Add KernelSynth script

README.md

@@ -10,7 +10,7 @@
 
 ## 🚀 News
 
-- **10 May 2024**: 🚀 We added the code for pretraining and fine-tuning Chronos models. You can find it in [this folder](./scripts/training).
+- **10 May 2024**: 🚀 We added the code for pretraining and fine-tuning Chronos models. You can find it in [this folder](./scripts/training). We also added [a script](./scripts/kernel-synth.py) for generating synthetic time series data from Gaussian processes (KernelSynth; see Section 4.2 in the paper for details).
 - **19 Apr 2024**: 🚀 Chronos is now supported on [AutoGluon-TimeSeries](https://auto.gluon.ai/stable/tutorials/timeseries/index.html), the powerful AutoML package for time series forecasting which enables model ensembles, cloud deployments, and much more. Get started with the [tutorial](https://auto.gluon.ai/stable/tutorials/timeseries/forecasting-chronos.html).
 - **08 Apr 2024**: 🧪 Experimental [MLX inference support](https://github.com/amazon-science/chronos-forecasting/tree/mlx) added. If you have an Apple Silicon Mac, you can now obtain significantly faster forecasts from Chronos compared to CPU inference. This provides an alternative way to exploit the GPU on your Apple Silicon Macs together with the "mps" support in PyTorch.
 - **25 Mar 2024**: [v1.1.0 released](https://github.com/amazon-science/chronos-forecasting/releases/tag/v1.1.0) with inference optimizations and `pipeline.embed` to extract encoder embeddings from Chronos.

pyproject.toml

@@ -13,6 +13,7 @@ dependencies = [
 test = ["pytest~=8.0", "numpy~=1.21"]
 typecheck = ["mypy~=1.9"]
 training = ["gluonts[pro]", "numpy", "tensorboard", "typer", "typer-config"]
+kernel-synth = ["gluonts[pro]", "joblib", "scikit-learn"]
 
