# 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
# Create data generator
val_generator = SeriesDataGenerator(dlwp, validation_data, add_insolation=add_insolation[m],
input_sel=input_selection[m], output_sel=output_selection[m],
input_time_steps=input_time_steps[m], output_time_steps=output_time_steps[m],
batch_size=64)
# Create TimeSeriesEstimator
estimator = TimeSeriesEstimator(dlwp, val_generator)
# Very crude but for this test I want to exclude the predicted thickness from being added back
if model_labels[m] == r'$\tau$ LSTM16':
estimator._outputs_in_inputs = {'varlev': np.array(['HGT/500'])}
# Make a time series prediction
print('Predicting with model %s...' % model_labels[m])
time_series = estimator.predict(num_forecast_steps, verbose=1)
# Slice the arrays as we want
time_series = time_series.sel(**selection, lat=((time_series.lat >= lat_min) & (time_series.lat <= lat_max)))
dlwp.is_convolutional = True
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
# Create data generator
generator = SeriesDataGenerator(dlwp,
data,
batch_size=216,
input_sel=input_selection,
output_sel=output_selection,
input_time_steps=input_time_steps,
output_time_steps=output_time_steps,
add_insolation=add_insolation,
sequence=sequence)
p_val, t_val = generator.generate([], scale_and_impute=False)
# Create TimeSeriesEstimator
estimator = TimeSeriesEstimator(dlwp, generator)
# Make a time series prediction and convert the predictors for comparison
print('Predicting with model %s...' % model_label)
forecast = estimator.predict(num_plot_steps, verbose=1)
forecast = forecast.sel(**selection)
# Scale the forecast
invert=True)
validation_data = data.sel(sample=validation_set)
train_data = data.sel(sample=train_set)
# For multiple GPUs, increase the batch size
batch_size = n_gpu * batch_size
# Build the data generators
if load_memory or use_keras_fit:
print('Loading data to memory...')
generator = SeriesDataGenerator(dlwp,
train_data,
input_sel=input_selection,
output_sel=output_selection,
input_time_steps=input_time_steps,
output_time_steps=output_time_steps,
batch_size=batch_size,
add_insolation=add_solar,
load=load_memory,
shuffle=shuffle,
interval=step_interval)
if use_keras_fit:
p_train, t_train = generator.generate([])
if validation_data is not None:
val_generator = SeriesDataGenerator(dlwp,
validation_data,
input_sel=input_selection,
output_sel=output_selection,
input_time_steps=input_time_steps,
output_time_steps=output_time_steps,
batch_size=batch_size,
train_set = np.isin(data.sample.values,
np.array(validation_set, dtype='datetime64[ns]'),
assume_unique=True,
invert=True)
validation_data = data.sel(sample=validation_set)
train_data = data.sel(sample=train_set)
# Build the data generators
if load_memory or use_keras_fit:
print('Loading data to memory...')
generator = SeriesDataGenerator(dlwp,
train_data,
input_sel=io_selection,
output_sel=io_selection,
input_time_steps=io_time_steps,
output_time_steps=io_time_steps,
sequence=integration_steps,
add_insolation=add_solar,
batch_size=batch_size,
load=load_memory,
shuffle=shuffle)
if use_keras_fit:
p_train, t_train = generator.generate([])
if validation_data is not None:
val_generator = SeriesDataGenerator(dlwp,
validation_data,
input_sel=io_selection,
output_sel=io_selection,
input_time_steps=io_time_steps,
output_time_steps=io_time_steps,
sequence=integration_steps,
