TURN-TAP-pytorch

This is a pytorch implementation of TURN TAP: Temporal Unit Regression Network for Temporal Action Proposals. This code is for research purpose and suggestions are welcome.

Enviorment

Pytorch 0.4.0

CUDA 8.0

Python 2.7.6

References

The tensorflow implementation code provided by the authors: https://github.com/jiyanggao/TURN-TAP/tree/master/turn_codes

Prepare the features

DenseFlow Features in the Google Drive：val set, test set

Setup

use git to clone this repository

$ git clone --recursive https://github.com/JunxuanZhang/TURN-TAP-pytorch/

Then create two necessary folders

$ mkdir features results

Move the downloaded features to the 'features' folder

Training and evaluation

To train and evaluate the TURN model, run the 'main.py' script

$ python main.py

if you want to continue training from the specfic checkpoint, use '--resume' option

$ python main.py --resume CHECKPOINT_PATH