Practical overview
When it fits
Use it when you are building a Python video-classification training workflow and can prepare the labeled data and model configuration yourself.
Who it is for
- Python developers adapting pretrained video models to labeled classification datasets.
- Machine-learning practitioners who want a training workflow with data loading, evaluation and experiment tracking components.
Documented uses
- Fine-tune a video classifier on a custom set of labeled clips.
- Compose a model from a vision backbone, classification head and temporal neck.
- Train across devices, evaluate model performance and track experiments.
- Prepare fine-tuned models for Hugging Face Hub or an ONNX-oriented workflow.
Requirements & limits
The public GitHub repository declares an MIT license.
- The documented dataset layout uses separate training and validation roots with class-named folders; AVI and MP4 extensions are supported.
- The README installation instructions include PyTorch, PyTorchVideo from its main branch, Transformers from its main branch and video-transformers.
- ONNX export and Hugging Face Hub workflows are described, while an ONNX Runtime inference workflow is not shown.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Fine-tune a video classifier from labeled folders
Configure and start a single-label video classification training run using the README's Timesformer example.
Before you start
- Python environment with PyTorch, PyTorchVideo, Transformers and video-transformers installed.
- Training and validation folders organized into class subfolders; AVI and MP4 files are documented as supported.
- A pretrained model accessible through the Hugging Face model identifier used in the example.
Steps
Install PyTorch, PyTorchVideo and Transformers, then install video-transformers as shown in the README.
Arrange training and validation videos under separate roots, with one subfolder per label.
Create a TransformersBackbone, VideoDataModule, LinearHead and VideoModel, setting the data roots and preprocessing parameters.
Create the single_label_classification trainer with an optimizer and epoch limit, then call trainer.fit().
Expected outcome
The example configures and starts fine-tuning a Timesformer video classifier on the prepared labeled data.
Guide limitations
- The guide follows the documented Timesformer single-label example; dataset-specific preprocessing and model choices may require adaptation.
- The README does not specify hardware requirements or demonstrate a completed run.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- fcakyon/video-transformers · Official homepage title · retained in its source language