def main():
# https://keras.io/layers/recurrent/
num_samples, time_steps, input_dim, output_dim = 50_000, 10, 1, 1
# Definition of the model.
model_keras = NBeatsKeras(backcast_length=time_steps, forecast_length=output_dim,
stack_types=(NBeatsKeras.GENERIC_BLOCK, NBeatsKeras.GENERIC_BLOCK),
nb_blocks_per_stack=2, thetas_dim=(4, 4), share_weights_in_stack=True,
hidden_layer_units=64)
model_pytorch = NBeatsPytorch(backcast_length=time_steps, forecast_length=output_dim,
stack_types=(NBeatsPytorch.GENERIC_BLOCK, NBeatsPytorch.GENERIC_BLOCK),
nb_blocks_per_stack=2, thetas_dim=(4, 4), share_weights_in_stack=True,
hidden_layer_units=64)
# Definition of the objective function and the optimizer.
model_keras.compile(loss='mae', optimizer='adam')
model_pytorch.compile(loss='mae', optimizer='adam')
# Definition of the data. The problem to solve is to find f such as | f(x) - y | -> 0.
# where f = np.mean.
x = np.random.uniform(size=(num_samples, time_steps, input_dim))
y = np.mean(x, axis=1, keepdims=True)
# Split data into training and testing datasets.
c = num_samples // 10
x_train, y_train, x_test, y_test = x[c:], y[c:], x[:c], y[:c]
test_size = len(x_test)
# Train the model.
print('Keras training...')
model_keras.fit(x_train, y_train, validation_data=(x_test, y_test), epochs=20, batch_size=128)
print('Pytorch training...')
model_pytorch.fit(x_train, y_train, validation_data=(x_test, y_test), epochs=20, batch_size=128)
# Save the model for later.
model_keras.save('n_beats_model.h5')
model_pytorch.save('n_beats_pytorch.th')
# Predict on the testing set (forecast).
predictions_keras_forecast = model_keras.predict(x_test)
predictions_pytorch_forecast = model_pytorch.predict(x_test)
np.testing.assert_equal(predictions_keras_forecast.shape, (test_size, model_keras.forecast_length, output_dim))
np.testing.assert_equal(predictions_pytorch_forecast.shape, (test_size, model_pytorch.forecast_length, output_dim))
# Predict on the testing set (backcast).
predictions_keras_backcast = model_keras.predict(x_test, return_backcast=True)
predictions_pytorch_backcast = model_pytorch.predict(x_test, return_backcast=True)
np.testing.assert_equal(predictions_keras_backcast.shape, (test_size, model_keras.backcast_length, output_dim))
np.testing.assert_equal(predictions_pytorch_backcast.shape, (test_size, model_pytorch.backcast_length, output_dim))
# Load the model.
model_keras_2 = NBeatsKeras.load('n_beats_model.h5')
model_pytorch_2 = NBeatsPytorch.load('n_beats_pytorch.th')
np.testing.assert_almost_equal(predictions_keras_forecast, model_keras_2.predict(x_test))
np.testing.assert_almost_equal(predictions_pytorch_forecast, model_pytorch_2.predict(x_test))
def main():
args = get_script_arguments()
device = torch.device(
'cuda') if not args.disable_cuda and torch.cuda.is_available(
) else torch.device('cpu')
forecast_length = 10
backcast_length = 5 * forecast_length
batch_size = 4 # greater than 4 for viz
if args.task == 'm4':
data_gen = batcher(get_m4_data(backcast_length, forecast_length),
batch_size=batch_size,
infinite=True)
elif args.task == 'dummy':
data_gen = dummy_data_generator(backcast_length,
forecast_length,
signal_type='seasonality',
random=True,
batch_size=batch_size)
else:
raise Exception('Unknown task.')
print('--- Model ---')
net = NBeatsNet(
device=device,
stack_types=[NBeatsNet.TREND_BLOCK, NBeatsNet.SEASONALITY_BLOCK],
forecast_length=forecast_length,
thetas_dims=[2, 8],
nb_blocks_per_stack=3,
backcast_length=backcast_length,
hidden_layer_units=1024,
share_weights_in_stack=False,
nb_harmonics=None)
# net = NBeatsNet(device=device,
# stack_types=[NBeatsNet.GENERIC_BLOCK, NBeatsNet.GENERIC_BLOCK],
# forecast_length=forecast_length,
# thetas_dims=[7, 8],
# nb_blocks_per_stack=3,
# backcast_length=backcast_length,
# hidden_layer_units=128,
# share_weights_in_stack=False
# )
optimiser = optim.Adam(net.parameters())
def plot_model(x, target, grad_step):
if not args.disable_plot:
print('plot()')
plot(net, x, target, backcast_length, forecast_length, grad_step)
max_grad_steps = 10000
if args.test:
max_grad_steps = 5
simple_fit(net, optimiser, data_gen, plot_model, device, max_grad_steps)
def train(self, **kwargs):
"""Train NBEATS Model"""
if os.path.isfile(CHECKPOINT_NAME):
os.remove(CHECKPOINT_NAME)
steps = kwargs.get("steps", 50)
batch_size = kwargs.get("batch_size", 10)
patience = kwargs.get("patience", 5)
device = self.get_device()
# There seems to be an issue with
net = NBeatsNet(
stack_types=[
NBeatsNet.TREND_BLOCK,
NBeatsNet.SEASONALITY_BLOCK,
NBeatsNet.GENERIC_BLOCK,
],
forecast_length=self.forecast_len,
thetas_dims=kwargs.get("thetas_dims", [2, 8, 3]),
nb_blocks_per_stack=kwargs.get("nb_blocks_per_stack", 3),
backcast_length=self.forecast_len,
hidden_layer_units=kwargs.get("hidden_layer_units", 128),
share_weights_in_stack=False,
device=device,
)
x_batch, y_batch, norm_constant = self.format_input(
self.dataframe, self.forecast_len)
c = len(x_batch)
if self.in_sample:
c -= self.forecast_len
optimiser = optim.Adam(net.parameters())
data = data_generator(x_batch[:c], y_batch[:c], batch_size)
best_loss = float("inf")
counter = 0
for _ in tqdm(range(steps)):
loss = self.train_100_grad_steps(data, device, net, optimiser)
if loss < best_loss:
best_loss, counter = loss, 0
else:
counter += 1
if counter >= patience:
break
self.model = net
self.constant = norm_constant
def __init__(self,period_to_forecast,data=None, backcast_length=None,save_checkpoint=False,path='',checkpoint_file_name='nbeats-training-checkpoint.th',mode='cpu',batch_size=None,thetas_dims=[7, 8],nb_blocks_per_stack=3,share_weights_in_stack=False,train_percent=0.8,hidden_layer_units=128,stack=None):
if (data is None):
print('For Prediction as no data passed')
batch_size=np.nan
else:
self.data=data
if(len(self.data.shape)!=2):
raise Exception('Numpy array should be of nx1 shape')
if self.data.shape[1]!=1:
raise Exception('Numpy array should be of nx1 shape')
self.forecast_length=period_to_forecast
if backcast_length==None:
self.backcast_length = 3 * self.forecast_length
else:
self.backcast_length=backcast_length
if batch_size==None:
self.batch_size=int(self.data.shape[0]/15)
else:
self.batch_size=batch_size
self.hidden_layer_units=hidden_layer_units
if stack==None:
self.stacks= [NBeatsNet.GENERIC_BLOCK, NBeatsNet.GENERIC_BLOCK]
else:
Dict = {1: NBeatsNet.GENERIC_BLOCK, 2: NBeatsNet.TREND_BLOCK , 3: NBeatsNet.SEASONALITY_BLOCK}
self.stacks = [Dict.get(item, item) for item in stack]
self.CHECKPOINT_NAME= path+checkpoint_file_name
self.device = torch.device(mode) #change to gpu if gpu present for better performance
self.thetas_dims=thetas_dims
self.nb_blocks_per_stack=nb_blocks_per_stack
self.train_size=train_percent
self.share_weights_in_stack=share_weights_in_stack
self.net = NBeatsNet(stack_types=self.stacks,
forecast_length=self.forecast_length,
thetas_dims=self.thetas_dims,
nb_blocks_per_stack=self.nb_blocks_per_stack,
backcast_length=self.backcast_length,
hidden_layer_units=self.hidden_layer_units,
share_weights_in_stack=self.share_weights_in_stack,
device=self.device)
self.parameters=self.net.parameters()
self.global_step_cl=0
self.check_save=save_checkpoint
self.loaded=False
self.saved=True
def main():
