deepsensor 0.3.8__tar.gz → 0.4.0__tar.gz

{deepsensor-0.3.8 → deepsensor-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: deepsensor
-Version: 0.3.8
+Version: 0.4.0
 Summary: A Python package for modelling xarray and pandas data with neural processes.
 Home-page: https://github.com/alan-turing-institute/deepsensor
 Author: Tom R. Andersson
@@ -44,7 +44,7 @@ data with neural processes</p>
 
 -----------
 
-[![release](https://img.shields.io/badge/release-v0.3.8-green?logo=github)](https://github.com/alan-turing-institute/deepsensor/releases)
+[![release](https://img.shields.io/badge/release-v0.4.0-green?logo=github)](https://github.com/alan-turing-institute/deepsensor/releases)
 [![Latest Docs](https://img.shields.io/badge/docs-latest-blue.svg)](https://alan-turing-institute.github.io/deepsensor/)
 ![Tests](https://github.com/alan-turing-institute/deepsensor/actions/workflows/tests.yml/badge.svg)
 [![Coverage Status](https://coveralls.io/repos/github/alan-turing-institute/deepsensor/badge.svg?branch=main)](https://coveralls.io/github/alan-turing-institute/deepsensor?branch=main)
@@ -240,7 +240,7 @@ if you would like to join this list!
 <table>
   <tbody>
     <tr>
-      <td align="center" valign="top" width="14.28%"><a href="https://github.com/acocac"><img src="https://avatars.githubusercontent.com/u/13321552?v=4?s=100" width="100px;" alt="Alejandro ©"/><br /><sub><b>Alejandro ©</b></sub></a><br /><a href="#userTesting-acocac" title="User Testing">📓</a> <a href="#bug-acocac" title="Bug reports">🐛</a> <a href="#mentoring-acocac" title="Mentoring">🧑‍🏫</a> <a href="#ideas-acocac" title="Ideas, Planning, & Feedback">🤔</a> <a href="#research-acocac" title="Research">🔬</a></td>
+      <td align="center" valign="top" width="14.28%"><a href="https://github.com/acocac"><img src="https://avatars.githubusercontent.com/u/13321552?v=4?s=100" width="100px;" alt="Alejandro ©"/><br /><sub><b>Alejandro ©</b></sub></a><br /><a href="#userTesting-acocac" title="User Testing">📓</a> <a href="#bug-acocac" title="Bug reports">🐛</a> <a href="#mentoring-acocac" title="Mentoring">🧑‍🏫</a> <a href="#ideas-acocac" title="Ideas, Planning, & Feedback">🤔</a> <a href="#research-acocac" title="Research">🔬</a> <a href="#code-acocac" title="Code">💻</a> <a href="#test-acocac" title="Tests">⚠️</a></td>
       <td align="center" valign="top" width="14.28%"><a href="https://github.com/annavaughan"><img src="https://avatars.githubusercontent.com/u/45528489?v=4?s=100" width="100px;" alt="Anna Vaughan"/><br /><sub><b>Anna Vaughan</b></sub></a><br /><a href="#research-annavaughan" title="Research">🔬</a></td>
       <td align="center" valign="top" width="14.28%"><a href="http://davidwilby.dev"><img src="https://avatars.githubusercontent.com/u/24752124?v=4?s=100" width="100px;" alt="David Wilby"/><br /><sub><b>David Wilby</b></sub></a><br /><a href="#doc-davidwilby" title="Documentation">📖</a> <a href="#test-davidwilby" title="Tests">⚠️</a> <a href="#maintenance-davidwilby" title="Maintenance">🚧</a></td>
       <td align="center" valign="top" width="14.28%"><a href="http://inconsistentrecords.co.uk"><img src="https://avatars.githubusercontent.com/u/731727?v=4?s=100" width="100px;" alt="Jim Circadian"/><br /><sub><b>Jim Circadian</b></sub></a><br /><a href="#ideas-JimCircadian" title="Ideas, Planning, & Feedback">🤔</a> <a href="#projectManagement-JimCircadian" title="Project Management">📆</a> <a href="#maintenance-JimCircadian" title="Maintenance">🚧</a></td>

