ocf-data-sampler 0.5.15__tar.gz → 0.5.16__tar.gz

{ocf_data_sampler-0.5.15 → ocf_data_sampler-0.5.16}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ocf-data-sampler
-Version: 0.5.15
+Version: 0.5.16
 Author: James Fulton, Peter Dudfield
 Author-email: Open Climate Fix team <info@openclimatefix.org>
 License: MIT License

{ocf_data_sampler-0.5.15 → ocf_data_sampler-0.5.16}/ocf_data_sampler/numpy_sample/nwp.py

@@ -28,7 +28,7 @@ def convert_nwp_to_numpy_sample(da: xr.DataArray, t0_idx: int | None = None) ->
         NWPSampleKey.channel_names: da.channel.values,
         NWPSampleKey.init_time_utc: da.init_time_utc.values.astype(float),
         NWPSampleKey.step: (da.step.values / 3600).astype(int),
-        NWPSampleKey.target_time_utc: da.target_time_utc.values.astype(float),
+        NWPSampleKey.target_time_utc: (da.init_time_utc.values + da.step.values).astype(float),
     }
 
     if t0_idx is not None:

ocf_data_sampler-0.5.16/ocf_data_sampler/select/diff_channels.py

@@ -0,0 +1,25 @@
+"""Takes the diff along the step axis for a given set of channels."""
+
+import numpy as np
+import xarray as xr
+
+
+def diff_channels(da: xr.DataArray, accum_channels: list[str]) -> xr.DataArray:
+    """Perform in-place diff of the given channels of the DataArray in the steps dimension.
+
+    Args:
+        da: The DataArray to slice from
+        accum_channels: Channels which are accumulated and need to be differenced
+    """
+    if da.dims[:2] != ("step", "channel"):
+        raise ValueError("This function assumes the first two dimensions are step then channel")
+
+    all_channels = da.channel.values
+    accum_channel_inds = [i for i, c in enumerate(all_channels) if c in accum_channels]
+
+    # Make a copy of the values to avoid changing the underlying numpy array
+    vals = da.values.copy()
+    vals[:-1, accum_channel_inds] = np.diff(vals[:, accum_channel_inds], axis=0)
+    da.values = vals
+
+    return da.isel(step=slice(0, -1))

ocf_data_sampler-0.5.16/ocf_data_sampler/select/select_time_slice.py

@@ -0,0 +1,107 @@
+"""Select a time slice from a Dataset or DataArray."""
+
+import numpy as np
+import pandas as pd
+import xarray as xr
+
+
+def select_time_slice(
+    da: xr.DataArray,
+    t0: pd.Timestamp,
+    interval_start: pd.Timedelta,
+    interval_end: pd.Timedelta,
+    time_resolution: pd.Timedelta,
+) -> xr.DataArray:
+    """Select a time slice from a DataArray.
+
+    Args:
+        da: The DataArray to slice from
+        t0: The init-time
+        interval_start: The start of the interval with respect to t0
+        interval_end: The end of the interval with respect to t0
+        time_resolution: Distance between neighbouring timestamps
+    """
+    start_dt = t0 + interval_start
+    end_dt = t0 + interval_end
+
+    start_dt = start_dt.ceil(time_resolution)
+    end_dt = end_dt.ceil(time_resolution)
+
+    return da.sel(time_utc=slice(start_dt, end_dt))
+
+
+def select_time_slice_nwp(
+    da: xr.DataArray,
+    t0: pd.Timestamp,
+    interval_start: pd.Timedelta,
+    interval_end: pd.Timedelta,
+    time_resolution: pd.Timedelta,
+    dropout_timedeltas: list[pd.Timedelta] | None = None,
+    dropout_frac: float | None = 0,
+) -> xr.DataArray:
+    """Select a time slice from an NWP DataArray.
+
+    Args:
+        da: The DataArray to slice from
+        t0: The init-time
+        interval_start: The start of the interval with respect to t0
+        interval_end: The end of the interval with respect to t0
+        time_resolution: Distance between neighbouring timestamps
+        dropout_timedeltas: List of possible timedeltas before t0 where data availability may start
+        dropout_frac: Probability to apply dropout
+    """
+    # Input checking
+    if dropout_timedeltas is None:
+        dropout_timedeltas = []
+
+    if len(dropout_timedeltas)>0:
+        if not all(t < pd.Timedelta(0) for t in dropout_timedeltas):
+            raise ValueError("dropout timedeltas must be negative")
+        if len(dropout_timedeltas) < 1:
+            raise ValueError("dropout timedeltas must have at least one element")
+
+    if not (0 <= dropout_frac <= 1):
+        raise ValueError("dropout_frac must be between 0 and 1")
+
+    consider_dropout = len(dropout_timedeltas) > 0 and dropout_frac > 0
+
+    start_dt = (t0 + interval_start).ceil(time_resolution)
+    end_dt = (t0 + interval_end).ceil(time_resolution)
+    target_times = pd.date_range(start_dt, end_dt, freq=time_resolution)
+
+    # Potentially apply NWP dropout
+    if consider_dropout and (np.random.uniform() < dropout_frac):
+        t0_available = t0 + np.random.choice(dropout_timedeltas)
+    else:
+        t0_available = t0
+
+    # Get the available and relevant init-times
+    t_min = target_times[0] - da.step.values[-1]
+    init_times = da.init_time_utc.values
+    available_init_times = init_times[(t_min<=init_times) & (init_times<=t0_available)]
+
+    # Find the most recent available init-times for all target-times
+    selected_init_times = np.array(
+        [available_init_times[available_init_times<=t][-1] for t in target_times],
+    )
+
+    # Find the required steps for all target-times
+    steps = target_times - selected_init_times
+
+    # If we are only selecting from one init-time we can construct the slice so its faster
+    if len(np.unique(selected_init_times))==1:
+        da_sel = da.sel(init_time_utc=selected_init_times[0], step=slice(steps[0], steps[-1]))
+
+    # If we are selecting from multiple init times this more complex and slower
+    else:
+        # We want one timestep for each target_time_hourly (obviously!) If we simply do
+        # nwp.sel(init_time=init_times, step=steps) then we'll get the *product* of
+        # init_times and steps, which is not what we want! Instead, we use xarray's
+        # vectorised-indexing mode via using a DataArray indexer.  See the last example here:
+        # https://docs.xarray.dev/en/latest/user-guide/indexing.html#more-advanced-indexing
+        coords = {"step": steps}
+        init_time_indexer = xr.DataArray(selected_init_times, coords=coords)
+        step_indexer = xr.DataArray(steps, coords=coords)
+        da_sel = da.sel(init_time_utc=init_time_indexer, step=step_indexer)
+
+    return da_sel

{ocf_data_sampler-0.5.15 → ocf_data_sampler-0.5.16}/ocf_data_sampler/torch_datasets/datasets/pvnet_uk.py

@@ -22,6 +22,7 @@ from ocf_data_sampler.select import Location, fill_time_periods
 from ocf_data_sampler.torch_datasets.utils import (
     add_alterate_coordinate_projections,
     config_normalization_values_to_dicts,
+    diff_nwp_data,
     fill_nans_in_arrays,
     find_valid_time_periods,
     merge_dicts,
@@ -259,7 +260,7 @@ class PVNetUKRegionalDataset(AbstractPVNetUKDataset):
         sample_dict = slice_datasets_by_space(self.datasets_dict, location, self.config)
         sample_dict = slice_datasets_by_time(sample_dict, t0, self.config)
         sample_dict = tensorstore_compute(sample_dict)
-
+        sample_dict = diff_nwp_data(sample_dict, self.config)
         return self.process_and_combine_datasets(sample_dict, t0, location)
 
     @override
@@ -318,6 +319,7 @@ class PVNetUKConcurrentDataset(AbstractPVNetUKDataset):
         # Slice by time then load to avoid loading the data multiple times from disk
         sample_dict = slice_datasets_by_time(self.datasets_dict, t0, self.config)
         sample_dict = tensorstore_compute(sample_dict)
+        sample_dict = diff_nwp_data(sample_dict, self.config)
 
         gsp_samples = []
 

{ocf_data_sampler-0.5.15 → ocf_data_sampler-0.5.16}/ocf_data_sampler/torch_datasets/datasets/site.py