num_samples, time_steps, input_dim, output_dim = 50000, 10, 1, 1
# Definition of the model.
model_pytorch = NBeatsPytorch(backcast_length=time_steps, forecast_length=output_dim,
stack_types=(NBeatsPytorch.GENERIC_BLOCK, NBeatsPytorch.GENERIC_BLOCK),
nb_blocks_per_stack=2, thetas_dim=(4, 4), share_weights_in_stack=True,
hidden_layer_units=64)
# Definition of the objective function and the optimizer.
model_pytorch.compile_model(loss='mae', learning_rate=1e-4)
# Definition of the data. The problem to solve is to find f such as | f(x) - y | -> 0.
# where f = np.mean.
x = np.random.uniform(size=(num_samples, time_steps, input_dim))
y = np.mean(x, axis=1, keepdims=True)
# Split data into training and testing datasets.
c = num_samples // 10
x_train, y_train, x_test, y_test = x[c:], y[c:], x[:c], y[:c]
# Train the model.
print('Pytorch training...')
model_pytorch.fit(x_train, y_train, validation_data=(x_test, y_test), epochs=20, batch_size=128)
# Save the model for later.
model_pytorch.save('n_beats_pytorch.th')
# Predict on the testing set.
predictions_pytorch = model_pytorch.predict(x_test)
print(predictions_pytorch.shape)
# Load the model.
model_pytorch_2 = NBeatsPytorch.load('n_beats_pytorch.th')
np.testing.assert_almost_equal(predictions_pytorch, model_pytorch_2.predict(x_test))
def load(self, path=""):
for d in self.classes:
checkpoint = d + "_" + self.checkpoint_name_BASE + f'bl{self.nb_blocks_per_stack}-f{self.forecast_length}-b{self.backcast_length}-btch{self.batch_size}-h{self.hidden_layer_units}.th'
if path:
checkpoint = path + checkpoint
print(
f'From where to load models: {path}, and the current checkpoint is: {checkpoint}'
)
if not os.path.exists(checkpoint):
print("-" * 50)
print(
"* /n * /n * /n * /n * /n * /t THERE IS NO CHECKPOINT LIKE THIS! * /n * /n * /n * /n * "
)
print("-" * 50)
raise ValueError("wrong checkpoint name!")
else:
print("-" * 50)
print(
"*/n*/n*/n*/n*/n*/t CHECKPOINT IS VALID WOOOHOO ! */n*/n*/n*/n*"
)
print("-" * 50)
net = NBeatsNet(
stack_types=[NBeatsNet.GENERIC_BLOCK, NBeatsNet.GENERIC_BLOCK],
forecast_length=self.forecast_length,
thetas_dims=self.thetas,
nb_blocks_per_stack=self.nb_blocks_per_stack,
backcast_length=self.backcast_length,
hidden_layer_units=self.hidden_layer_units,
share_weights_in_stack=False,
device=self.device)
optimiser = optim.Adam(net.parameters())
naf.load(checkpoint, net, optimiser)
net.to(self.device)
self.nets[d] = net
def nbeats_dataframe(df,
forecast_length,
in_sample,
device,
train_portion=0.75):
backcast_length = 1 * forecast_length
df = df.values # just keep np array here for simplicity.
norm_constant = np.max(df)
df = df / norm_constant # small leak to the test set here.