{deepsensor-0.3.8 → deepsensor-0.4.0}/README.md

@@ -11,7 +11,7 @@ data with neural processes</p>
 
 -----------
 
-[![release](https://img.shields.io/badge/release-v0.3.8-green?logo=github)](https://github.com/alan-turing-institute/deepsensor/releases)
+[![release](https://img.shields.io/badge/release-v0.4.0-green?logo=github)](https://github.com/alan-turing-institute/deepsensor/releases)
 [![Latest Docs](https://img.shields.io/badge/docs-latest-blue.svg)](https://alan-turing-institute.github.io/deepsensor/)
 ![Tests](https://github.com/alan-turing-institute/deepsensor/actions/workflows/tests.yml/badge.svg)
 [![Coverage Status](https://coveralls.io/repos/github/alan-turing-institute/deepsensor/badge.svg?branch=main)](https://coveralls.io/github/alan-turing-institute/deepsensor?branch=main)
@@ -207,7 +207,7 @@ if you would like to join this list!
 <table>
   <tbody>
     <tr>
-      <td align="center" valign="top" width="14.28%"><a href="https://github.com/acocac"><img src="https://avatars.githubusercontent.com/u/13321552?v=4?s=100" width="100px;" alt="Alejandro ©"/><br /><sub><b>Alejandro ©</b></sub></a><br /><a href="#userTesting-acocac" title="User Testing">📓</a> <a href="#bug-acocac" title="Bug reports">🐛</a> <a href="#mentoring-acocac" title="Mentoring">🧑‍🏫</a> <a href="#ideas-acocac" title="Ideas, Planning, & Feedback">🤔</a> <a href="#research-acocac" title="Research">🔬</a></td>
+      <td align="center" valign="top" width="14.28%"><a href="https://github.com/acocac"><img src="https://avatars.githubusercontent.com/u/13321552?v=4?s=100" width="100px;" alt="Alejandro ©"/><br /><sub><b>Alejandro ©</b></sub></a><br /><a href="#userTesting-acocac" title="User Testing">📓</a> <a href="#bug-acocac" title="Bug reports">🐛</a> <a href="#mentoring-acocac" title="Mentoring">🧑‍🏫</a> <a href="#ideas-acocac" title="Ideas, Planning, & Feedback">🤔</a> <a href="#research-acocac" title="Research">🔬</a> <a href="#code-acocac" title="Code">💻</a> <a href="#test-acocac" title="Tests">⚠️</a></td>
       <td align="center" valign="top" width="14.28%"><a href="https://github.com/annavaughan"><img src="https://avatars.githubusercontent.com/u/45528489?v=4?s=100" width="100px;" alt="Anna Vaughan"/><br /><sub><b>Anna Vaughan</b></sub></a><br /><a href="#research-annavaughan" title="Research">🔬</a></td>
       <td align="center" valign="top" width="14.28%"><a href="http://davidwilby.dev"><img src="https://avatars.githubusercontent.com/u/24752124?v=4?s=100" width="100px;" alt="David Wilby"/><br /><sub><b>David Wilby</b></sub></a><br /><a href="#doc-davidwilby" title="Documentation">📖</a> <a href="#test-davidwilby" title="Tests">⚠️</a> <a href="#maintenance-davidwilby" title="Maintenance">🚧</a></td>
       <td align="center" valign="top" width="14.28%"><a href="http://inconsistentrecords.co.uk"><img src="https://avatars.githubusercontent.com/u/731727?v=4?s=100" width="100px;" alt="Jim Circadian"/><br /><sub><b>Jim Circadian</b></sub></a><br /><a href="#ideas-JimCircadian" title="Ideas, Planning, & Feedback">🤔</a> <a href="#projectManagement-JimCircadian" title="Project Management">📆</a> <a href="#maintenance-JimCircadian" title="Maintenance">🚧</a></td>

{deepsensor-0.3.8 → deepsensor-0.4.0}/deepsensor/data/loader.py

@@ -678,11 +678,11 @@ class TaskLoader:
         seed: Optional[int] = None,
     ) -> (np.ndarray, np.ndarray):
         """
-        Sample a DataArray according to a given strategy.
+        Sample a DataFrame according to a given strategy.
 
         Args:
             df (:class:`pandas.DataFrame` | :class:`pandas.Series`):
-                DataArray to sample, assumed to be time-sliced for the task
+                Dataframe to sample, assumed to be time-sliced for the task
                 already.
             sampling_strat (str | int | float | :class:`numpy:numpy.ndarray`):
                 Sampling strategy, either "all" or an integer for random grid
@@ -720,20 +720,24 @@ class TaskLoader:
             X_c = df.reset_index()[["x1", "x2"]].values.T.astype(self.dtype)
             Y_c = df.values.T
         elif isinstance(sampling_strat, np.ndarray):
+            if df.index.get_level_values("x1").dtype != sampling_strat.dtype:
+                raise InvalidSamplingStrategyError(
+                    "Passed a numpy coordinate array to sample pandas DataFrame, "
+                    "but the coordinate array has a different dtype than the DataFrame. "
+                    f"Got {sampling_strat.dtype} but expected {df.index.get_level_values('x1').dtype}."
+                )
             X_c = sampling_strat.astype(self.dtype)
-            x1match = np.in1d(df.index.get_level_values("x1"), X_c[0])
-            x2match = np.in1d(df.index.get_level_values("x2"), X_c[1])
-            num_matches = np.sum(x1match & x2match)
-
-            # Check that we got all the samples we asked for
-            if num_matches != X_c.shape[1]:
+            try:
+                Y_c = df.loc[pd.IndexSlice[:, X_c[0], X_c[1]]].values.T
+            except KeyError:
                 raise InvalidSamplingStrategyError(
-                    f"Passed a numpy coordinate array to sample pandas DataFrame, "
-                    f"but the DataFrame did not contain all the requested samples. "
-                    f"Requested {X_c.shape[1]} samples but only got {num_matches}."
+                    "Passed a numpy coordinate array to sample pandas DataFrame, "
+                    "but the DataFrame did not contain all the requested samples.\n"
+                    f"Indexes: {df.index}\n"
+                    f"Sampling coords: {X_c}\n"
+                    "If this is unexpected, check that your numpy sampling array matches "
+                    "the DataFrame index values *exactly*."
                 )
-
-            Y_c = df[x1match & x2match].values.T
         else:
             raise InvalidSamplingStrategyError(
                 f"Unknown sampling strategy {sampling_strat}"

{deepsensor-0.3.8 → deepsensor-0.4.0}/deepsensor/data/sources.py

@@ -277,7 +277,7 @@ def get_era5_reanalysis_data(
     if num_processes == 1:
         # Just download in one go
         if verbose:
-            print("Downloading ERA5 data in without parallelisation... ")
+            print("Downloading ERA5 data without parallelisation... ")
         era5_da = _get_era5_reanalysis_data_parallel(
             date_range=date_range,
             var_IDs=var_IDs,
@@ -432,7 +432,7 @@ def get_gldas_land_mask(
         with urllib.request.urlopen(req) as response:
             with open(fname, "wb") as f:
                 f.write(response.read())
-        da = xr.open_dataset(fname)["GLDAS_mask"].isel(time=0).drop("time").load()
+        da = xr.open_dataset(fname)["GLDAS_mask"].isel(time=0).drop_vars("time").load()
 
         if isinstance(extent, str):
             extent = extent_str_to_tuple(extent)
@@ -577,7 +577,7 @@ def get_earthenv_auxiliary_data(
             # Read data
             da = xr.open_dataset(fname).to_array().squeeze().load()
             da = da.rename({"y": "lat", "x": "lon"})
-            da = da.drop(["band", "spatial_ref", "variable"])
+            da = da.drop_vars(["band", "spatial_ref", "variable"])
             da.name = var_ID
             da = da.sel(lat=slice(lat_max, lat_min), lon=slice(lon_min, lon_max))
             da_dict[var_ID] = da

deepsensor-0.4.0/deepsensor/eval/__init__.py

@@ -0,0 +1 @@
+from .metrics import *

deepsensor-0.4.0/deepsensor/eval/metrics.py

@@ -0,0 +1,24 @@
+import xarray as xr
+from deepsensor.model.pred import Prediction
+
+
+def compute_errors(pred: Prediction, target: xr.Dataset) -> xr.Dataset:
+    """
+    Compute errors between predictions and targets.
+
+    Args:
+        pred: Prediction object.
+        target: Target data.
+
+    Returns:
+        xr.Dataset: Dataset of pointwise differences between predictions and targets
+        at the same valid time in the predictions. Note, the difference is positive
+        when the prediction is greater than the target.
+    """
+    errors = {}
+    for var_ID, pred_var in pred.items():
+        target_var = target[var_ID]
+        error = pred_var["mean"] - target_var.sel(time=pred_var.time)
+        error.name = f"{var_ID}"
+        errors[var_ID] = error
+    return xr.Dataset(errors)

{deepsensor-0.3.8 → deepsensor-0.4.0}/deepsensor/model/model.py

@@ -348,6 +348,34 @@ class DeepSensorModel(ProbabilisticModel):
         if ar_sample and n_samples < 1:
             raise ValueError("Must pass `n_samples` > 0 to use `ar_sample`.")
 
+        target_delta_t = self.task_loader.target_delta_t
+        dts = [pd.Timedelta(dt) for dt in target_delta_t]
+        dts_all_zero = all([dt == pd.Timedelta(seconds=0) for dt in dts])
+        if target_delta_t is not None and dts_all_zero:
+            forecasting_mode = False
+            lead_times = None
+        elif target_delta_t is not None and not dts_all_zero:
+            target_var_IDs_set = set(self.task_loader.target_var_IDs)
+            msg = f"""
+            Got more than one set of target variables in target sets,
+            but predictions can only be made with one set of target variables
+            to simplify implementation.
+            Got {target_var_IDs_set}.
+            """
+            assert len(target_var_IDs_set) == 1, msg
+            # Repeat lead_tim for each variable in each target set
+            lead_times = []
+            for target_set_idx, dt in enumerate(target_delta_t):
+                target_set_dim = self.task_loader.target_dims[target_set_idx]
+                lead_times += [
+                    pd.Timedelta(dt, unit=self.task_loader.time_freq)
+                    for _ in range(target_set_dim)
+                ]
+            forecasting_mode = True
+        else:
+            forecasting_mode = False
+            lead_times = None
+
         if type(tasks) is Task:
             tasks = [tasks]
 
@@ -355,12 +383,14 @@ class DeepSensorModel(ProbabilisticModel):
             B.set_random_seed(seed)
             np.random.seed(seed)
 
-        dates = [task["time"] for task in tasks]
+        init_dates = [task["time"] for task in tasks]
 
         # Flatten tuple of tuples to single list
         target_var_IDs = [
             var_ID for set in self.task_loader.target_var_IDs for var_ID in set
         ]
+        if lead_times is not None:
+            assert len(lead_times) == len(target_var_IDs)
 
         # TODO consider removing this logic, can we just depend on the dim names in X_t?
         if not unnormalise:
@@ -385,7 +415,7 @@ class DeepSensorModel(ProbabilisticModel):
         elif isinstance(X_t, (xr.DataArray, xr.Dataset)):
             # Remove time dimension if present
             if "time" in X_t.coords:
-                X_t = X_t.isel(time=0).drop("time")
+                X_t = X_t.isel(time=0).drop_vars("time")
 
         if mode == "off-grid" and append_indexes is not None:
             # Check append_indexes are all same length as X_t
@@ -450,11 +480,13 @@ class DeepSensorModel(ProbabilisticModel):
         pred = Prediction(
             target_var_IDs,
             pred_params_to_store,
-            dates,
+            init_dates,
             X_t,
             X_t_mask,
             coord_names,
             n_samples=n_samples,
+            forecasting_mode=forecasting_mode,
+            lead_times=lead_times,
         )
 
         def unnormalise_pred_array(arr, **kwargs):
@@ -605,14 +637,22 @@ class DeepSensorModel(ProbabilisticModel):
             # Assign predictions to Prediction object
             for param, arr in prediction_arrs.items():
                 if param != "mixture_probs":
-                    pred.assign(param, task["time"], arr)
+                    pred.assign(param, task["time"], arr, lead_times=lead_times)
                 elif param == "mixture_probs":
                     assert arr.shape[0] == self.N_mixture_components, (
                         f"Number of mixture components ({arr.shape[0]}) does not match "
                         f"model attribute N_mixture_components ({self.N_mixture_components})."
                     )
                     for component_i, probs in enumerate(arr):
-                        pred.assign(f"{param}_{component_i}", task["time"], probs)
+                        pred.assign(
+                            f"{param}_{component_i}",
+                            task["time"],
+                            probs,
+                            lead_times=lead_times,
+                        )
+
+        if forecasting_mode:
+            pred = add_valid_time_coord_to_pred(pred)
 
         if verbose:
             dur = time.time() - tic
@@ -621,6 +661,35 @@ class DeepSensorModel(ProbabilisticModel):
         return pred
 
 
+def add_valid_time_coord_to_pred(pred: Prediction) -> Prediction:
+    """
+    Add a valid time coordinate "time" to a Prediction object based on the
+    initialisation times "init_time" and lead times "lead_time".
+
+    Args:
+        pred (:class:`~.model.pred.Prediction`):
+            Prediction object to add valid time coordinate to.
+
+    Returns:
+        :class:`~.model.pred.Prediction`:
+            Prediction object with valid time coordinate added.
+    """
+    for var_ID in pred.keys():
+        if isinstance(pred[var_ID], pd.DataFrame):
+            x = pred[var_ID].reset_index()
+            pred[var_ID]["time"] = (x["lead_time"] + x["init_time"]).values
+            print(f"{x}")
+            print(f"{x.dtypes}")
+        elif isinstance(pred[var_ID], xr.Dataset):
+            x = pred[var_ID]
+            pred[var_ID] = pred[var_ID].assign_coords(
+                time=x["lead_time"] + x["init_time"]
+            )
+        else:
+            raise ValueError(f"Unsupported prediction type {type(pred[var_ID])}.")
+    return pred
+
+
 def main():  # pragma: no cover
     import deepsensor.tensorflow
     from deepsensor.data.loader import TaskLoader

{deepsensor-0.3.8 → deepsensor-0.4.0}/deepsensor/model/pred.py

@@ -4,6 +4,8 @@ import numpy as np
 import pandas as pd
 import xarray as xr
 
+Timestamp = Union[str, pd.Timestamp, np.datetime64]
+
 
 class Prediction(dict):
     """
@@ -32,13 +34,20 @@ class Prediction(dict):
         n_samples (int)
             Number of joint samples to draw from the model. If 0, will not
             draw samples. Default 0.
+        forecasting_mode (bool)
+            If True, stored forecast predictions with an init_time and lead_time dimension,
+            and a valid_time coordinate. If False, stores prediction at t=0 only
+            (i.e. spatial interpolation), with only a single time dimension. Default False.
+        lead_times (List[pd.Timedelta], optional)
+            List of lead times to store in predictions. Must be provided if
+            forecasting_mode is True. Default None.
     """
 
     def __init__(
         self,
         target_var_IDs: List[str],
         pred_params: List[str],
-        dates: List[Union[str, pd.Timestamp]],
+        dates: List[Timestamp],
         X_t: Union[
             xr.Dataset,
             xr.DataArray,
@@ -50,6 +59,8 @@ class Prediction(dict):
         X_t_mask: Optional[Union[xr.Dataset, xr.DataArray]] = None,
         coord_names: dict = None,
         n_samples: int = 0,
+        forecasting_mode: bool = False,
+        lead_times: Optional[List[pd.Timedelta]] = None,
     ):
         self.target_var_IDs = target_var_IDs
         self.X_t_mask = X_t_mask
@@ -58,6 +69,13 @@ class Prediction(dict):
         self.x1_name = coord_names["x1"]
         self.x2_name = coord_names["x2"]
 
+        self.forecasting_mode = forecasting_mode
+        if forecasting_mode:
+            assert (
+                lead_times is not None
+            ), "If forecasting_mode is True, lead_times must be provided."
+        self.lead_times = lead_times
+
         self.mode = infer_prediction_modality_from_X_t(X_t)
 
         self.pred_params = pred_params
@@ -67,15 +85,25 @@ class Prediction(dict):
                 *[f"sample_{i}" for i in range(n_samples)],
             ]
 
+        # Create empty xarray/pandas objects to store predictions
         if self.mode == "on-grid":
             for var_ID in self.target_var_IDs:
-                # Create empty xarray/pandas objects to store predictions
+                if self.forecasting_mode:
+                    prepend_dims = ["lead_time"]
+                    prepend_coords = {"lead_time": lead_times}
+                else:
+                    prepend_dims = None
+                    prepend_coords = None
                 self[var_ID] = create_empty_spatiotemporal_xarray(
                     X_t,
                     dates,
                     data_vars=self.pred_params,
                     coord_names=coord_names,
+                    prepend_dims=prepend_dims,
+                    prepend_coords=prepend_coords,
                 )
+                if self.forecasting_mode:
+                    self[var_ID] = self[var_ID].rename(time="init_time")
             if self.X_t_mask is None:
                 # Create 2D boolean array of True values to simplify indexing
                 self.X_t_mask = (
@@ -86,8 +114,18 @@ class Prediction(dict):
                 )
         elif self.mode == "off-grid":
             # Repeat target locs for each date to create multiindex
-            idxs = [(date, *idxs) for date in dates for idxs in X_t.index]
-            index = pd.MultiIndex.from_tuples(idxs, names=["time", *X_t.index.names])
+            if self.forecasting_mode:
+                index_names = ["lead_time", "init_time", *X_t.index.names]
+                idxs = [
+                    (lt, date, *idxs)
+                    for lt in lead_times
+                    for date in dates
+                    for idxs in X_t.index
+                ]
+            else:
+                index_names = ["time", *X_t.index.names]
+                idxs = [(date, *idxs) for date in dates for idxs in X_t.index]
+            index = pd.MultiIndex.from_tuples(idxs, names=index_names)
             for var_ID in self.target_var_IDs:
                 self[var_ID] = pd.DataFrame(index=index, columns=self.pred_params)
 
@@ -106,6 +144,7 @@ class Prediction(dict):
         prediction_parameter: str,
         date: Union[str, pd.Timestamp],
         data: np.ndarray,
+        lead_times: Optional[List[pd.Timedelta]] = None,
     ):
         """
 
@@ -117,11 +156,29 @@ class Prediction(dict):
             data (np.ndarray)
                 If off-grid: Shape (N_var, N_targets) or (N_samples, N_var, N_targets).
                 If on-grid: Shape (N_var, N_x1, N_x2) or (N_samples, N_var, N_x1, N_x2).
+            lead_time (pd.Timedelta, optional)
+                Lead time of the forecast. Required if forecasting_mode is True. Default None.
         """
+        if self.forecasting_mode:
+            assert (
+                lead_times is not None
+            ), "If forecasting_mode is True, lead_times must be provided."
+
+            msg = f"""
+            If forecasting_mode is True, lead_times must be of equal length to the number of
+            variables in the data (the first dimension). Got {lead_times=} of length
+            {len(lead_times)} lead times and data shape {data.shape}.
+            """
+            assert len(lead_times) == data.shape[0], msg
+
         if self.mode == "on-grid":
             if prediction_parameter != "samples":
-                for var_ID, pred in zip(self.target_var_IDs, data):
-                    self[var_ID][prediction_parameter].loc[date].data[
+                for i, (var_ID, pred) in enumerate(zip(self.target_var_IDs, data)):
+                    if self.forecasting_mode:
+                        index = (lead_times[i], date)
+                    else:
+                        index = date
+                    self[var_ID][prediction_parameter].loc[index].data[
                         self.X_t_mask.data
                     ] = pred.ravel()
             elif prediction_parameter == "samples":
@@ -130,28 +187,44 @@ class Prediction(dict):
                     f"have shape (N_samples, N_var, N_x1, N_x2). Got {data.shape}."
                 )
                 for sample_i, sample in enumerate(data):
-                    for var_ID, pred in zip(self.target_var_IDs, sample):
-                        self[var_ID][f"sample_{sample_i}"].loc[date].data[
+                    for i, (var_ID, pred) in enumerate(
+                        zip(self.target_var_IDs, sample)
+                    ):
+                        if self.forecasting_mode:
+                            index = (lead_times[i], date)
+                        else:
+                            index = date
+                        self[var_ID][f"sample_{sample_i}"].loc[index].data[
                             self.X_t_mask.data
                         ] = pred.ravel()
 
