variable_std = data.sel(**selection).variables['std'].values
os.makedirs(plot_directory, exist_ok=True)
#%% Make the verification and forecast
verification = verify.verification_from_samples(data.sel(**selection),
forecast_steps=num_plot_steps,
dt=model_dt)
# Scale the verification
if scale_variables:
verification = verification * variable_std + variable_mean
print('Loading model %s...' % model)
dlwp = load_model(model)
# Build in some tolerance for old models trained with former APIs missing the is_convolutional and is_recurrent
# attributes. This may not always work!
if not hasattr(dlwp, 'is_recurrent'):
dlwp.is_recurrent = False
for layer in dlwp.model.layers:
if 'LSTM' in layer.name.upper() or 'LST_M' in layer.name.upper():
dlwp.is_recurrent = True
if not hasattr(dlwp, 'is_convolutional'):
dlwp.is_convolutional = False
for layer in dlwp.model.layers:
if 'CONV' in layer.name.upper():
dlwp.is_convolutional = True
if isinstance(dlwp, DLWPFunctional):
if not hasattr(dlwp, '_n_steps'):
for m, model in enumerate(models):
print('Loading model %s...' % model)
# Some tolerance for using a weighted loss function. Unreliable but doesn't hurt.
if 'weight' in model.lower():
lats = validation_data.lat.values
output_shape = (validation_data.dims['lat'], validation_data.dims['lon'])
if crop_north_pole:
lats = lats[1:]
customs = {'loss': latitude_weighted_loss(mean_squared_error, lats, output_shape, axis=-2,
weighting='midlatitude')}
else:
customs = None
# Load the model
dlwp, history = load_model('%s/%s' % (root_directory, model), True, custom_objects=customs)
# Assign forecast hour coordinate
if isinstance(dlwp, DLWPFunctional):
if not hasattr(dlwp, '_n_steps'):
dlwp._n_steps = 6
if not hasattr(dlwp, 'time_dim'):
dlwp.time_dim = 2
sequence = dlwp._n_steps
else:
sequence = None
f_hours.append(np.arange(dt, num_forecast_steps * dt + 1., dt))
# Build in some tolerance for old models trained with former APIs missing the is_convolutional and is_recurrent
# attributes. This may not always work!
if not hasattr(dlwp, 'is_recurrent'):
#%% Make forecasts
model_forecasts = []
if plot_laplace:
laplace_forecasts = []
else:
laplace_fill = None
num_forecast_steps = int(np.ceil(plot_forecast_hour / model_dt))
f_hour = np.arange(model_dt, num_forecast_steps * model_dt + 1, model_dt)
dlwp, p_val, t_val = None, None, None
for mod, model in enumerate(models):
print('Loading model %s...' % model)
dlwp, history = load_model('%s/%s' % (root_directory, model), True)
# Create data generator
if predictor_sel[mod] is not None:
val_ds = data.sel(**predictor_sel[mod])
else:
val_ds = data.copy()
if crop_north_pole:
val_ds = val_ds.isel(lat=(val_ds.lat < 90.0))
val_generator = DataGenerator(dlwp, val_ds, batch_size=216)
p_val, t_val = val_generator.generate([], scale_and_impute=False)
# Make a time series prediction and convert the predictors for comparison
print('Predicting with model %s...' % model_labels[mod])
time_series = dlwp.predict_timeseries(p_val, num_forecast_steps)
if scale_variables: