def _load_combined_attributes(self):
"""This function loads data set specific attributes and combines them with hydroatlas attributes"""
dfs = []
# load dataset specific attributes from the subclass
df = self._load_attributes()
if df is not None:
# in case of training (not finetuning) check for NaNs in feature std.
if self._compute_scaler:
utils.attributes_sanity_check(df=df)
dfs.append(df)
# Hydroatlas attributes can be used everywhere
if self.cfg.hydroatlas_attributes:
dfs.append(self._load_hydroatlas_attributes())
if dfs:
# combine all attributes into a single dataframe
df = pd.concat(dfs, axis=1)
# check if any attribute specified in the config is not available in the dataframes
combined_attributes = self.cfg.camels_attributes + self.cfg.hydroatlas_attributes
missing_columns = [
attr for attr in combined_attributes if attr not in df.columns
]
if missing_columns:
raise ValueError(
f"The following attributes are not available in the dataset: {missing_columns}"
)
# fix the order of the columns to be alphabetically
df = df.sort_index(axis=1)
# calculate statistics and normalize features
if self._compute_scaler:
self.scaler["attribute_means"] = df.mean()
self.scaler["attribute_stds"] = df.std()
if any([k.startswith("camels_attr") for k in self.scaler.keys()]):
LOGGER.warning(
"Deprecation warning: Using old scaler files won't be supported in the upcoming release."
)
# Here we assume that only camels attributes are used
df = (df - self.scaler['camels_attr_means']
) / self.scaler["camels_attr_stds"]
else:
df = (df - self.scaler['attribute_means']
) / self.scaler["attribute_stds"]
# preprocess each basin feature vector as pytorch tensor
for basin in self.basins:
attributes = df.loc[df.index == basin].values.flatten()
self.attributes[basin] = torch.from_numpy(
attributes.astype(np.float32))
def _load_hydroatlas_attributes(self):
df = utils.load_hydroatlas_attributes(self.cfg.data_dir, basins=self.basins)
# remove all attributes not defined in the config
drop_cols = [c for c in df.columns if c not in self.cfg.hydroatlas_attributes]
df = df.drop(drop_cols, axis=1)
if self.is_train:
# sanity check attributes for NaN in per-feature standard deviation
utils.attributes_sanity_check(df=df)
return df