x_train_batch, y = [], []
for i in range(backcast_length + 1,
len(df) - forecast_length +
1): # 25% to 75% so 50% #### Watch out I had to plus one.
x_train_batch.append(df[i - backcast_length:i])
y.append(df[i:i + forecast_length])
x_train_batch = np.array(x_train_batch)[..., 0]
y = np.array(y)[..., 0]
# x_train, y_train = x_train_batch, y
net = NBeatsNet(
stack_types=[NBeatsNet.GENERIC_BLOCK, NBeatsNet.GENERIC_BLOCK],
forecast_length=forecast_length,
thetas_dims=[7, 8],
nb_blocks_per_stack=3,
backcast_length=backcast_length,
hidden_layer_units=128,
share_weights_in_stack=False,
device=device,
)
if in_sample == True:
c = int(len(x_train_batch) * train_portion)
x_train, y_train = x_train_batch[:c], y[:c]
x_test, y_test = x_train_batch[c:], y[c:]
return x_train, y_train, x_test, y_test, net, norm_constant
else:
c = int(len(x_train_batch) * 1)
x_train, y_train = x_train_batch[:c], y[:c]
return x_train, y_train, net, norm_constant
forecast_length = 500
backcast_length = 3 * forecast_length
batch_size = 256
for folder_name in dirs:
name = folder_name.split("/")[-1]
checkpoint_name = name + "_" + checkpoint_name_BASE
if os.path.isfile(checkpoint_name):
continue
net = NBeatsNet(
stack_types=[NBeatsNet.GENERIC_BLOCK, NBeatsNet.GENERIC_BLOCK],
forecast_length=forecast_length,
thetas_dims=[7, 8],
nb_blocks_per_stack=3,
backcast_length=backcast_length,
hidden_layer_units=128,
share_weights_in_stack=False,
device=device)
optimiser = optim.Adam(net.parameters())
test_losses = []
actual_class_dir = data_dir + "/" + name + "/"
for (_, dirs, files) in os.walk(actual_class_dir):
iteration = 0
for file in files:
if 'mat' in file:
continue
class NBeats: #UNIVARIATE DATA TO BE PASSED AS NUMPY ARRAY
def __init__(self,period_to_forecast,data=None, backcast_length=None,save_checkpoint=False,path='',checkpoint_file_name='nbeats-training-checkpoint.th',mode='cpu',batch_size=None,thetas_dims=[7, 8],nb_blocks_per_stack=3,share_weights_in_stack=False,train_percent=0.8,hidden_layer_units=128,stack=None):
if (data is None):
print('For Prediction as no data passed')
batch_size=np.nan
else:
self.data=data
if(len(self.data.shape)!=2):
raise Exception('Numpy array should be of nx1 shape')
if self.data.shape[1]!=1:
raise Exception('Numpy array should be of nx1 shape')
self.forecast_length=period_to_forecast
if backcast_length==None:
self.backcast_length = 3 * self.forecast_length
else:
self.backcast_length=backcast_length
if batch_size==None:
self.batch_size=int(self.data.shape[0]/15)
else:
self.batch_size=batch_size
self.hidden_layer_units=hidden_layer_units
if stack==None:
self.stacks= [NBeatsNet.GENERIC_BLOCK, NBeatsNet.GENERIC_BLOCK]
else:
Dict = {1: NBeatsNet.GENERIC_BLOCK, 2: NBeatsNet.TREND_BLOCK , 3: NBeatsNet.SEASONALITY_BLOCK}
self.stacks = [Dict.get(item, item) for item in stack]
self.CHECKPOINT_NAME= path+checkpoint_file_name
self.device = torch.device(mode) #change to gpu if gpu present for better performance
self.thetas_dims=thetas_dims
self.nb_blocks_per_stack=nb_blocks_per_stack
self.train_size=train_percent
self.share_weights_in_stack=share_weights_in_stack
self.net = NBeatsNet(stack_types=self.stacks,
forecast_length=self.forecast_length,
thetas_dims=self.thetas_dims,
nb_blocks_per_stack=self.nb_blocks_per_stack,
backcast_length=self.backcast_length,
hidden_layer_units=self.hidden_layer_units,
