def evaluate_basin(model: nn.Module,
loader: DataLoader) -> Tuple[np.ndarray, np.ndarray]:
"""Evaluate model on a single basin
Parameters
----------
model : nn.Module
The PyTorch model to train
loader : DataLoader
PyTorch DataLoader containing the basin data in batches.
Returns
-------
preds : np.ndarray
Array containing the (rescaled) network prediction for the entire data period
obs : np.ndarray
Array containing the observed discharge for the entire data period
"""
model.eval()
preds, obs = None, None
with torch.no_grad():
for data in loader:
if len(data) == 2:
x, y = data
x, y = x.to(DEVICE), y.to(DEVICE)
p = model(x)[0]
elif len(data) == 3:
x_d, x_s, y = data
x_d, x_s, y = x_d.to(DEVICE), x_s.to(DEVICE), y.to(DEVICE)
p = model(x_d, x_s[:, 0, :])[0]
if preds is None:
preds = p.detach().cpu()
obs = y.detach().cpu()
else:
preds = torch.cat((preds, p.detach().cpu()), 0)
obs = torch.cat((obs, y.detach().cpu()), 0)
preds = rescale_features(preds.numpy(), variable='output')
obs = obs.numpy()
# set discharges < 0 to zero
preds[preds < 0] = 0
return preds, obs
def eval_with_added_noise(model: torch.nn.Module, loader: DataLoader,
noise: torch.Tensor) -> float:
"""Evaluate model on a single basin with added noise
Parameters
----------
model : nn.Module
The PyTorch model to train
loader : DataLoader
PyTorch DataLoader containing the basin data in batches.
noise : torch.Tensor
Tensor containing the noise for this evaluation run.
Returns
-------
float
Nash-Sutcliff-Efficiency of the simulations with added noise.
"""
model.eval()
preds, obs = None, None
with torch.no_grad():
for x_d, x_s, y in loader:
x_d, x_s, y = x_d.to(DEVICE), x_s.to(DEVICE), y.to(DEVICE)
batch_noise = noise.repeat(*x_s.size()[:2], 1)
x_s = x_s.add(batch_noise)
y_hat = model(x_d, x_s[:, 0, :])[0]
if preds is None:
preds = y_hat.detach().cpu()
obs = y.detach().cpu()
else:
preds = torch.cat((preds, y_hat.detach().cpu()), 0)
obs = torch.cat((obs, y.detach().cpu()), 0)
obs = obs.numpy()
preds = rescale_features(preds.numpy(), variable='output')
# set discharges < 0 to zero
preds[preds < 0] = 0
nse = calc_nse(obs[obs >= 0], preds[obs >= 0])
return nse
def eval_with_added_noise(model: xgb.XGBRegressor, ds_test: CamelsTXT,
noise: torch.Tensor) -> float:
"""Evaluate model on a single basin with added noise
Parameters
----------
model : xgb.XGBRegressor
The XGBoost model to evaluate
ds_test : CamelsTXT
Dataset containing the basin data.
noise : torch.Tensor
Tensor containing the noise for this evaluation run.
Returns
-------
float
Nash-Sutcliffe-Efficiency of the simulations with added noise.
"""
preds, obs = None, None
x = ds_test.x.reshape(len(ds_test.x), -1).numpy()
obs = ds_test.y.numpy()
attributes = ds_test.attributes.repeat(len(x), 1).clone()
batch_noise = noise.repeat(len(attributes), 1)
attributes = attributes.add(batch_noise)
x = np.concatenate([x, attributes.numpy()], axis=1)
preds = model.predict(x)
preds = rescale_features(preds, variable='output')
# set discharges < 0 to zero
preds[preds < 0] = 0
nse = calc_nse(obs[obs >= 0], preds[obs.reshape(-1) >= 0])
return nse
def evaluate_basin(model: xgb.XGBRegressor, ds_test: CamelsTXT,
no_static: bool) -> Tuple[np.ndarray, np.ndarray]:
"""Evaluate model on a single basin
Parameters
----------
model : xgb.XGBRegressor
The XGBoost model to evaluate
ds_test : CamelsTXT
CAMELS dataset containing the basin data
no_static: bool
If True, will not include static attributes as input features
Returns
-------
preds : np.ndarray
Array containing the (rescaled) prediction for the entire data period
obs : np.ndarray
Array containing the observed discharge for the entire data period
"""
preds, obs = None, None
x = ds_test.x.reshape(len(ds_test.x), -1).numpy()
obs = ds_test.y.numpy()
if not no_static:
x = np.concatenate([x, ds_test.attributes.repeat(len(x), 1).numpy()],
axis=1)
preds = model.predict(x)
preds = rescale_features(preds, variable='output')
# set discharges < 0 to zero
preds[preds < 0] = 0
return preds, obs