         elif self.mode == "off-grid":
             if prediction_parameter != "samples":
-                for var_ID, pred in zip(self.target_var_IDs, data):
-                    self[var_ID][prediction_parameter].loc[date] = pred
+                for i, (var_ID, pred) in enumerate(zip(self.target_var_IDs, data)):
+                    if self.forecasting_mode:
+                        index = (lead_times[i], date)
+                    else:
+                        index = date
+                    self[var_ID][prediction_parameter].loc[index] = pred
             elif prediction_parameter == "samples":
                 assert len(data.shape) == 3, (
                     f"If prediction_parameter is 'samples', and mode is 'off-grid', data must"
                     f"have shape (N_samples, N_var, N_targets). Got {data.shape}."
                 )
                 for sample_i, sample in enumerate(data):
-                    for var_ID, pred in zip(self.target_var_IDs, sample):
-                        self[var_ID][f"sample_{sample_i}"].loc[date] = pred
+                    for i, (var_ID, pred) in enumerate(
+                        zip(self.target_var_IDs, sample)
+                    ):
+                        if self.forecasting_mode:
+                            index = (lead_times[i], date)
+                        else:
+                            index = date
+                        self[var_ID][f"sample_{sample_i}"].loc[index] = pred
 
 
 def create_empty_spatiotemporal_xarray(
     X: Union[xr.Dataset, xr.DataArray],
-    dates: List,
+    dates: List[Timestamp],
     coord_names: dict = None,
     data_vars: List[str] = None,
     prepend_dims: Optional[List[str]] = None,
@@ -231,10 +304,6 @@ def create_empty_spatiotemporal_xarray(
     # Convert time coord to pandas timestamps
     pred_ds = pred_ds.assign_coords(time=pd.to_datetime(pred_ds.time.values))
 
-    # TODO: Convert init time to forecast time?
-    # pred_ds = pred_ds.assign_coords(
-    #     time=pred_ds['time'] + pd.Timedelta(days=task_loader.target_delta_t[0]))
-
     return pred_ds
 
 

{deepsensor-0.3.8 → deepsensor-0.4.0}/deepsensor.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: deepsensor
-Version: 0.3.8
+Version: 0.4.0
 Summary: A Python package for modelling xarray and pandas data with neural processes.
 Home-page: https://github.com/alan-turing-institute/deepsensor
 Author: Tom R. Andersson
@@ -44,7 +44,7 @@ data with neural processes</p>
 
