LifeQA data and code

This repo contains the data and PyTorch code that accompanies our LREC 2020 paper:

LifeQA: A Real-life Dataset for Video Question Answering

Santiago Castro, Mahmoud Azab, Jonathan C. Stroud, Cristina Noujaim, Ruoyao Wang, Jia Deng, Rada Mihalcea

More information is available at the LifeQA website.

Setup

To run it, setup a new environment with Conda and activate it:

conda env create -f environment.yml
conda activate lifeqa

Data

The dataset is under data/, in lqa_train.json, lqa_dev.json, and lqa_test.json. Even though it's divided into train/dev/test, for most experiments we merge them and use a five-fold cross-validation, with the folds indicated in data/folds.

Visual features

You can download the already extracted features or do the following to extract them yourself.

1. Download the videos. Due to YouTube's Terms of Service, we can't provide the video files. However, we provide the IDs and timestamps to obtain the same data. Download the YouTube videos indicated in the field parent_video_id from the JSON files, cut them based on the fields start_time and end_time, and save them based on the JSON key (e.g., 213) to data/videos, placing the files there without subdirectories.

2. Run save_frames.sh to extract the frames in the video files:

   bash feature_extraction/save_frames.sh
   

3. Download pretrained weights from Sports1M for C3D and save it in data/features/c3d.pickle.

4. To extract the features (e.g. from an ImageNet-pretrained ResNet-152) and save them in big H5 files:

   mkdir data/features
   python feature_extraction/extract_features.py resnet
   

Baselines

Check the scripts under run_scripts to run the available baselines.

TVQA

Running the TVQA baseline is different from running the rest of the baselines.

We copied TVQA's repo content from commit 2c98044 into the TVQA/ folder.

Changes from upstream

It has been changed to support 4 answer choices instead of 5. Some other minor modifications have been done as well.

Setup

1. Convert LifeQA dataset to TVQA format

   python scripts/to_tvqa_format.py
   

2. Enter TVQA/ directory:

   cd TVQA/
   

3. Setup the interpreter:

   conda env create -f environment.yml
   conda activate tvqa
   

4. Do some pre-processing:

   python preprocessing.py --data_dir ../data/tvqa_format
   
   for i in 0 1 2 3 4; do
      python preprocessing.py --data_dir ../data/tvqa_format/fold${i}
   done
   
   mkdir cache_lifeqa
   python tvqa_dataset.py \
     --input_streams sub \
     --no_ts \
     --vcpt_path ../data/tvqa_format/det_visual_concepts_hq.pickle \
     --train_path ../data/tvqa_format/lqa_train_processed.json \
     --valid_path ../data/tvqa_format/lqa_dev_processed.json \
     --test_path ../data/tvqa_format/lqa_test_processed.json \
     --word2idx_path cache_lifeqa/word2idx.pickle \
     --idx2word_path cache_lifeqa/idx2word.pickle \
     --vocab_embedding_path cache_lifeqa/vocab_embedding.pickle
   

Train and test on LifeQA dataset from scratch

For 5-fold cross-validation:

for i in 0 1 2 3 4; do
    python main.py \
      --input_streams sub vcpt \
      --no_ts \
      --vcpt_path ../data/tvqa_format/det_visual_concepts_hq.pickle \
      --train_path ../data/tvqa_format/fold${i}/train_processed.json \
      --valid_path ../data/tvqa_format/fold${i}/validation_processed.json \
      --test_path ../data/tvqa_format/fold${i}/test_processed.json \
      --word2idx_path cache_lifeqa/word2idx.pickle \
      --idx2word_path cache_lifeqa/idx2word.pickle \
      --vocab_embedding_path cache_lifeqa/vocab_embedding.pickle

    python test.py --model_dir $(ls -t results/ | head -1) --mode test
done

For train, dev, and test partitions:

python main.py \
  --input_streams sub vcpt \
  --no_ts \
  --vcpt_path ../data/tvqa_format/det_visual_concepts_hq.pickle \
  --train_path ../data/tvqa_format/lqa_train_processed.json \
  --valid_path ../data/tvqa_format/lqa_dev_processed.json \
  --test_path ../data/tvqa_format/lqa_test_processed.json \
  --word2idx_path cache_lifeqa/word2idx.pickle \
  --idx2word_path cache_lifeqa/idx2word.pickle \
  --vocab_embedding_path cache_lifeqa/vocab_embedding.pickle

python test.py --model_dir $(ls -t results/ | head -1) --mode test

Train on TVQA dataset

python preprocessing.py

mkdir cache_original
python tvqa_dataset.py \
  --input_streams sub \
  --no_ts \
  --word2idx_path cache_original/word2idx.pickle \
  --idx2word_path cache_original/idx2word.pickle \
  --vocab_embedding_path cache_lifeqa/vocab_embedding.pickle
python main.py \
  --input_streams sub vcpt \
  --no_ts
RESULTS_FOLDER_NAME=$(ls -t results/ | head -1)

The result from this part was saved in results_2019_05_16_23_02_15 in Google Drive. Note it corresponds to S+V+Q, with cpt as the video feature and w/o ts.

Test on LifeQA dataset

For 5-fold cross-validation:

for i in 0 1 2 3 4; do
  python test.py \
    --vcpt_path ../data/tvqa_format/det_visual_concepts_hq.pickle \
    --test_path ../data/tvqa_format/fold${i}/test_processed.json \
    --model_dir "${RESULTS_FOLDER_NAME}" \
    --mode test
done

For the test partition:

python test.py \
  --vcpt_path ../data/tvqa_format/det_visual_concepts_hq.pickle \
  --test_path ../data/tvqa_format/lqa_test_processed.json \
  --model_dir "${RESULTS_FOLDER_NAME}" \
  --mode test

Fine-tune on LifeQA dataset

For 5-fold cross-validation:

python main.py \
  --input_streams sub vcpt \
  --no_ts \
  --vcpt_path ../data/tvqa_format/det_visual_concepts_hq.pickle \
  --train_path ../data/tvqa_format/fold${i}/train_processed.json \
  --valid_path ../data/tvqa_format/fold${i}/validation_processed.json \
  --test_path ../data/tvqa_format/fold${i}/test_processed.json \
  --word2idx_path cache_original/word2idx.pickle \
  --idx2word_path cache_original/idx2word.pickle \
  --vocab_embedding_path cache_original/vocab_embedding.pickle \
  --pretrained_model_dir "${RESULTS_FOLDER_NAME}" \
  --new_word2idx_path cache_lifeqa/word2idx.pickle

For train, dev, and test partitions:

python main.py \
  --input_streams sub vcpt \
  --no_ts \
  --vcpt_path ../data/tvqa_format/det_visual_concepts_hq.pickle \
  --train_path ../data/tvqa_format/lqa_train_processed.json \
  --valid_path ../data/tvqa_format/lqa_dev_processed.json \
  --test_path ../data/tvqa_format/lqa_test_processed.json \
  --word2idx_path cache_original/word2idx.pickle \
  --idx2word_path cache_original/idx2word.pickle \
  --vocab_embedding_path cache_original/vocab_embedding.pickle \
  --pretrained_model_dir "${RESULTS_FOLDER_NAME}" \
  --new_word2idx_path cache_lifeqa/word2idx.pickle

Issues

If you encounter issues while using our data or code, please open an issue in this repo.

Citation

BibTeX entry.