Practical overview
When it fits
Use Mantis when building a Python time-series classification workflow and when input sequences can be prepared at lengths that are multiples of 32.
Who it is for
- Researchers evaluating foundation models for time-series classification.
- Developers building Python workflows that train classifiers from time-series features.
Documented uses
- Extract features from time-series datasets with a frozen pretrained model.
- Train a separate classifier using extracted features.
- Fine-tune a model for a new time-series classification task.
Requirements & limits
The project is open source under the Apache-2.0 license. The README documents installation through pip and lists Python 3.10 through 3.14.
- Input sequence lengths must be multiples of 32; interpolation is suggested, with 512 described as a reasonable target in many cases.
- The repository provides separate MantisV1 and MantisV2 modules for its listed checkpoints.
- Feature normalization is important when training a linear classifier on extracted features.
- The project describes its models as having a few million parameters and says fine-tuning is possible on a single GPU, or feasible on a CPU.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Install Mantis with pip
Install the Mantis time-series classification package.
Before you start
- Python environment with pip.
Steps
Run the command below.
pip install mantis-tsfmConsult the repository's
pyproject.tomlfor package requirements.
Expected outcome
The README provides a pip-based installation route for the Mantis package.
Guide limitations
- Installation alone does not load a model or classify data.
- The README does not specify additional environment setup for this installation route.
Extract features with a pretrained Mantis model
Prepare time-series data, load a pretrained Mantis checkpoint, and extract features using MantisTrainer.
Before you start
- Install the
mantis-tsfmpackage. - Time-series input with sequence length a multiple of 32; the README suggests interpolation to length 512.
- A selected pretrained checkpoint from the listed Hugging Face models.
Steps
Prepare the time series so its sequence length is a multiple of 32; interpolation to 512 is suggested.
Load a pretrained model, for example with
from mantis.architecture import MantisV1,network = MantisV1(device='cuda'), andnetwork = network.from_pretrained("paris-noah/Mantis-8M").Create the feature-extraction wrapper with
model = MantisTrainer(device='cuda', network=network).Extract features with
Z = model.transform(X), then train a classifier of your choice on those features.
Expected outcome
Mantis produces feature representations for the supplied time-series dataset, which can be used to train a classifier.
Guide limitations
- Input sequence length must be a multiple of 32; interpolation size is a hyperparameter, with 512 suggested based on pretraining.
- The example uses
device='cuda'; other runtime choices are not established by this workflow. - Feature normalization is important when using a linear classifier.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- vfeofanov/mantis · Official homepage title · retained in its source language