IntentVizor

Project Structure

./solvers # training solvers based on different settings.
./scripts # scripts for running the training
./models # models used in the project
./utils # utility code for the video summarization
./notebooks # notebooks for data preprocessing
./runners # runners for training the models
./factory # factory code for model and solvers (factory mode)
./loggers # training loggers
./exceptions # exceptions
./deployment # code for deploying the models.
./data # data loaders
./cache # code for the cache.
./evaluation # evaluation metrics

Installation

pip install requirements.txt

Running

The entrypoint of our project is the file train_qfvsum.py and train_sqfvsum.py. We also provide two bash scripts in scripts directory.

bash ./scripts/train_qfvsum.sh # train models on textual query dataset

bash ./scripts/train_vqfvsum.sh # train models on visual query dataset

Guidance on Reading the Code

Topic-Net and Intent-Net

In the early stage of developing the models, we use the term "topic" and "topic-net" to refer the currently used terms "intent" and "intent-net". As a result, the "topic" means "intent" in our paper. We do not change the term because the change of the name may make our trained model checkpoint fail.

Model Directory

We divide the code into six sub-directories, i.e, baselines, feature_encoder, gcn, intent_net, score_net and visual_query. The model.py and shot_model.py are the entire models for textual and visual query datasets. The gcn directory stores the code for implementing GCN. Some of the code is collected from G-TAD.

Visual Query Dataset Generation

We put the preprocessing code for visual query dataset in the notebooks directory.

Checkpoint

The checkpoint is available on Google Drive.