@@ -27,6 +27,7 @@ from ocf_data_sampler.select import (
 from ocf_data_sampler.torch_datasets.utils import (
     add_alterate_coordinate_projections,
     config_normalization_values_to_dicts,
+    diff_nwp_data,
     fill_nans_in_arrays,
     find_valid_time_periods,
     merge_dicts,
@@ -57,6 +58,7 @@ def get_locations(site_xr: xr.Dataset) -> list[Location]:
 
     return locations
 
+
 def process_and_combine_datasets(
     dataset_dict: dict,
     config: Configuration,
@@ -80,8 +82,6 @@ def process_and_combine_datasets(
 
         for nwp_key, da_nwp in dataset_dict["nwp"].items():
 
-            # Standardise and convert to NumpyBatch
-
             channel_means = means_dict["nwp"][nwp_key]
             channel_stds = stds_dict["nwp"][nwp_key]
 
@@ -276,8 +276,8 @@ class SitesDataset(Dataset):
         """
         sample_dict = slice_datasets_by_space(self.datasets_dict, location, self.config)
         sample_dict = slice_datasets_by_time(sample_dict, t0, self.config)
-
         sample_dict = tensorstore_compute(sample_dict)
+        sample_dict = diff_nwp_data(sample_dict, self.config)
 
         return process_and_combine_datasets(
             sample_dict,
@@ -414,6 +414,7 @@ class SitesDatasetConcurrent(Dataset):
         # slice by time first as we want to keep all site id info
         sample_dict = slice_datasets_by_time(self.datasets_dict, t0, self.config)
         sample_dict = tensorstore_compute(sample_dict)
+        sample_dict = diff_nwp_data(sample_dict, self.config)
 
         site_samples = []
 

{ocf_data_sampler-0.5.15 → ocf_data_sampler-0.5.16}/ocf_data_sampler/torch_datasets/utils/__init__.py

@@ -3,4 +3,5 @@ from .merge_and_fill_utils import fill_nans_in_arrays, merge_dicts
 from .valid_time_periods import find_valid_time_periods
 from .spatial_slice_for_dataset import slice_datasets_by_space
 from .time_slice_for_dataset import slice_datasets_by_time
-from .add_alterate_coordinate_projections import add_alterate_coordinate_projections
+from .add_alterate_coordinate_projections import add_alterate_coordinate_projections
+from .diff_nwp_data import diff_nwp_data

ocf_data_sampler-0.5.16/ocf_data_sampler/torch_datasets/utils/diff_nwp_data.py

@@ -0,0 +1,20 @@
+"""Take the in-place diff of some channels of the NWP data."""
+
+from ocf_data_sampler.config import Configuration
+from ocf_data_sampler.select.diff_channels import diff_channels
+
+
+def diff_nwp_data(dataset_dict: dict, config: Configuration) -> dict:
+    """Take the in-place diff of some channels of the NWP data.
+
+    Args:
+        dataset_dict: Dictionary of xarray datasets
+        config: Configuration object
+    """
+    if "nwp" in dataset_dict:
+        for nwp_key, da_nwp in dataset_dict["nwp"].items():
+            accum_channels = config.input_data.nwp[nwp_key].accum_channels
+            if len(accum_channels)>0:
+                # diff_channels() is an in-place operation and modifies the input
+                dataset_dict["nwp"][nwp_key] = diff_channels(da_nwp, accum_channels)
+    return dataset_dict

{ocf_data_sampler-0.5.15 → ocf_data_sampler-0.5.16}/ocf_data_sampler/torch_datasets/utils/time_slice_for_dataset.py

@@ -28,15 +28,23 @@ def slice_datasets_by_time(
         for nwp_key, da_nwp in datasets_dict["nwp"].items():
             nwp_config = config.input_data.nwp[nwp_key]
 
