LSTM Regression Model¶

In [1]:

from deep_river.regression import RollingRegressor
from river import (
    metrics,
    compose,
    preprocessing,
    datasets,
    stats,
    feature_extraction,
)
from torch import nn
from tqdm import tqdm

from deep_river.regression import RollingRegressor from river import ( metrics, compose, preprocessing, datasets, stats, feature_extraction, ) from torch import nn from tqdm import tqdm

In [2]:

def get_hour(x):
    x["hour"] = x["moment"].hour
    return x

def get_hour(x): x["hour"] = x["moment"].hour return x

Simple RNN Regression Model¶

In [3]:

class RnnModule(nn.Module):

    def __init__(self, n_features, hidden_size):
        super().__init__()
        self.n_features = n_features
        self.rnn = nn.RNN(
            input_size=n_features, hidden_size=hidden_size, num_layers=1
        )
        self.fc = nn.Linear(in_features=hidden_size, out_features=1)

    def forward(self, X, **kwargs):
        output, hn = self.rnn(X)  # lstm with input, hidden, and internal state
        return self.fc(output[-1, :])

class RnnModule(nn.Module): def __init__(self, n_features, hidden_size): super().__init__() self.n_features = n_features self.rnn = nn.RNN( input_size=n_features, hidden_size=hidden_size, num_layers=1 ) self.fc = nn.Linear(in_features=hidden_size, out_features=1) def forward(self, X, **kwargs): output, hn = self.rnn(X) # lstm with input, hidden, and internal state return self.fc(output[-1, :])

In [4]:

dataset = datasets.Bikes()
metric = metrics.MAE()

model_pipeline = compose.Select(
    "clouds", "humidity", "pressure", "temperature", "wind"
)
model_pipeline += get_hour | feature_extraction.TargetAgg(
    by=["station", "hour"], how=stats.Mean()
)
model_pipeline |= preprocessing.StandardScaler()
model_pipeline |= RollingRegressor(
    module=RnnModule,
    loss_fn="mse",
    optimizer_fn="sgd",
    window_size=20,
    lr=1e-2,
    hidden_size=32,  # parameters of MyModule can be overwritten
    append_predict=True,
)
model_pipeline

dataset = datasets.Bikes() metric = metrics.MAE() model_pipeline = compose.Select( "clouds", "humidity", "pressure", "temperature", "wind" ) model_pipeline += get_hour | feature_extraction.TargetAgg( by=["station", "hour"], how=stats.Mean() ) model_pipeline |= preprocessing.StandardScaler() model_pipeline |= RollingRegressor( module=RnnModule, loss_fn="mse", optimizer_fn="sgd", window_size=20, lr=1e-2, hidden_size=32, # parameters of MyModule can be overwritten append_predict=True, ) model_pipeline

Out[4]:

['clouds', 'humidity', 'pressure', 'temperature', 'wind']

( clouds humidity pressure temperature wind )

get_hour

def get_hour(x): x['hour'] = x['moment'].hour return x

y_mean_by_station_and_hour

( by=['station', 'hour'] how=Mean () target_name="y" )

StandardScaler

( with_std=True )

RollingRegressor

( module=None loss_fn="mse_loss" optimizer_fn=<class 'torch.optim.sgd.SGD'> lr=0.01 window_size=20 append_predict=True device="cpu" seed=42 )

In [5]:

for x, y in tqdm(dataset.take(5000)):
    y_pred = model_pipeline.predict_one(x)
    metric.update(y_true=y, y_pred=y_pred)
    model_pipeline.learn_one(x=x, y=y)
print(f"MAE: {metric.get():.2f}")

for x, y in tqdm(dataset.take(5000)): y_pred = model_pipeline.predict_one(x) metric.update(y_true=y, y_pred=y_pred) model_pipeline.learn_one(x=x, y=y) print(f"MAE: {metric.get():.2f}")

5000it [00:11, 451.42it/s]

MAE: 3.94

LSTM Regression Model¶

In [6]:

class LstmModule(nn.Module):

    def __init__(self, n_features, hidden_size=1):
        super().__init__()
        self.n_features = n_features
        self.hidden_size = hidden_size
        self.lstm = nn.LSTM(
            input_size=n_features,
            hidden_size=hidden_size,
            num_layers=1,
            bidirectional=False,
        )
        self.fc = nn.Linear(in_features=hidden_size, out_features=1)

    def forward(self, X, **kwargs):
        output, (hn, cn) = self.lstm(
            X
        )  # lstm with input, hidden, and internal state
        return self.fc(output[-1, :])

class LstmModule(nn.Module): def __init__(self, n_features, hidden_size=1): super().__init__() self.n_features = n_features self.hidden_size = hidden_size self.lstm = nn.LSTM( input_size=n_features, hidden_size=hidden_size, num_layers=1, bidirectional=False, ) self.fc = nn.Linear(in_features=hidden_size, out_features=1) def forward(self, X, **kwargs): output, (hn, cn) = self.lstm( X ) # lstm with input, hidden, and internal state return self.fc(output[-1, :])

In [7]:

dataset = datasets.Bikes()
metric = metrics.MAE()

model_pipeline = compose.Select(
    "clouds", "humidity", "pressure", "temperature", "wind"
)
model_pipeline += get_hour | feature_extraction.TargetAgg(
    by=["station", "hour"], how=stats.Mean()
)
model_pipeline |= preprocessing.StandardScaler()
model_pipeline |= RollingRegressor(
    module=LstmModule,
    loss_fn="mse",
    optimizer_fn="sgd",
    window_size=20,
    lr=1e-2,
    hidden_size=32,  # parameters of MyModule can be overwritten
    append_predict=True,
)
model_pipeline

dataset = datasets.Bikes() metric = metrics.MAE() model_pipeline = compose.Select( "clouds", "humidity", "pressure", "temperature", "wind" ) model_pipeline += get_hour | feature_extraction.TargetAgg( by=["station", "hour"], how=stats.Mean() ) model_pipeline |= preprocessing.StandardScaler() model_pipeline |= RollingRegressor( module=LstmModule, loss_fn="mse", optimizer_fn="sgd", window_size=20, lr=1e-2, hidden_size=32, # parameters of MyModule can be overwritten append_predict=True, ) model_pipeline

Out[7]:

['clouds', 'humidity', 'pressure', 'temperature', 'wind']

( clouds humidity pressure temperature wind )

get_hour

def get_hour(x): x['hour'] = x['moment'].hour return x

y_mean_by_station_and_hour

( by=['station', 'hour'] how=Mean () target_name="y" )

StandardScaler

( with_std=True )

RollingRegressor

( module=None loss_fn="mse_loss" optimizer_fn=<class 'torch.optim.sgd.SGD'> lr=0.01 window_size=20 append_predict=True device="cpu" seed=42 )

In [8]:

for x, y in tqdm(dataset.take(5000)):
    y_pred = model_pipeline.predict_one(x)
    metric.update(y_true=y, y_pred=y_pred)
    model_pipeline.learn_one(x=x, y=y)
print(f"MAE: {metric.get():.2f}")

for x, y in tqdm(dataset.take(5000)): y_pred = model_pipeline.predict_one(x) metric.update(y_true=y, y_pred=y_pred) model_pipeline.learn_one(x=x, y=y) print(f"MAE: {metric.get():.2f}")

5000it [00:22, 225.22it/s]

MAE: 2.81