Code for the ICML 2021 paper: Optimization Planning for 3D ConvNets.
Authors: Zhaofan Qiu, Ting Yao, Chong-Wah Ngo, Tao Mei
The provided codes have been tested with Python-3.9.5 & Pytorch-1.9.0 on four Tesla-V100s.
├─ base_config # Pre-set config file for each dataset
├─ dataset # Video lists (NOT provided) and code to load video data
├─ jpgs # Images for README
├─ layers # Custom network layers
├─ model # Network architectures
├─ record # Config file for each run
├─ utils # Basic functions
├─ extract_score_3d.py # Main script to extract predicted score
├─ helpers.py # Helper functions for main scripts
├─ merge_score.py # Main script to merge scores from different clips
├─ train_3d.py # Main script to launch a training using given strategy
├─ train_3d_op.py # Main script to launch a searching of best strategy
└─ run.sh # Shell script for training-extracting-merging pipeline
- Pre-process the target dataset and put the lists in to the
dataset
folder. Codes in dataset/video_dataset.py can load three video formats (raw video, jpeg frames and video LMDB) and can be simply modified to support the custom format. - Make config file in the
record
folder. The config examples includeop-*.yml
for pre-searched strategy,kinetics-*.yml
for simple strategy on Kinetics-400, - Run run.sh for the training-extracting-merging pipeline or replace train_3d.py with train_3d_op.py for searching the optimal strategy.
Add more explainations and examples.
Please feel free to email to Zhaofan Qiu if you have any question regarding the paper or any suggestions for further improvements.
If you find this code helpful, thanks for citing our work as
@inproceedings{qiu2021optimization,
title={Optimization Planning for 3D ConvNets},
author={Qiu, Zhaofan and Yao, Ting and Ngo, Chong-Wah and Mei, Tao},
booktitle={Proceedings of the 38th International Conference on Machine Learning (ICML)},
publisher={PMLR},
year={2021}
}
Please also pay attention to the citations of the included networks/algorithms.