps_per_epoch=8 (random sampling), not evaluation.max_batches. Parquet requires pyarrow or fastparquet.

Visualization: visualization¶

The visualization/default.yaml group is not in the root defaults. +visualization=default includes it, but the CLI evaluator only uses visualization.output_dir; it does not pass the style profile to the figure generator.

These style fields are for the separate API/runner, which expects metrics.json and/or sample_*.npz. Do not assume the evaluation CSVs feed this runner.

Hydra, paths, and environment¶

hydra.run.dir does not automatically redirect every artifact. Configure training.checkpoint.dir, inference.output_dir, evaluation.output_dir, and visualization.output_dir for the workflow. Hydra logs/configurations record overrides, but do not guarantee scientific determinism.

There are no active CLI parameters for AMP, gradient accumulation, optimizer selection, complete training resumption, or DDP initialization. Use the fields actually consumed above; adding an arbitrary key with + does not implement the corresponding functionality.