Practical overview
When it fits
Use it when you want to experiment with the repository’s unsupervised reinforcement-learning approach and already have, or can prepare, video features in the expected HDF5 format.
Who it is for
- Researchers studying unsupervised video summarization and deep reinforcement learning.
- Developers who can prepare video-feature datasets in the documented HDF5 format and want to train or evaluate the model.
Documented uses
- Create randomized dataset splits for training and evaluation.
- Train a video-summary policy network on preprocessed feature data.
- Evaluate a saved model and store results in HDF5.
- Plot evaluation scores or training reward curves.
- Render selected frames into an MP4 using an existing frame directory.
Requirements & limits
The repository is public and carries an MIT license.
- The README specifies Python 2.7 and PyTorch 0.4.0; it also identifies h5py and tabulate as dependencies that may need installation.
- The documented model input is preprocessed video features stored in HDF5, not raw video.
- The summary-rendering script requires a frame directory and matching filename configuration.
- The README reports that the original QMUL dataset server is inaccessible and links to an alternative dataset download.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare dataset splits for the SumMe workflow
Create the five dataset splits used by the documented SumMe training and evaluation commands.
Before you start
- Python 2.7 and PyTorch 0.4.0, as specified by the README.
- The repository cloned locally.
- The preprocessed SumMe HDF5 dataset at datasets/eccv16_dataset_summe_google_pool5.h5.
Steps
Clone https://github.com/KaiyangZhou/pytorch-vsumm-reinforce and change into the repository directory.
Obtain and extract the preprocessed datasets archive from the updated Google Drive location linked in the README.
Run the command below.
python create_split.py -d datasets/eccv16_dataset_summe_google_pool5.h5 --save-dir datasets --save-name summe_splits --num-splits 5
Expected outcome
Five randomized splits are saved as a JSON file named summe_splits.json in the datasets directory.
Guide limitations
- The README says the original QMUL dataset server is inaccessible and points to a Google Drive download instead.
- This guide prepares splits only; it does not run training or evaluation.
Render selected frames as an MP4 summary
Convert a saved machine summary and a directory of video frames into an MP4 file.
Before you start
- A results HDF5 file containing the summary.
- A directory containing the corresponding video frames.
- Frame filenames matching the format configured in summary2video.py.
Steps
Set the frame naming format in summary2video.py to match the video-frame files.
Run the command below.
python summary2video.py -p path_to/result.h5 -d path_to/video_frames -i 0 --fps 30 --save-dir log --save-name summary.mp4
Expected outcome
The script writes an MP4 summary to the specified save directory and filename.
Guide limitations
- The required frame directory and an existing result file are not created by this rendering command.
- The README states that frame naming must be configured to match the files.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- KaiyangZhou/pytorch-vsumm-reinforce · Official homepage title · retained in its source language
Related options
KaiyangZhou/vsumm-reinforce — The README identifies this as the original Theano implementation of the method.