 [tool.mypy]
 ignore_missing_imports = true

scripts/kernel-synth.py

@@ -0,0 +1,197 @@
+import argparse
+import functools
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+from gluonts.dataset.arrow import ArrowWriter
+from joblib import Parallel, delayed
+from sklearn.gaussian_process import GaussianProcessRegressor
+from sklearn.gaussian_process.kernels import (
+    RBF,
+    ConstantKernel,
+    DotProduct,
+    ExpSineSquared,
+    Kernel,
+    RationalQuadratic,
+    WhiteKernel,
+)
+from tqdm.auto import tqdm
+
+LENGTH = 1024
+KERNEL_BANK = [
+    ExpSineSquared(periodicity=24 / LENGTH),  # H
+    ExpSineSquared(periodicity=48 / LENGTH),  # 0.5H
+    ExpSineSquared(periodicity=96 / LENGTH),  # 0.25H
+    ExpSineSquared(periodicity=24 * 7 / LENGTH),  # H
+    ExpSineSquared(periodicity=48 * 7 / LENGTH),  # 0.5H
+    ExpSineSquared(periodicity=96 * 7 / LENGTH),  # 0.25H
+    ExpSineSquared(periodicity=7 / LENGTH),  # D
+    ExpSineSquared(periodicity=14 / LENGTH),  # 0.5D
+    ExpSineSquared(periodicity=30 / LENGTH),  # D
+    ExpSineSquared(periodicity=60 / LENGTH),  # 0.5D
+    ExpSineSquared(periodicity=365 / LENGTH),  # D
+    ExpSineSquared(periodicity=365 * 2 / LENGTH),  # 0.5D
+    ExpSineSquared(periodicity=4 / LENGTH),  # W
+    ExpSineSquared(periodicity=26 / LENGTH),  # W
+    ExpSineSquared(periodicity=52 / LENGTH),  # W
+    ExpSineSquared(periodicity=4 / LENGTH),  # M
+    ExpSineSquared(periodicity=6 / LENGTH),  # M
+    ExpSineSquared(periodicity=12 / LENGTH),  # M
+    ExpSineSquared(periodicity=4 / LENGTH),  # Q
+    ExpSineSquared(periodicity=4 * 10 / LENGTH),  # Q
+    ExpSineSquared(periodicity=10 / LENGTH),  # Y
+    DotProduct(sigma_0=0.0),
+    DotProduct(sigma_0=1.0),
+    DotProduct(sigma_0=10.0),
+    RBF(length_scale=0.1),
+    RBF(length_scale=1.0),
+    RBF(length_scale=10.0),
+    RationalQuadratic(alpha=0.1),
+    RationalQuadratic(alpha=1.0),
+    RationalQuadratic(alpha=10.0),
+    WhiteKernel(noise_level=0.1),
+    WhiteKernel(noise_level=1.0),
+    ConstantKernel(),
+]
+
+
+def random_binary_map(a: Kernel, b: Kernel):
+    """
+    Applies a random binary operator (+ or *) with equal probability
+    on kernels ``a`` and ``b``.
+
+    Parameters
+    ----------
+    a
+        A GP kernel.
+    b
+        A GP kernel.
+
+    Returns
+    -------
+        The composite kernel `a + b` or `a * b`.
+    """
+    binary_maps = [lambda x, y: x + y, lambda x, y: x * y]
+    return np.random.choice(binary_maps)(a, b)
+
+
+def sample_from_gp_prior(
+    kernel: Kernel, X: np.ndarray, random_seed: Optional[int] = None
+):
+    """
+    Draw a sample from a GP prior.
+
+    Parameters
+    ----------
+    kernel
+        The GP covaraince kernel.
+    X
+        The input "time" points.
+    random_seed, optional
+        The random seed for sampling, by default None.
+
+    Returns
+    -------
+        A time series sampled from the GP prior.
+    """
+    if X.ndim == 1:
+        X = X[:, None]
+
+    assert X.ndim == 2
+    gpr = GaussianProcessRegressor(kernel=kernel)
+    ts = gpr.sample_y(X, n_samples=1, random_state=random_seed)
+
+    return ts
+
+
+def sample_from_gp_prior_efficient(
+    kernel: Kernel,
+    X: np.ndarray,
+    random_seed: Optional[int] = None,
+    method: str = "eigh",
+):
+    """
+    Draw a sample from a GP prior. An efficient version that allows specification
+    of the sampling method. The default sampling method used in GaussianProcessRegressor
+    is based on SVD which is significantly slower that alternatives such as `eigh` and
+    `cholesky`.
+
+    Parameters
+    ----------
+    kernel
+        The GP covaraince kernel.
+    X
+        The input "time" points.
+    random_seed, optional
+        The random seed for sampling, by default None.
+    method, optional
+        The sampling method for multivariate_normal, by default `eigh`.
+
+    Returns
+    -------
+        A time series sampled from the GP prior.
+    """
+    if X.ndim == 1:
+        X = X[:, None]
+
+    assert X.ndim == 2
+
+    cov = kernel(X)
+    ts = np.random.default_rng(seed=random_seed).multivariate_normal(
+        mean=np.zeros(X.shape[0]), cov=cov, method=method
+    )
+
+    return ts
+
+
+def generate_time_series(max_kernels: int = 5):
+    """Generate a synthetic time series from KernelSynth.
+
+    Parameters
+    ----------
+    max_kernels, optional
+        The maximum number of base kernels to use for each time series, by default 5
+
+    Returns
+    -------
+        A time series generated by KernelSynth.
+    """
+    while True:
+        X = np.linspace(0, 1, LENGTH)
+
+        # Randomly select upto max_kernels kernels from the KERNEL_BANK
+        selected_kernels = np.random.choice(
+            KERNEL_BANK, np.random.randint(1, max_kernels + 1), replace=True
+        )
+
+        # Combine the sampled kernels using random binary operators
+        kernel = functools.reduce(random_binary_map, selected_kernels)
+
+        # Sample a time series from the GP prior
+        try:
+            ts = sample_from_gp_prior(kernel=kernel, X=X)
+        except np.linalg.LinAlgError as err:
+            print("Error caught:", err)
+            continue
+
+        # The timestamp is arbitrary
+        return {"start": np.datetime64("2000-01-01 00:00", "s"), "target": ts.squeeze()}
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-N", "--num-series", type=int, default=1000_000)
+    parser.add_argument("-J", "--max-kernels", type=int, default=5)
+    args = parser.parse_args()
+    path = Path(__file__).parent / "kernelsynth-data.arrow"
+
+    generated_dataset = Parallel(n_jobs=-1)(
+        delayed(generate_time_series)(max_kernels=args.max_kernels)
+        for _ in tqdm(range(args.num_series))
+    )
+
+    ArrowWriter(compression="lz4").write_to_file(
+        generated_dataset,
+        path=path,
+    )