share_weights_in_stack=self.share_weights_in_stack,
device=self.device)
self.parameters=self.net.parameters()
self.global_step_cl=0
self.check_save=save_checkpoint
self.loaded=False
self.saved=True
def plot_scatter(self,*args, **kwargs):
plt.plot(*args, **kwargs)
plt.scatter(*args, **kwargs)
def data_generator(self,x_full, y_full, bs):
def split(arr, size):
arrays = []
while len(arr) > size:
slice_ = arr[:size]
arrays.append(slice_)
arr = arr[size:]
arrays.append(arr)
return arrays
while True:
for rr in split((x_full, y_full), bs):
yield rr
def loader(self,model, optimiser):
if os.path.exists(self.CHECKPOINT_NAME):
checkpoint = torch.load(self.CHECKPOINT_NAME)
model.load_state_dict(checkpoint['model_state_dict'])
optimiser.load_state_dict(checkpoint['optimizer_state_dict'])
grad_step = checkpoint['grad_step']
if self.loaded:
self.norm_constant=checkpoint['norm_constant']
return grad_step
return 0
def saver(self,model, optimiser, grad_step):
if self.saved:
torch.save({
'norm_constant':self.norm_constant,
'grad_step': grad_step,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimiser.state_dict(),
}, self.CHECKPOINT_NAME)
else:
torch.save({
'grad_step': grad_step,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimiser.state_dict(),
}, self.CHECKPOINT_NAME)
def load(self,file=None,optimiser=None):
if file==None:
raise Exception('Empty File Name')
elif file=='':
raise Exception('Empty File Name')
else:
if optimiser==None:
self.optimiser = optim.Adam(self.net.parameters())
else:
self.optimiser = optimiser
self.CHECKPOINT_NAME=file
self.loaded=True
self.global_step_cl=self.loader(self.net,self.optimiser)
self.CHECKPOINT_NAME='nbeats-training-checkpoint.th'
def save(self,file=None):
if file==None:
raise Exception('Empty File Name')
elif file=='':
raise Exception('Empty File Name')
else:
self.CHECKPOINT_NAME=file
self.saver(self.net,self.optimiser,self.global_step_cl)
self.saved=True
def train_100_grad_steps(self,data, test_losses):
if not self.loaded:
global_step = self.loader(self.net, self.optimiser)
self.loaded=False
self.global_step_cl=global_step
for x_train_batch, y_train_batch in data:
self.global_step_cl += 1
self.optimiser.zero_grad()
self.net.train()
_, forecast = self.net(torch.tensor(x_train_batch, dtype=torch.float).to(self.device))
loss = F.mse_loss(forecast, torch.tensor(y_train_batch, dtype=torch.float).to(self.device))
loss.backward()
self.optimiser.step()
if self.verbose:
if self.global_step_cl % 30 == 0:
print(f'grad_step = {str(self.global_step_cl).zfill(6)}, tr_loss = {loss.item():.6f}, te_loss = {test_losses[-1]:.6f}')
if self.global_step_cl > 0 and self.global_step_cl % 100 == 0:
with torch.no_grad():
self.saver(self.net, self.optimiser, self.global_step_cl)
break
return forecast
def eval_test(self, test_losses, x_test, y_test):
self.net.eval()
_, forecast = self.net(torch.tensor(x_test, dtype=torch.float))
test_losses.append(F.mse_loss(forecast, torch.tensor(y_test, dtype=torch.float)).item())
p = forecast.detach().numpy()
if self.plot:
subplots = [221, 222, 223, 224]
plt.figure(1)
for plot_id, i in enumerate(np.random.choice(range(len(x_test)), size=4, replace=False)):
ff, xx, yy = p[i] * self.norm_constant, x_test[i] * self.norm_constant, y_test[i] * self.norm_constant
plt.subplot(subplots[plot_id])
plt.grid()
self.plot_scatter(range(0, self.backcast_length), xx, color='b')
