Supported models¶
tslens is architecture-agnostic. It attributes any callable that maps
(batch, seq_len, n_features) — or a tuple
of tensors — to predictions. There is no model registry and nothing to subclass: if the
forward pass is a differentiable-or-not PyTorch computation, it can be explained.
That said, real forecasting/classification models rarely take a single clean tensor.
The tables below are the models this package (and its baselines) have actually been
run against, and exactly how to wire each calling convention into .attribute(...).
Not listed?
The tables record tested integrations, not a compatibility limit. If your callable accepts tensors and returns predictions, start with the integration cookbook.
Single-input models¶
Pass the series straight through — nothing to split.
| Model | Family | Paper |
|---|---|---|
| DLinear | Linear | Zeng et al., AAAI 2023 |
| LightTS | Linear/MLP | Zhang et al., arXiv:2207.01186 |
| TiDE | Linear/MLP | Das et al., TMLR 2023 |
| FiLM | Linear/MLP | Zhou et al., NeurIPS 2022 |
| TSMixer | MLP-Mixer | Chen et al., TMLR 2023 |
| FreTS | Frequency-domain MLP | Yi et al., NeurIPS 2023 |
| MICN | Convolutional | Wang et al., ICLR 2023 |
| Crossformer | Transformer | Zhang & Yan, ICLR 2023 |
| PatchTST | Transformer | Nie et al., ICLR 2023 |
| Pyraformer | Transformer | Liu et al., ICLR 2022 |
| SegRNN | Recurrent | Lin et al., arXiv:2308.11200 |
| Koopa | Koopman operator | Liu et al., NeurIPS 2023 |
| LSTM | Recurrent | Hochreiter & Schmidhuber, Neural Computation 1997 |
| TCN | Convolutional | Bai et al., arXiv:1803.01271 |
| CALF | LLM-backed foundation model | Liu et al., arXiv:2403.07300 |
| OFA (GPT4TS) | LLM-backed foundation model | Zhou et al., NeurIPS 2023 |
| TimeLLM | LLM-backed foundation model | Jin et al., ICLR 2024 |
attr = WinTSR(model).attribute(inputs=x_enc, baselines=torch.zeros_like(x_enc))
Dual-input (TSlib) models¶
These TSlib models consume every
forward argument themselves (x_enc, x_mark_enc, x_dec, x_mark_dec), so the calendar
features get attributed alongside the series. tslens.attr.tsr.DUAL_INPUT_USERS is the
canonical list — pass dual_input_users=[...] to TSR if you're explaining a model not
on it.
| Model | Family | Paper |
|---|---|---|
| Transformer | Transformer | Vaswani et al., NeurIPS 2017 |
| Informer | Transformer | Zhou et al., AAAI 2021 |
| Autoformer | Transformer | Wu et al., NeurIPS 2021 |
| FEDformer | Transformer | Zhou et al., ICML 2022 |
| ETSformer | Transformer | Woo et al., arXiv:2202.01381 |
| Nonstationary Transformer | Transformer | Liu et al., NeurIPS 2022 |
| Reformer | Transformer | Kitaev et al., ICLR 2020 |
| iTransformer | Transformer | Liu et al., ICLR 2024 |
| TimeXer | Transformer | Wang et al., NeurIPS 2024 |
| TimeMixer | MLP-Mixer | Wang et al., ICLR 2024 |
| TimesNet | Convolutional | Wu et al., ICLR 2023 |
| RNN | Recurrent | Hochreiter & Schmidhuber, Neural Computation 1997 |
attr_enc, attr_mark = WinTSR(model).attribute(
inputs=(x_enc, x_mark_enc),
baselines=(torch.zeros_like(x_enc), torch.zeros_like(x_mark_enc)),
additional_forward_args=(x_dec, x_mark_dec),
)
See the integration cookbook for the full walkthrough, or the TSlib models notebook to run it.
Pretrained foundation models¶
Timer¶
Timer is a generative pretrained Transformer for
zero-shot time-series forecasting. The
Pretrained Timer notebook
loads the public 84M-parameter checkpoint, forecasts ETTh2, and adapts its
generate() interface for perturbation-based WinTSR attribution.
MOMENT¶
MOMENT is an open foundation-model family for forecasting and other time-series tasks. The Pretrained MOMENT notebook loads MOMENT-1-small, forecasts ETTh2, and bridges its channel-first 512-step input to WinTSR's time-first attribution interface.
Tiny Time Mixers (TTM)¶
TTM is IBM's compact pretrained model family for zero- and few-shot forecasting. The Pretrained TTM notebook loads the Granite TTM-R2 512/96 checkpoint, attributes its native time-first input, and visualizes how WinTSR's relevance threshold changes the result.
GPT4TS / One Fits All¶
GPT4TS adapts a mostly frozen pretrained GPT-2 backbone to time-series tasks through trainable input and output layers. The Pretrained GPT4TS notebook fits those forecasting layers on ETTh2, verifies a held-out forecast, and explains the result through GPT4TS's single-input convention.
Model zoo and training harness¶
Trained checkpoints, dataset loaders, and experiment scripts for all of the above live
in WinTSR-research — the paper's
training/interpretation harness, which depends on this package the same way any user
would (pip install tslens). This repository ships the attribution methods only; it
does not vendor model code.