[timeseries] Speed up data preparation for local models

[shchur]

Description of changes:

• Convert the target column into a pd.Series before splitting individual time series. This is much faster than calling .loc on a TimeSeriesDataFrame. For TSDF we end up subsetting the static_features for each iteration in the loop, but these static features aren't used by the underlying models.

Testing on a subset of 5000 items from the M5 competition dataset:

Using code currently on master:

Loaded dataset with 7559974 rows and 5000 items.
Fitting Naive: 35.2s

After current PR:

Loaded dataset with 7559974 rows and 5000 items.
Fitting Naive: 19.9s

Code for reproducing the results

import time
import pandas as pd
from autogluon.timeseries import TimeSeriesDataFrame
from autogluon.timeseries.models import NaiveModel


prediction_length = 28
raw_data = pd.read_parquet("../m5/data/subset.parquet")
static = pd.read_parquet("../m5/data/static.parquet")

raw_data["item_id"] = raw_data["item_id"].astype("str")
static["item_id"] = static["item_id"].astype("str")
static.set_index("item_id", inplace=True)

print(f"Loaded dataset with {len(raw_data)} rows and {raw_data['item_id'].nunique()} items.")
df = TimeSeriesDataFrame(raw_data, static_features=static)

model = NaiveModel(prediction_length=28, target="demand")
start_time = time.time()
model.fit(train_data=df)
preds = model.predict(data=df)
print(f"Fitting Naive: {time.time() - start_time:.1f}s")

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Job PR-2587-efc023c is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-2587/efc023c/index.html