LSTM Regression Model¶

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

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

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, :])

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(10, 16),
    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

for x, y in 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}")

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, :])

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(10, 16),
    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

for x, y in 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}")