---
license: apache-2.0
---
# Mantis: Lightweight Calibrated Foundation Model for User-Friendly Time Series Classification
> **🚨 NEW Version 0.2.0: Mantis pre-training is now available! 🚨**
## Overview
**MANTIS** is an open-source python package with a pre-trained time series classification foundation model implemented by Huawei Noah's Ark Lab.
This is a repository the model checkpoint. Please refer to the [GitHub](https://github.com/vfeofanov/mantis/tree/main) repository of the package and
the technical report on [arXiv](https://arxiv.org/abs/2502.15637) for more details.
## Installation
### Pip installation
It can be installed via `pip` by running:
```
pip install mantis-tsfm
```
The requirements can be verified at [`pyproject.toml`](https://github.com/vfeofanov/mantis/blob/main/pyproject.toml)
### Editable mode using Poetry
First, install Poetry and add the path to the binary file to your shell configuration file.
For example, on Linux systems, you can do this by running:
```bash
curl -sSL https://install.python-poetry.org | python3 -
export PATH="/home/username/.local/bin:$PATH"
```
Now you can create a virtual environment that is based on one of your already installed Python interpreters.
For example, if your default Python is 3.9, then create the environment by running:
```bash
poetry env use 3.9
```
Alternatively, you can specify a path to the interpreter. For example, to use an Anaconda Python interpreter:
```bash
poetry env use /path/to/anaconda3/envs/my_env/bin/python
```
If you want to run any command within the environment, instead of activating the environment manually, you can use `poetry run`:
```bash
poetry run
```
For example, to install the dependencies and run tests:
```bash
poetry install
poetry run pytest
```
If dependencies are not resolving correctly, try re-generating the lock file:
```bash
poetry lock
poetry install
```
## Getting started
Please refer to [`getting_started/`](https://github.com/vfeofanov/mantis/tree/main/getting_started) to see reproducible examples of how the package can be used.
Below we summarize the basic commands needed to use the package.
### Prepare Data.
As an input, Mantis accepts any time series with sequence length **proportional** to 32, which corresponds to the number of tokens fixed in our model.
We found that resizing time series via interpolation is generally a good choice:
``` python
import torch
import torch.nn.functional as F
def resize(X):
X_scaled = F.interpolate(torch.tensor(X, dtype=torch.float), size=512, mode='linear', align_corners=False)
return X_scaled.numpy()
```
Generally speaking, the interpolation size is a hyperparameter to play with. Nevertheless, since Mantis was pre-trained on sequences of length 512, interpolating to this length looks reasonable in most of cases.
### Initialization.
To load our pre-trained model from the HuggingFace, it is sufficient to run:
``` python
from mantis.architecture import Mantis8M
network = Mantis8M(device='cuda')
network = network.from_pretrained("paris-noah/Mantis-8M")
```
### Feature Extraction.
We provide a scikit-learn-like wrapper `MantisTrainer` that allows to use Mantis as a feature extractor by running the following commands:
``` python
from mantis.trainer import MantisTrainer
model = MantisTrainer(device='cuda', network=network)
Z = model.transform(X) # X is your time series dataset
```
### Fine-tuning.
If you want to fine-tune the model on your supervised dataset, you can use `fit` method of `MantisTrainer`:
``` python
from mantis.trainer import MantisTrainer
model = MantisTrainer(device='cuda', network=network)
model.fit(X, y) # y is a vector with class labels
probs = model.predict_proba(X)
y_pred = model.predict(X)
```
### Adapters.
We have integrated into the framework the possibility to pass the input to an adapter before sending it to the foundation model. This may be useful for time series data sets with a large number of channels. More specifically, large number of channels may induce the curse of dimensionality or make model's fine-tuning unfeasible.
A straightforward way to overcome these issues is to use a dimension reduction approach like PCA:
``` python
from mantis.adapters import MultichannelProjector
adapter = MultichannelProjector(new_num_channels=5, base_projector='pca')
adapter.fit(X)
X_transformed = adapter.transform(X)
model = MantisTrainer(device='cuda', network=network)
Z = model.transform(X_transformed)
```
Another wat is to add learnable layers before the foundation model and fine-tune them with the prediction head:
``` python
from mantis.adapters import LinearChannelCombiner
model = MantisTrainer(device='cuda', network=network)
adapter = LinearChannelCombiner(num_channels=X.shape[1], new_num_channels=5)
model.fit(X, y, adapter=adapter, fine_tuning_type='adapter_head')
```
### Pre-training.
The model can be pre-trained using the `pretrain` method of `MantisTrainer` that supports data parallelization. You can see a pre-training demo at `getting_started/pretrain.py`.
For example, to pre-train the model on 4 GPUs, you can run the following commands:
```
cd getting_started/
python -m torch.distributed.run --nproc_per_node=4 --nnodes=1 pretrain.py --seed 42
```
## Citing Mantis 📚
If you use Mantis in your work, please cite this technical report:
```bibtex
@article{feofanov2025mantis,
title={Mantis: Lightweight Calibrated Foundation Model for User-Friendly Time Series Classification},
author={Vasilii Feofanov and Songkang Wen and Marius Alonso and Romain Ilbert and Hongbo Guo and Malik Tiomoko and Lujia Pan and Jianfeng Zhang and Ievgen Redko},
journal={arXiv preprint arXiv:2502.15637},
year={2025},
}
```