 -----------
 
-[![release](https://img.shields.io/badge/release-v0.3.8-green?logo=github)](https://github.com/alan-turing-institute/deepsensor/releases)
+[![release](https://img.shields.io/badge/release-v0.4.0-green?logo=github)](https://github.com/alan-turing-institute/deepsensor/releases)
 [![Latest Docs](https://img.shields.io/badge/docs-latest-blue.svg)](https://alan-turing-institute.github.io/deepsensor/)
 ![Tests](https://github.com/alan-turing-institute/deepsensor/actions/workflows/tests.yml/badge.svg)
 [![Coverage Status](https://coveralls.io/repos/github/alan-turing-institute/deepsensor/badge.svg?branch=main)](https://coveralls.io/github/alan-turing-institute/deepsensor?branch=main)
@@ -240,7 +240,7 @@ if you would like to join this list!
 <table>
   <tbody>
     <tr>
-      <td align="center" valign="top" width="14.28%"><a href="https://github.com/acocac"><img src="https://avatars.githubusercontent.com/u/13321552?v=4?s=100" width="100px;" alt="Alejandro ©"/><br /><sub><b>Alejandro ©</b></sub></a><br /><a href="#userTesting-acocac" title="User Testing">📓</a> <a href="#bug-acocac" title="Bug reports">🐛</a> <a href="#mentoring-acocac" title="Mentoring">🧑‍🏫</a> <a href="#ideas-acocac" title="Ideas, Planning, & Feedback">🤔</a> <a href="#research-acocac" title="Research">🔬</a></td>
+      <td align="center" valign="top" width="14.28%"><a href="https://github.com/acocac"><img src="https://avatars.githubusercontent.com/u/13321552?v=4?s=100" width="100px;" alt="Alejandro ©"/><br /><sub><b>Alejandro ©</b></sub></a><br /><a href="#userTesting-acocac" title="User Testing">📓</a> <a href="#bug-acocac" title="Bug reports">🐛</a> <a href="#mentoring-acocac" title="Mentoring">🧑‍🏫</a> <a href="#ideas-acocac" title="Ideas, Planning, & Feedback">🤔</a> <a href="#research-acocac" title="Research">🔬</a> <a href="#code-acocac" title="Code">💻</a> <a href="#test-acocac" title="Tests">⚠️</a></td>
       <td align="center" valign="top" width="14.28%"><a href="https://github.com/annavaughan"><img src="https://avatars.githubusercontent.com/u/45528489?v=4?s=100" width="100px;" alt="Anna Vaughan"/><br /><sub><b>Anna Vaughan</b></sub></a><br /><a href="#research-annavaughan" title="Research">🔬</a></td>
       <td align="center" valign="top" width="14.28%"><a href="http://davidwilby.dev"><img src="https://avatars.githubusercontent.com/u/24752124?v=4?s=100" width="100px;" alt="David Wilby"/><br /><sub><b>David Wilby</b></sub></a><br /><a href="#doc-davidwilby" title="Documentation">📖</a> <a href="#test-davidwilby" title="Tests">⚠️</a> <a href="#maintenance-davidwilby" title="Maintenance">🚧</a></td>
       <td align="center" valign="top" width="14.28%"><a href="http://inconsistentrecords.co.uk"><img src="https://avatars.githubusercontent.com/u/731727?v=4?s=100" width="100px;" alt="Jim Circadian"/><br /><sub><b>Jim Circadian</b></sub></a><br /><a href="#ideas-JimCircadian" title="Ideas, Planning, & Feedback">🤔</a> <a href="#projectManagement-JimCircadian" title="Project Management">📆</a> <a href="#maintenance-JimCircadian" title="Maintenance">🚧</a></td>

{deepsensor-0.3.8 → deepsensor-0.4.0}/deepsensor.egg-info/SOURCES.txt

@@ -22,6 +22,8 @@ deepsensor/data/processor.py
 deepsensor/data/sources.py
 deepsensor/data/task.py
 deepsensor/data/utils.py
+deepsensor/eval/__init__.py
+deepsensor/eval/metrics.py
 deepsensor/model/__init__.py
 deepsensor/model/convnp.py
 deepsensor/model/defaults.py

{deepsensor-0.3.8 → deepsensor-0.4.0}/setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = deepsensor
-version = 0.3.8
+version = 0.4.0
 author = Tom R. Andersson
 author_email = tomandersson3@gmail.com
 description = A Python package for modelling xarray and pandas data with neural processes.

{deepsensor-0.3.8 → deepsensor-0.4.0}/tests/test_model.py

@@ -18,6 +18,7 @@ from deepsensor.data.processor import DataProcessor
 from deepsensor.data.loader import TaskLoader
 from deepsensor.model.convnp import ConvNP
 from deepsensor.train.train import Trainer
+from deepsensor.eval.metrics import compute_errors
 
 from tests.utils import gen_random_data_xr, gen_random_data_pandas
 
@@ -55,8 +56,15 @@ class TestModel(unittest.TestCase):
     def setUpClass(cls):
         # super().__init__(*args, **kwargs)
         # It's safe to share data between tests because the TaskLoader does not modify data
+        cls.var_ID = "2m_temp"
         cls.da = _gen_data_xr()
+        cls.da.name = cls.var_ID
         cls.df = _gen_data_pandas()
+        cls.df.name = cls.var_ID
+        # Various tests assume we have a single target set with a single variable.
+        # If a test requires multiple target sets or variables, this is set up in the test.
+        assert isinstance(cls.da, xr.DataArray)
+        assert isinstance(cls.df, pd.Series)
 
         cls.dp = DataProcessor()
         _ = cls.dp([cls.da, cls.df])  # Compute normalisation parameters
@@ -417,10 +425,10 @@ class TestModel(unittest.TestCase):
         task = tl("2020-01-01")
         pred = model.predict(task, X_t=da_raw)
 
-        assert np.array_equal(
+        np.testing.assert_array_equal(
             pred["dummy_data"]["mean"]["latitude"], da_raw["latitude"]
         )
-        assert np.array_equal(
+        np.testing.assert_array_equal(
             pred["dummy_data"]["mean"]["longitude"], da_raw["longitude"]
         )
 
@@ -493,14 +501,14 @@ class TestModel(unittest.TestCase):
             # Check that nothing breaks and the correct parameters are returned
             pred = model.predict(task, X_t=X_t, pred_params=pred_params)
             for pred_param in pred_params:
-                assert pred_param in pred["var"]
+                assert pred_param in pred[self.var_ID]
 
