[timeseries] Speed up prediction for GluonTS models

[shchur]

Description of changes:

• Significantly speed up prediction for GluonTS models. The two main sources are:
  • Use of pd.Series instead of TimeSeriesDataFrame inside SimpleGluonTSDataset to avoid copying the static features.
  • Speed up get_forecast_horizon_index_ts_dataframe with a more efficient groupby operation
• Clean up redundant logic in AbstractGluonTSModel. predict.

Benchmarking results on the M5 dataset

(only training for 1 epoch / 1 batch, so the training time/performance are meaningless here)

With this PR

5K items / 7M rows

Training timeseries model DeepAR.
        -0.9107       = Validation score (-mean_wQuantileLoss)
        5.99    s     = Training runtime
        47.57   s     = Validation (prediction) runtime

30K items / 47M rows

Training timeseries model DeepAR.
        -0.7019       = Validation score (-mean_wQuantileLoss)
        11.92   s     = Training runtime
        298.31  s     = Validation (prediction) runtime

Current master branch
5K items / 7M rows

Training timeseries model DeepAR.
        -0.8331       = Validation score (-mean_wQuantileLoss)
        6.61    s     = Training runtime
        81.16   s     = Validation (prediction) runtime

30K items / 47M rows

Training timeseries model DeepAR.
        -0.8672       = Validation score (-mean_wQuantileLoss)
        15.64   s     = Training runtime
        1107.97 s     = Validation (prediction) runtime

Code to reproduce the results

import pandas as pd
from autogluon.timeseries import TimeSeriesDataFrame, TimeSeriesPredictor


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

raw_data["item_id"] = raw_data["item_id"].astype("str")
raw_data["timestamp"] = pd.to_datetime(raw_data["timestamp"])
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)

predictor = TimeSeriesPredictor(prediction_length=prediction_length, target="demand")
predictor.fit(
    train_data=df,
    hyperparameters={"DeepAR": {"epochs": 1, "num_batches_per_epoch": 1}},
)

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-2593-4386fe1 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-2593/4386fe1/index.html

[canerturkmen]

Looks great, and very impressive speedups!! Some minor comments.

[canerturkmen]

LGTM! Thanks!

[github-actions]

Job PR-2593-6bb80b7 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-2593/6bb80b7/index.html