31st ACM International Conference on Information & Knowledge Management, Georgia, USA, 2022
- Ubuntu 18.04.5 LTS
- CUDA 11.3
- Python 3.8.12
Download our subttitle and meta data, training and test data from our homepage. Then organize the dataset folder as follows:
./data
|--YOUTUBE_data
|---meta_data
|---transcripts
Download and embed the thumbnail images using InceptionNet, or you can download embedding vector directly from our homepage.
$ python thumbnail_download.py
Embed all the features: headline, tags, style, thumbnail by running:
$ python meta_embedding.py
Embedding data is exported to meta_embedding.hdf5 that have structure:
{video_id: 1-D list,
thumbnail: 2-D matrix,
headline: 2-D matrix,
style: 2-D matrix,
tags:2-D matrix,
y: 1-D array}
Note: This process may take time since the pre-trained word2vec model used to embedding subtitle.
Using video index in the *big_meta.csv file and and put subtitle in that order. Change the input csv file name at line 54. This file has to have video_id and subtitle columns. Then, run:
$ python subtitle_pre.py
Modify the ML method in config/config_meta.yaml file at line 3. Start training by runining:
$ python meta_train.py
Modify the ML method in config/config_word2vec.yaml file at line 3. Start training by runining:
$ python subtitle_word2vec_train.py
Follow the README document in pretrain folder to create subtitles' embeddings. The run:
$ python train.py
Our pre-trained model can be obtained here
Star if you find it useful. ⭐