Hobin Ryu, Sunghun Kang, Haeyong Kang, and Chang D. Yoo. AAAI 2021. [arxiv]

• Ubuntu 16.04
• CUDA 9.2
• cuDNN 7.4.2
• Java 8
• Python 2.7.12
  • PyTorch 1.1.0
  • Other python packages specified in requirements.txt

$ pip install -r requirements.txt

1. Download the GloVe Embedding from here and locate it at data/Embeddings/GloVe/GloVe_300.json.

2. Extract features from datasets and locate them at data/<DATASET>/features/<NETWORK>.hdf5.

   e.g. ResNet101 features of the MSVD dataset will be located at data/MSVD/features/ResNet101.hdf5.

   I refer to this repo for extracting the ResNet101 features, and this repo for extracting the 3D-ResNext101 features.

3. Split the features into train, val, and test sets by running following commands.

   $ python -m split.MSVD
   $ python -m split.MSR-VTT
   

You can skip step 2-3 and download below files

• MSVD
  • ResNet-101 [train] [val] [test]
  • 3D-ResNext-101 [train] [val] [test]
• MSR-VTT
  • ResNet-101 [train] [val] [test]
  • 3D-ResNext-101 [train] [val] [test]

Clone the evaluation code from the official coco-evaluation repo.

$ git clone https://github.com/tylin/coco-caption.git
$ mv coco-caption/pycocoevalcap .
$ rm -rf coco-caption

$ python extract_negative_videos.py

or you can skip this step as the output files are already uploaded at data/<DATASET>/metadata/neg_vids_<SPLIT>.json

$ python train.py

You can change some hyperparameters by modifying config.py.

Pretrained Models - SGN(R101+RN)

• MSVD: https://drive.google.com/file/d/12Xjd8VdDiyvBxM9sPnnXz87Wa_eVv0ii/view?usp=sharing
• MSR-VTT: https://drive.google.com/file/d/1kx7FBi2UBCgIP7R9ideMpwXY0Gnqn7Yx/view?usp=sharing

*Disclaimer: The models above do not have the same weight as the models used in the paper (I trained them again because I lost).

$ python evaluate.py --ckpt_fpath <MODEL_CHECKPOINT_PATH>

The source-code in this repository is released under MIT License.