TIMEE: Time Series Classification via In-Context Learning

TIMEE is a pretrained transformer for time series classification. It classifies test series in a single forward pass given labeled training examples — no per-dataset training or fine-tuning required.

Installation

pip install timee-ts

Requirements: Python ≥ 3.10, PyTorch ≥ 2.0.

Quickstart

from timee import TimeeClassifier
import numpy as np

clf = TimeeClassifier.from_pretrained()  # downloads from HuggingFace on first use

# X: (n_samples, n_channels, seq_len) float32
X_train = np.random.randn(20, 1, 256).astype(np.float32)
y_train = np.array([0, 1] * 10)
X_test  = np.random.randn(5, 1, 256).astype(np.float32)

predictions, probabilities = clf.predict(X_train, y_train, X_test)

Labels can be any type (int, str, etc.). Datasets with more than 10 classes are handled automatically via one-vs-rest.

API

TimeeClassifier.from_pretrained(path, device=None, use_ensemble=True)

Loads a checkpoint from a directory containing model.safetensors.

Parameter	Default	Description
path	"liamsbhoo/timee"	Local directory or HuggingFace Hub repo ID.
device	auto	"cuda", "cpu", or torch.device. Defaults to CUDA > MPS > CPU.
use_ensemble	True	4-member preprocessing ensemble (interpolate×{256,512} × {raw, diff}). Set False for faster single-pass inference.

clf.predict(X_train, y_train, X_test)

Returns (predictions, probabilities):

• predictions: (n_test,), same type as y_train
• probabilities: (n_test, n_classes), rows sum to 1

Multivariate Support (beta)

Note: TIMEE is trained and evaluated as a univariate classifier — multivariate is not its focus (yet).

As mentioned in the paper, TIMEE supports multivariate input through two mechanisms:

1. channel-independent: each channel is classified separately and the per-channel class probabilities are averaged.
2. late-channel-mixing: channels are embedded separately before going through an attention pooling layer (acting as a mixer).

Mechanism (1) is already implemented in TimeeClassifier: pass (n_samples, n_channels, seq_len) input (use n_channels=1 for the univariate case) and it averages over channels automatically.

Mechanism (2) is implemented as TimeeMultivariateClassifier, which fuses channels via attention pooling and takes the same (n_samples, n_channels, seq_len) input.

from timee import TimeeMultivariateClassifier

clf = TimeeMultivariateClassifier.from_pretrained()
predictions, probabilities = clf.predict(X_train, y_train, X_test)

Citation

@misc{küken2026timeeendtoendtimeseries,
      title={TimEE: End-to-end Time Series Classification via In-Context Learning},
      author={Jaris Küken and Shi Bin Hoo and Martin Mráz and Frank Hutter and Lennart Purucker},
      year={2026},
      eprint={2607.07500},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2607.07500},
}