sup3r.preprocessing.samplers.base.Sampler

sup3r.preprocessing.samplers.base.Sampler#

class Sampler(data, sample_shape: tuple | None = None, batch_size: int = 16, feature_sets: dict | None = None, proxy_obs_kwargs: dict | None = None, mode: str = 'lazy')[source]#

Bases: Container

Basic Sampler class for iterating through batches of samples from the contained data.

Parameters:

data (Union[Sup3rX, Sup3rDataset],) – Object with data that will be sampled from. Usually the .data attribute of various Container objects. i.e. Loader, Rasterizer, Deriver, as long as the spatial dimensions are not flattened.
sample_shape (tuple) – Size of arrays to sample from the contained data.
batch_size (int) – Number of samples to get to build a single batch. A sample of (sample_shape[0], sample_shape[1], batch_size * sample_shape[2]) is first selected from underlying dataset and then reshaped into (batch_size, *sample_shape) to get a single batch. This is more efficient than getting N = batch_size samples and then stacking.
feature_sets (Optional[dict]) – Optional dictionary describing how the full set of features is split between lr_features, hr_exo_features, and hr_out_features.

lr_featureslist | tuple
List of feature names or patt*erns to use as low-resolution model inputs. If no entry is provided then all available features from the data will be used.

hr_out_featureslist | tuple
List of feature names or patt*erns that should be output by the generative model and available as ground truth targets. If no entry is provided then all features in lr_features will be used.

hr_exo_featureslist | tuple
List of feature names or patt*erns that should be available as high-resolution model inputs (like topography or observations) or for bespoke loss functions. Features used as inputs are injected into the model mid-network to condition output on high-resolution information. The model configuration should have the appropriate layers to use these features. e.g. Sup3rConcat for topography injection, Sup3rObsModel or Sup3rCrossAttention for obs injection. If no entry is provided then hr_exo_features will be empty.

*To include sparse features as inputs or targets the features must have an “_obs” suffix.
proxy_obs_kwargs (dict | None) – Optional dictionary of keyword arguments to pass to the proxy observation generator. This is only used when training with proxy observations. Top-level keys (onshore_obs_frac, offshore_obs_frac, perturbation_scale) apply to all obs features as defaults. A source-feature-named sub-dict (keyed by the gridded feature name, e.g. u_100m for u_100m_obs) overrides any of those keys for that specific feature:
```
proxy_obs_kwargs = {
    'onshore_obs_frac': {'spatial': [0.3, 0.7], 'temporal': 1},
    'perturbation_scale': 0.01,
    'u_100m': {
        'onshore_obs_frac': {'spatial': 0.9},
        'perturbation_scale': 0.05,
    },
}
```
perturbation_scalefloat
If non-zero, gaussian noise scaled by this value times the per-feature batch standard deviation is added to proxy obs.

onshore_obs_fracdict
Fraction of onshore observations per batch. Keys are ‘spatial’ and ‘temporal’. Each value is a float (fixed fraction) or a [min, max] list to sample uniformly per batch.

offshore_obs_fracdict
Same as onshore_obs_frac but applied where topography <= 0. Ignored when topography is not a source feature.
mode (str) – Mode for sampling data. Options are ‘lazy’ or ‘eager’. ‘eager’ mode pre-loads all data into memory as numpy arrays for faster access. ‘lazy’ mode samples directly from the underlying data object, which could be backed by dask arrays or on-disk netCDF files.

Methods

`check_feature_consistency`()	Check that the feature sets are consistent with each other and the obs features are configured correctly.
`check_proxy_obs_consistency`()	Check that the obs features are configured correctly for proxy observations.
`derive`(feature[, strict])	Resolve feature name to a feature in the underlying data.
`get_sample_index`([n_obs])	Randomly gets spatiotemporal sample index.
`post_init_log`([args_dict])	Log additional arguments after initialization.
`preflight`()	Perform shape and feature checks.
`wrap`(data)	Return a `Sup3rDataset` object or tuple of such.

Attributes

`timer`
`data`	Return underlying data.
`hr_exo_features`	Get a list of exogenous high-resolution features that are only used for training e.g., mid-network high-res topo injection.
`hr_features`	List of feature names or patt*erns that the model is shown at high-resolution.
`hr_features_ind`	Get the high-resolution feature channel indices that should be included for loss calculations.
`hr_out_features`	List of feature names or patt*erns that should be output by the generative model.
`hr_sample_features`	List of feature names used in the sample index for the high-resolution training data.
`hr_sample_shape`	Shape of the data sample to select when __next__() is called.
`hr_source_features`	List of feature names or patt*erns that should be available natively as high-resolution.
`lr_features`	List of feature names or patt*erns to use as low-resolution model inputs.
`lr_features_ind`	Get the low-resolution feature channel indices that should be included for training.
`obs_features`	List of feature names or patt*erns that should be treated as observations.
`obs_features_ind`	Get the source feature indices for each obs feature's corresponding base feature (e.g. `u_100m` for `u_100m_obs`).
`real_obs_features`	Obs features that actually exist in the data.
`sample_shape`	Shape of the data sample to select when `__next__()` is called.
`shape`	Get shape of underlying data.
`use_proxy_obs`	Whether to use proxy observations.

property use_proxy_obs#: Whether to use proxy observations. When True, proxy observation features are generated by masking the corresponding gridded ground truth data and are appended to the samples. The obs features are specified by the obs_features argument and should have a corresponding source feature in the data features that is used for sampling. For example, an obs feature named temperature_obs would be generated from the gridded ground truth feature named temperature.

get_sample_index(n_obs=None)[source]#

Randomly gets spatiotemporal sample index.

Returns:: sample_index (tuple) – Tuple of latitude slice, longitude slice, time slice, and features. Used to get single observation like self.data[sample_index]

Notes

If n_obs > 1 this will get a time slice with n_obs * self.sample_shape[2] time steps, which will then be reshaped into n_obs samples each with self.sample_shape[2] time steps. This is a much more efficient way of getting batches of samples but only works if there are enough continuous time steps to sample.

preflight()[source]#: Perform shape and feature checks.

check_proxy_obs_consistency()[source]#: Check that the obs features are configured correctly for proxy observations.

check_feature_consistency()[source]#: Check that the feature sets are consistent with each other and the obs features are configured correctly.

property sample_shape: tuple#: Shape of the data sample to select when __next__() is called.

property hr_sample_shape: tuple#: Shape of the data sample to select when __next__() is called. Same as sample_shape

property lr_features#: List of feature names or patt*erns to use as low-resolution model inputs. If no entry is provided then all available features from the data will be used.

property hr_source_features#: List of feature names or patt*erns that should be available natively as high-resolution. For a non-dual sampler this is all features, since even features only provided to the model as low-resolution still need to be coarsened from the high-resolution data. This is in contrast to dual samplers (DualSampler), where there are separate high-resolution and low-resolution data members.

property hr_features#: List of feature names or patt*erns that the model is shown at high-resolution. This does not include features that are only shown to the model after coarsening. Thus, this includes hr_out_features and and hr_exo_features but not lr_features.

property real_obs_features#: Obs features that actually exist in the data. In contrast to proxy obs, which are generated by masking the gridded data, these features are present in the raw data and are sampled directly. These are a subset of obs_features.

property hr_sample_features#: List of feature names used in the sample index for the high-resolution training data. When using proxy obs, obs features that are not present in the raw data are excluded from sampling (they will be generated as proxy obs). Obs features that ARE in the data are included so they get sampled.

property hr_out_features#: List of feature names or patt*erns that should be output by the generative model. If no entry is provided then all features in hr_features will be used.

property hr_exo_features#: Get a list of exogenous high-resolution features that are only used for training e.g., mid-network high-res topo injection. These must come at the end of the high-res feature set. These can also be input to the model as low-res features.

property obs_features#: List of feature names or patt*erns that should be treated as observations. These features will be included in the high-res data but not the low-res data and won’t necessarily be expected to be output by the generative model. These are different from other hr_exo_features in that they are intended to be used as observation features with NaN values where observations are not available.

property hr_features_ind#: Get the high-resolution feature channel indices that should be included for loss calculations. This includes hr_out_features and hr_exo_features, Any high-resolution features that are only included in the data handler to be coarsened for the low-res input are removed.

property lr_features_ind#: Get the low-resolution feature channel indices that should be included for training. This includes lr_features.

property obs_features_ind#

Get the source feature indices for each obs feature’s corresponding base feature (e.g. u_100m for u_100m_obs). Used to extract gridded truth for proxy obs generation.

Returns:: list[int] – Indices into hr_sample_features for each obs base feature.

property data#

Return underlying data.

Returns:: Sup3rDataset

sup3r.preprocessing.samplers.base.Sampler

Contents

sup3r.preprocessing.samplers.base.Sampler#