+            # Add a buffer if we need to diff some of the channels in time
+            if len(nwp_config.accum_channels)>0:
+                interval_end_mins = (
+                    nwp_config.interval_end_minutes
+                    + nwp_config.time_resolution_minutes
+                )
+            else:
+                interval_end_mins = nwp_config.interval_end_minutes
+
             sliced_datasets_dict["nwp"][nwp_key] = select_time_slice_nwp(
                 da_nwp,
                 t0,
                 time_resolution=minutes(nwp_config.time_resolution_minutes),
                 interval_start=minutes(nwp_config.interval_start_minutes),
-                interval_end=minutes(nwp_config.interval_end_minutes),
+                interval_end=minutes(interval_end_mins),
                 dropout_timedeltas=minutes(nwp_config.dropout_timedeltas_minutes),
                 dropout_frac=nwp_config.dropout_fraction,
-                accum_channels=nwp_config.accum_channels,
             )
 
     if "sat" in datasets_dict:

{ocf_data_sampler-0.5.15 → ocf_data_sampler-0.5.16}/ocf_data_sampler.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ocf-data-sampler
-Version: 0.5.15
+Version: 0.5.16
 Author: James Fulton, Peter Dudfield
 Author-email: Open Climate Fix team <info@openclimatefix.org>
 License: MIT License

{ocf_data_sampler-0.5.15 → ocf_data_sampler-0.5.16}/ocf_data_sampler.egg-info/SOURCES.txt

@@ -40,6 +40,7 @@ ocf_data_sampler/numpy_sample/satellite.py
 ocf_data_sampler/numpy_sample/site.py
 ocf_data_sampler/numpy_sample/sun_position.py
 ocf_data_sampler/select/__init__.py
+ocf_data_sampler/select/diff_channels.py
 ocf_data_sampler/select/dropout.py
 ocf_data_sampler/select/fill_time_periods.py
 ocf_data_sampler/select/find_contiguous_time_periods.py
@@ -57,6 +58,7 @@ ocf_data_sampler/torch_datasets/sample/uk_regional.py
 ocf_data_sampler/torch_datasets/utils/__init__.py
 ocf_data_sampler/torch_datasets/utils/add_alterate_coordinate_projections.py
 ocf_data_sampler/torch_datasets/utils/config_normalization_values_to_dicts.py
+ocf_data_sampler/torch_datasets/utils/diff_nwp_data.py
 ocf_data_sampler/torch_datasets/utils/merge_and_fill_utils.py
 ocf_data_sampler/torch_datasets/utils/spatial_slice_for_dataset.py
 ocf_data_sampler/torch_datasets/utils/time_slice_for_dataset.py

