Forecasting with GluonTS & PyTorch on GPUs#

In this tutorial, we learn how to train and evaluate a time series forecasting model with GluonTS on GPUs. We’ll train a DeepAR modal, an auto-regressive deep network that produces probabilistic forecasts. On Union, we show how to easily request for an A100, specify dependencies for this workflow, and visualize our results with Flyte Deck.

Run on Union BYOC

Once you have a Union account, install union:

pip install union

Export the following environment variable to build and push images to your own container registry:

# replace with your registry name
export IMAGE_SPEC_REGISTRY="<your-container-registry>"

Then run the following commands to run the workflow:

git clone https://github.com/unionai/unionai-examples
cd unionai-examples/tutorials/gluonts_time_series
union run --remote gluonts_time_series.py gluonts_wf

The source code for this tutorial can be found here .

Managing Dependencies#

First, let’s import the workflow dependencies:

import os
from pathlib import Path

from flytekit import task, Resources, current_context, workflow, Deck, ImageSpec
from flytekit.types.directory import FlyteDirectory
from flytekit.types.file import FlyteFile
from flytekit.extras.accelerators import A100, L4

We use Flytekit’s ImageSpec to specify our Python dependencies as a list of packages. We’ll use gluonts which uses PyTorch Lightning for training and matplotlib for visualizations.

gluon_image = ImageSpec(
    "gluon-time-series",
    packages=[
        "torch==2.3.1",
        "gluonts[torch]==0.15.1",
        "matplotlib==3.9.1",
        "orjson==3.10.6",
        "union",
        "pandas==2.2.2",
    ],
    registry=os.environ.get("IMAGE_SPEC_REGISTRY"),
)

Download our time series dataset#

First, we download the m4_hourly dataset and cache it, so we can reuse the data in sequential task without re downloading it every time:

@task(container_image=gluon_image, requests=Resources(cpu="2", mem="6Gi"))
def download_m4_hourly_dataset() -> FlyteFile:
    from gluonts.dataset.repository import get_dataset

    working_dir = Path(current_context().working_directory)
    dataset_path = working_dir / "dataset_path"
    dataset_path.mkdir(exist_ok=True)

    get_dataset("m4_hourly", path=dataset_path)

    dataset_compressed = working_dir / "dataset_path.tar.gz"
    _compress(dataset_path, dataset_compressed)
    return dataset_compressed

Training DeepAR on A100#

For our model training task, we use flytekit's @task decorator to configure the hardware and image used to train our model. With requests=Resources(gpu="1") and accelerator=A00, we declare the hardware require to train out deep learning forecasting model. The body of the function is training code from DeepAR’s quick start tutorial. The task returns a FlyteDirectory containing the serialized predictor model.

@task(
    container_image=gluon_image,
    requests=Resources(gpu="1", cpu="2", mem="6Gi"),
    accelerator=A100,
)
def train_predictor(dataset: FlyteFile) -> FlyteDirectory:
    from gluonts.dataset.repository import get_dataset
    from gluonts.torch import DeepAREstimator

    dataset.download()
    working_dir = Path(current_context().working_directory)
    dataset_path = working_dir / "dataset"
    _decompress(dataset.path, dataset_path)
    dataset = get_dataset("m4_hourly", path=dataset_path)

    model = DeepAREstimator(
        prediction_length=dataset.metadata.prediction_length,
        freq="h",
        trainer_kwargs={"max_epochs": 5},
    )

    predictor = model.train(dataset.train)

    ctx = current_context()
    working_dir = Path(ctx.working_directory)
    predictor_path = working_dir / "predictor"
    predictor_path.mkdir(exist_ok=True)

    predictor.serialize(predictor_path)
    return predictor_path

Forecasting and Evaluating Model#

For computing forecast, we declare a L4 GPU as our accelerator and a specific amount of CPU and memory. We also pass in enable_deck=True to visualize our evaluation in a Flyte Deck.

@task(
    container_image=gluon_image,
    requests=Resources(gpu="1", cpu="4", mem="4Gi"),
    accelerator=L4,
    enable_deck=True,
)
def compute_forecasts(dataset: FlyteFile, predictor_directory: FlyteDirectory):
    from gluonts.dataset.repository import get_dataset
    from gluonts.torch.model.predictor import PyTorchPredictor
    from gluonts.evaluation import make_evaluation_predictions
    from gluonts.evaluation import Evaluator
    import pandas as pd

    # Download input data
    dataset.download()
    working_dir = Path(current_context().working_directory)
    dataset_path = working_dir / "dataset"
    _decompress(dataset.path, dataset_path)
    dataset = get_dataset("m4_hourly", path=dataset_path)

    predictor_directory.download()

    predictor = PyTorchPredictor.deserialize(Path(predictor_directory))

    # Create forecast
    forecast_it, ts_it = make_evaluation_predictions(
        dataset=dataset.test,
        predictor=predictor,
        num_samples=100,
    )
    forecasts = list(forecast_it)
    tss = list(ts_it)

    ts_entry = tss[0]
    forecast_entry = forecasts[0]

    # Plot forecast
    import matplotlib.pyplot as plt

    fig, ax = plt.subplots(figsize=(8, 6))
    ax.plot(ts_entry[-150:].to_timestamp())
    ax.tick_params(axis="x", rotation=30)
    forecast_entry.plot(show_label=True, ax=ax)
    fig.legend()

    ctx = current_context()
    forecast_deck = Deck("Forecast", _fig_to_html(fig))
    ctx.decks.insert(0, forecast_deck)

    # Evaluate
    evaluator = Evaluator(quantiles=[0.1, 0.5, 0.9])
    agg_metrics, _ = evaluator(tss, forecasts)

    agg_df = pd.DataFrame.from_dict(agg_metrics, orient="index")
    agg_df.columns = ["metric values"]

    metrics_deck = Deck("Metrics", agg_df.to_html())
    ctx.decks.insert(1, metrics_deck)

Complete Workflow#

Finally, we define the workflow that calls train_predictor and passes it’s output to compute_forecasts. We run the workflow by:

union run --remote gluonts_time_series.py gluonts_wf

@workflow
def gluonts_wf() -> FlyteDirectory:
    dataset = download_m4_hourly_dataset()
    predictor_directory = train_predictor(dataset=dataset)
    compute_forecasts(dataset=dataset, predictor_directory=predictor_directory)
    return predictor_directory

Appendix#

The following are helper functions used by our Flyte tasks. We include functions to convert a matplotlib figure to HTML, decompress and compress a directory.

def _fig_to_html(fig) -> str:
    """Convert matplotlib figure to HTML."""
    import io
    import base64

    fig_bytes = io.BytesIO()
    fig.savefig(fig_bytes, format="jpg")
    fig_bytes.seek(0)
    image_base64 = base64.b64encode(fig_bytes.read()).decode()
    return f'<img src="data:image/png;base64,{image_base64}" alt="Rendered Image" />'

def _decompress(src: Path, dest: Path):
    """Decompress src into dest."""
    import tarfile

    with tarfile.open(src, "r:gz") as tar:
        tar.extractall(path=dest)

def _compress(src: Path, dest: Path):
    """Compress src into dest."""
    import tarfile

    with tarfile.open(dest, "w:gz") as tar:
        for file in src.rglob("*"):
            tar.add(file, arcname=file.relative_to(src))