self.plot_scatter(range(self.backcast_length, self.backcast_length + self.forecast_length), yy, color='g')
self.plot_scatter(range(self.backcast_length, self.backcast_length + self.forecast_length), ff, color='r')
plt.show()
def fit(self,epoch=25,optimiser=None,plot=False,verbose=True):
self.plot=plot
self.verbose=verbose
self.epoch=epoch
self.norm_constant = np.max(self.data)
self.data = self.data / self.norm_constant
x_train_batch, y,self.x_forecast = [], [] ,[]
for i in range(self.backcast_length, len(self.data) - self.forecast_length):
x_train_batch.append(self.data[i - self.backcast_length:i])
y.append(self.data[i:i + self.forecast_length])
i+=self.forecast_length
self.x_forecast.append(self.data[i+1 - self.backcast_length:i+1])
x_train_batch = np.array(x_train_batch)[..., 0]
self.x_forecast = np.array(self.x_forecast)[..., 0]
y = np.array(y)[..., 0]
i+=self.forecast_length
if optimiser==None:
self.optimiser = optim.Adam(self.net.parameters())
else:
self.optimiser = optimiser
c = int(len(x_train_batch) * self.train_size)
x_train, y_train = x_train_batch[:c], y[:c]
x_test, y_test = x_train_batch[c:], y[c:]
train_data = self.data_generator(x_train, y_train, self.batch_size)
test_losses = []
for i in range(self.epoch):
self.eval_test(test_losses, x_test, y_test)
self.train_100_grad_steps(train_data, test_losses)
if self.check_save:
pass
else:
if os.path.exists(self.CHECKPOINT_NAME):
os.remove(self.CHECKPOINT_NAME)
def predict(self,predict_data=None):
if (predict_data is None):
_, forecasted_values = self.net(torch.tensor(self.x_forecast, dtype=torch.float))
forecasted_values= forecasted_values.detach().numpy()
forecasted_values = forecasted_values * self.norm_constant
else:
if (predict_data.shape[0]!=self.backcast_length):
raise Exception('Numpy array for prediction input should be of backcast_length: {} x 1 shape'.format(self.backcast_length))
else:
predict_data=predict_data/self.norm_constant
predict_data= np.reshape(predict_data, (self.backcast_length, 1))
_, forecasted_values = self.net(torch.tensor(predict_data.T, dtype=torch.float))
forecasted_values= forecasted_values.detach().numpy()
forecasted_values = forecasted_values * self.norm_constant
return forecasted_values.T
def train_and_score_model(state,
ili_data,
horizon=12,
lookback=120,
split=0.7,
batch_size=10,
model_dir='saved_models',
covid_19_onset='2019-11-01',
device=torch.device('cpu'),
report_interval=100,
num_training_intervals=20):
'''traing a state flu model'''
# check state data exists
supported_states = ili_data.STATE.unique().tolist()
assert state in supported_states, f'{state} not present in the ili_data'
# establish checkpoint
os.makedirs(model_dir, exist_ok=True)
checkpoint_path = f'{model_dir}/nbeats-training-checkpoint.th'
if os.path.isfile(checkpoint_path):
os.remove(checkpoint_path)
# map inputs to commonly used variables
forecast_length = horizon # num weeks in forecast horizon
backcast_length = lookback
norm_constant = 1
# get state data and eliminate COVID-19 period
state_data_df = ili_data.loc['2000-01-01':covid_19_onset].reset_index()
state_data_df.set_index(['STATE', 'DATE'], inplace=True)
state_data = state_data_df.loc[state].values
# create n-beats training data
x_train_batch, ylist = [], []
for i in range(backcast_length, len(state_data) - forecast_length):
x_train_batch.append(state_data[i - backcast_length:i])