             # Test mixture probs special case
             pred_params = ["mixture_probs"]
             pred = model.predict(task, X_t=self.da, pred_params=pred_params)
             for component in range(model.N_mixture_components):
                 pred_param = f"mixture_probs_{component}"
-                assert pred_param in pred["var"]
+                assert pred_param in pred[self.var_ID]
 
     def test_highlevel_predict_with_pred_params_xarray(self):
         """
@@ -528,14 +536,14 @@ class TestModel(unittest.TestCase):
             # Check that nothing breaks and the correct parameters are returned
             pred = model.predict(task, X_t=self.da, pred_params=pred_params)
             for pred_param in pred_params:
-                assert pred_param in pred["var"]
+                assert pred_param in pred[self.var_ID]
 
             # Test mixture probs special case
             pred_params = ["mixture_probs"]
             pred = model.predict(task, X_t=self.da, pred_params=pred_params)
             for component in range(model.N_mixture_components):
                 pred_param = f"mixture_probs_{component}"
-                assert pred_param in pred["var"]
+                assert pred_param in pred[self.var_ID]
 
     def test_highlevel_predict_with_invalid_pred_params(self):
         """Test that passing ``pred_params`` to ``.predict`` works."""
@@ -640,6 +648,66 @@ class TestModel(unittest.TestCase):
                 ar_sample=True,
             )
 
+    def test_forecasting_model_predict_return_valid_times(self):
+        """Test that the times returned by a forecasting model are valid."""
+        init_dates = ["2020-01-01", "2020-01-02"]
+        expected_init_times = np.array(init_dates).astype(np.datetime64)
+
+        lead_times_days = [1, 2, 3]
+        expected_lead_times = np.array(
+            [np.timedelta64(lt, "D") for lt in lead_times_days]
+        )
+
+        expected_valid_times = np.array(
+            expected_lead_times[:, None] + expected_init_times[None, :]
+        )
+
+        tl = TaskLoader(
+            context=self.da,
+            target=[
+                self.da,
+            ]
+            * len(lead_times_days),
+            target_delta_t=lead_times_days,
+            time_freq="D",
+        )
+        model = ConvNP(self.dp, tl, unet_channels=(5, 5, 5), verbose=False)
+        tasks = tl(init_dates, context_sampling=10)
+
+        X_ts = [
+            # Gridded predictions (xarray)
+            self.da,
+            # Off-grid prediction (pandas)
+            np.array([[0.0, 0.5, 1.0], [0.0, 0.5, 1.0]]),
+        ]
+        for X_t in X_ts:
+            pred = model.predict(tasks, X_t=X_t)
+
+            pred_var = pred[self.var_ID]
+
+            if isinstance(pred_var, xr.Dataset):
+                # Check we can compute errors using the valid time coord ('time')
+                errors = compute_errors(pred, self.da.to_dataset())
+                for var_ID in errors.keys():
+                    assert tuple(errors[var_ID].dims) == (
+                        "lead_time",
+                        "init_time",
+                        "x1",
+                        "x2",
+                    )
+                    assert errors[var_ID].shape == pred[var_ID]["mean"].shape
+            elif isinstance(pred_var, pd.DataFrame):
+                # Makes coordinate checking easier by avoiding repeat values
+                pred_var = pred_var.to_xarray().isel(x1=0, x2=0)
+
+            np.testing.assert_array_equal(
+                pred_var.lead_time.values, expected_lead_times
+            )
+            np.testing.assert_array_equal(
+                pred_var.init_time.values, expected_init_times
+            )
+            np.testing.assert_array_equal(pred_var.time.values, expected_valid_times)
+
 
 def assert_shape(x, shape: tuple):
     """Assert that the shape of ``x`` matches ``shape``."""

{deepsensor-0.3.8 → deepsensor-0.4.0}/tests/test_task_loader.py

@@ -192,6 +192,23 @@ class TestTaskLoader(unittest.TestCase):
                 with self.assertRaises(InvalidSamplingStrategyError):
                     task = tl("2020-01-01", invalid_sampling_strategy)
 
+    def test_different_dtype_when_sampling_offgrid_data_at_specific_numpy_locs(self):
+        """Test different dtype when sampling off-grid data at specific numpy locations."""
+        sampling_strat = np.array(
+            [np.linspace(0, 1, 10), np.linspace(0, 1, 10)], dtype=np.float16
+        )
+
+        tl = TaskLoader(
+            context=self.df,
+            target=self.df,
+        )
+
+        assert sampling_strat.dtype != tl.context[0].index.get_level_values("x1").dtype
+        assert sampling_strat.dtype != tl.context[0].index.get_level_values("x2").dtype
+
+        with self.assertRaises(InvalidSamplingStrategyError):
+            task = tl("2020-01-01", sampling_strat, sampling_strat)
+
     def test_wrong_links(self):
         """Test link indexes out of range."""
         with self.assertRaises(ValueError):