ocf_data_sampler-0.5.15/ocf_data_sampler/select/select_time_slice.py

@@ -1,143 +0,0 @@
-"""Select a time slice from a Dataset or DataArray."""
-
-import numpy as np
-import pandas as pd
-import xarray as xr
-
-
-def select_time_slice(
-    da: xr.DataArray,
-    t0: pd.Timestamp,
-    interval_start: pd.Timedelta,
-    interval_end: pd.Timedelta,
-    time_resolution: pd.Timedelta,
-) -> xr.DataArray:
-    """Select a time slice from a DataArray.
-
-    Args:
-        da: The DataArray to slice from
-        t0: The init-time
-        interval_start: The start of the interval with respect to t0
-        interval_end: The end of the interval with respect to t0
-        time_resolution: Distance between neighbouring timestamps
-    """
-    start_dt = t0 + interval_start
-    end_dt = t0 + interval_end
-
-    start_dt = start_dt.ceil(time_resolution)
-    end_dt = end_dt.ceil(time_resolution)
-
-    return da.sel(time_utc=slice(start_dt, end_dt))
-
-
-def select_time_slice_nwp(
-    da: xr.DataArray,
-    t0: pd.Timestamp,
-    interval_start: pd.Timedelta,
-    interval_end: pd.Timedelta,
-    time_resolution: pd.Timedelta,
-    dropout_timedeltas: list[pd.Timedelta] | None = None,
-    dropout_frac: float | None = 0,
-    accum_channels: list[str] | None = None,
-) -> xr.DataArray:
-    """Select a time slice from an NWP DataArray.
-
-    Args:
-        da: The DataArray to slice from
-        t0: The init-time
-        interval_start: The start of the interval with respect to t0
-        interval_end: The end of the interval with respect to t0
-        time_resolution: Distance between neighbouring timestamps
-        dropout_timedeltas: List of possible timedeltas before t0 where data availability may start
-        dropout_frac: Probability to apply dropout
-        accum_channels: Channels which are accumulated and need to be differenced
-    """
-    if accum_channels is None:
-        accum_channels = []
-
-    if dropout_timedeltas is None:
-        dropout_timedeltas = []
-
-    if len(dropout_timedeltas)>0:
-        if not all(t < pd.Timedelta(0) for t in dropout_timedeltas):
-            raise ValueError("dropout timedeltas must be negative")
-        if len(dropout_timedeltas) < 1:
-            raise ValueError("dropout timedeltas must have at least one element")
-
-    if not (0 <= dropout_frac <= 1):
-        raise ValueError("dropout_frac must be between 0 and 1")
-
-    consider_dropout = len(dropout_timedeltas) > 0 and dropout_frac > 0
-
-    # The accumatated and non-accumulated channels
-    accum_channels = np.intersect1d(da.channel.values, accum_channels)
-    non_accum_channels = np.setdiff1d(da.channel.values, accum_channels)
-
-    start_dt = (t0 + interval_start).ceil(time_resolution)
-    end_dt = (t0 + interval_end).ceil(time_resolution)
-    target_times = pd.date_range(start_dt, end_dt, freq=time_resolution)
-
-    # Potentially apply NWP dropout
-    if consider_dropout and (np.random.uniform() < dropout_frac):
-        dt = np.random.choice(dropout_timedeltas)
-        t0_available = t0 + dt
-    else:
-        t0_available = t0
-
-    # Forecasts made up to and including t0
-    available_init_times = da.init_time_utc.sel(init_time_utc=slice(None, t0_available))
-
-    # Find the most recent available init times for all target times
-    selected_init_times = available_init_times.sel(
-        init_time_utc=target_times,
-        method="ffill",  # forward fill from init times to target times
-    ).values
-
-    # Find the required steps for all target times
-    steps = target_times - selected_init_times
-
-    # We want one timestep for each target_time_hourly (obviously!) If we simply do
-    # nwp.sel(init_time=init_times, step=steps) then we'll get the *product* of
-    # init_times and steps, which is not what we want! Instead, we use xarray's
-    # vectorised-indexing mode via using a DataArray indexer.  See the last example here:
-    # https://docs.xarray.dev/en/latest/user-guide/indexing.html#more-advanced-indexing
-
-    coords = {"target_time_utc": target_times}
-    init_time_indexer = xr.DataArray(selected_init_times, coords=coords)
-    step_indexer = xr.DataArray(steps, coords=coords)
-
-    if len(accum_channels) == 0:
-        da_sel = da.sel(step=step_indexer, init_time_utc=init_time_indexer)
-    else:
-        # First minimise the size of the dataset we are diffing
-        # - find the init times we are slicing from
-        unique_init_times = np.unique(selected_init_times)
-        # - find the min and max steps we slice over. Max is extended due to diff
-        min_step = min(steps)
-        max_step = max(steps) + time_resolution
-
-        da_min = da.sel(init_time_utc=unique_init_times, step=slice(min_step, max_step))
-
-        # Slice out the data which does not need to be diffed
-        da_non_accum = da_min.sel(channel=non_accum_channels)
-        da_sel_non_accum = da_non_accum.sel(step=step_indexer, init_time_utc=init_time_indexer)
-
-        # Slice out the channels which need to be diffed
-        da_accum = da_min.sel(channel=accum_channels)
-
-        # Take the diff and slice requested data
-        da_accum = da_accum.diff(dim="step", label="lower")
-        da_sel_accum = da_accum.sel(step=step_indexer, init_time_utc=init_time_indexer)
-
-        # Join diffed and non-diffed variables
-        da_sel = xr.concat([da_sel_non_accum, da_sel_accum], dim="channel")
-
-        # Reorder the variable back to the original order
-        da_sel = da_sel.sel(channel=da.channel.values)
-
-        # Rename the diffed channels
-        da_sel["channel"] = [
-            f"diff_{v}" if v in accum_channels else v for v in da_sel.channel.values
-        ]
-
-    return da_sel