ylist.append(state_data[i:i + forecast_length])
x_train_batch = np.array(x_train_batch)[..., 0]
ylist = np.array(ylist)[..., 0]
cut = int(len(x_train_batch) * split)
x_train, y_train = x_train_batch[:cut], ylist[:cut]
x_test, y_test = x_train_batch[cut:], ylist[cut:]
# create n-beats model
net = NBeatsNet(
stack_types=[NBeatsNet.GENERIC_BLOCK, NBeatsNet.GENERIC_BLOCK],
forecast_length=forecast_length,
thetas_dims=[7, 8],
nb_blocks_per_stack=3,
backcast_length=backcast_length,
hidden_layer_units=512, #128,
share_weights_in_stack=False,
device=device)
# set optimizer
optimiser = optim.Adam(net.parameters())
# create data generator
data = data_generator(x_train, y_train, batch_size)
# train model
test_losses = []
training_losses = []
for i in range(num_training_intervals):
eval_test(backcast_length, forecast_length, net, norm_constant,
test_losses, x_test, y_test)
training_losses, test_losses = train_100_grad_steps(
data, device, net, optimiser, training_losses, test_losses,
report_interval, checkpoint_path)
_, forecast = net(torch.tensor(x_test, dtype=torch.float))
y_pred = forecast.detach().numpy()
return y_test, y_pred, checkpoint_path
def main(name, copy, device='cpu'):
xtrain = np.loadtxt('xtrain_{}.txt'.format(name))
ytrain = np.loadtxt('ytrain_{}.txt'.format(name))
xtest = np.loadtxt('xtest_{}.txt'.format(name))
ytest = np.loadtxt('ytest_{}.txt'.format(name))
print('-------we got the DATA dude : it works ------------')
#Reminder of the hyperparameter
backcast_length = 100
forecast_length = 100
#Definition of the seasonality model :
thetas_dim1 = 4,
stack_types1 = NBeatsNet.SEASONALITY_BLOCK,
model1 = NBeatsNet(device=torch.device(device),
backcast_length=backcast_length,
forecast_length=forecast_length,
stack_types=stack_types1,
nb_blocks_per_stack=2,
thetas_dim=thetas_dim1,
share_weights_in_stack=True,
hidden_layer_units=64)
model1.compile_model(loss='mae', learning_rate=1e-5)
plt.figure(figsize=(10, 10))
#model training
model1.fit(xtrain,
ytrain,
validation_data=(xtest, ytest),
epochs=100,
batch_size=10)
model1.save('nbeats_test_seasonality.h5')
predictions = model1.predict(xtest)
print(predictions.shape)
#Definition of the generic model :
thetas_dim2 = 4,
stack_types2 = NBeatsNet.GENERIC_BLOCK,
model2 = NBeatsNet(device=torch.device(device),
backcast_length=backcast_length,
forecast_length=forecast_length,
stack_types=stack_types2,
nb_blocks_per_stack=2,
thetas_dim=thetas_dim2,
share_weights_in_stack=True,
hidden_layer_units=64)
model2.compile_model(loss='mae', learning_rate=1e-5)
#model training
model2.fit(xtrain,
ytrain,
validation_data=(xtest, ytest),
epochs=100,
batch_size=10)
model2.save('nbeats_test_seasonality.h5')
predictions = model2.predict(xtest)
print(predictions.shape)
#Definition of the trend model :
thetas_dim3 = 4,
stack_types3 = NBeatsNet.TREND_BLOCK,
model3 = NBeatsNet(device=torch.device(device),
backcast_length=backcast_length,
forecast_length=forecast_length,
stack_types=stack_types3,
nb_blocks_per_stack=2,
thetas_dim=thetas_dim3,
share_weights_in_stack=True,
hidden_layer_units=64)
model3.compile_model(loss='mae', learning_rate=1e-5)
#model training
model3.fit(xtrain,
ytrain,
validation_data=(xtest, ytest),
epochs=100,
batch_size=10)
model3.save('nbeats_test_seasonality.h5')
predictions = model3.predict(xtest)
print(predictions.shape)
plt.show()