Causal Action--Effect Modeling

The official project repository for the IJCNLP-AACL 2023 paper: Implicit Affordance Acquisition via Causal Action–Effect Modeling in the Video Domain.
NOTE: the code has not been tested.

We built upon the HERO architecture and proposed two pre-training tasks for implicit affordance learning from the video domain.
Most codes are copied/modified from HERO's repository.
The visual frame feature extraction (SlowFast + ResNet-152) closely follows their recipe, which is available at CAE_Video_Feature_Extractor.

If you find this repository helpful, please cite:

@article{yang2023affordance,
  title={Implicit Affordance Acquisition via Causal Action–Effect Modeling in the Video Domain},
  author={Yang, Hsiu-Yu and Silberer, Carina},
  journal={Proceedings of the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 13th International Joint Conference on Natural Language Processing},
  year={2023}
}

Create an environment

conda env create -f env.yml

Quick Start

1. Get the pre-processed features:

To download the pre-processed input features of CAE dataset, please contact me via [email protected].
* Note:

• For the CAE dataset, please refer to this repository.
• To reproduce the whole dataset pre-processing steps, refer to step 5.

After obtain the preprocessed features, you will see the following folder structure:

    ├── eval_table
    │   └── eval_table.json
    ├── test
    │   └── test.json
    ├── train
    │   └── train.json
    ├── val
    │   └── val.json
    ├── txt_db
    │    ├── full
    │    │   ├── data.mdb
    │    │   ├── eval_table.json
    │    │   ├── lock.mdb
    │    │   ├── meta.json
    │    │   ├── test_ids.json
    │    │   ├── train_ids.json
    │    │   ├── val_ids.json
    │    │   ├── vid2frame_sub_len.json
    │    │   ├── neg
    │    │   │    ├── train_video_based_negatives.json
    │    │   │    ├── val_10_video_based_negatives.json
    │    │   │    └── test_video_based_negatives.json
    └── video_db
        ├── full
        │    ├── id2nframe.json
        │    ├── resnet_slowfast_2.0_compressed
        │    │    ├── data.mdb
        │    │    ├── lock.mdb
        │    └── video_feat_info.pkl

2. Run pretraining:

MAM

export PRETRAIN_EXP=model/cae/pretrain_cae_mam
CUDA_VISIBLE_DEVICES=1,2 HOROVOD_CACHE_CAPACITY=0  horovodrun -np 2 python pretrain.py \
--config config/example-pretrain-cae-mam-random-2gpu.json \
--output_dir $PRETRAIN_EXP

MEM

export PRETRAIN_EXP=model/cae/pretrain_cae_mem
CUDA_VISIBLE_DEVICES=1,2 HOROVOD_CACHE_CAPACITY=0  horovodrun -np 2 python pretrain.py \
--config config/example-pretrain-cae-mem-random-2gpu.json \
--output_dir $PRETRAIN_EXP

MULTI-CAE (joint-task training in the round-robin fashion)

*Note: provide the checkpoints of RoBERTa-base under the field checkpoint in example_pretrain-cae-multi-2gpu.json

export PRETRAIN_EXP=model/cae/pretrain_cae_multi
CUDA_VISIBLE_DEVICES=1,2 HOROVOD_CACHE_CAPACITY=0  horovodrun -np 2 python pretrain.py \
--config config/example-pretrain-cae-multi-2gpu.json \
--output_dir $PRETRAIN_EXP

3. Run intrinsic evaluations:

*Note: replace --output_dir and --checkpoint with your own model trained in step 2.

Masked Action Prediction (MAP)

export PRETRAIN_EXP=pretrain_cae_mam
export PATH_TO_EXP=path_to_single_result_verb_exp
export CKPT=98500
export EXP=full
export SPLIT=test

CUDA_VISIBLE_DEVICES=0 nohup horovodrun -np 1 python eval_map.py \
--sub_txt_db $PATH_TO_EXP/txt_db/$EXP \
--vfeat_db $PATH_TO_EXP/video_db/$EXP \
--split $SPLIT \
--subset $EXP \
--task cae_video_sub \
--checkpoint $CKPT \
--model_config config/hero_finetune.json \
--batch_size 64 \
--output_dir $PRETRAIN_EXP

Masked Effect Prediction (MEP)

export PRETRAIN_EXP=pretrain_cae_mem
export PATH_TO_EXP=path_to_single_result_verb_exp
export CKPT=95000
export EXP=full
export SPLIT=test

CUDA_VISIBLE_DEVICES=0 nohup horovodrun -np 1 python eval_map.py \
--sub_txt_db $PATH_TO_EXP=/txt_db/$EXP \
--vfeat_db $PATH_TO_EXP=/video_db/$EXP \
--split $SPLIT \
--subset $EXP \
--negative_sampling_strategy video-based \
--negative_examples neg/test_video_based_negatives.json \
--task cae_video_sub \
--checkpoint $CKPT \
--model_config config/hero_finetune.json \
--batch_size 64 \
--output_dir $PRETRAIN_EXP

4. Run the extrinsic evaluation on PROST in a zero-shot manner:

*Note: replace --output_dir and --checkpoint with your own model trained in step 2.

# Take the checkpoints trained with MAM only for example:
export PRETRAIN_EXP=pretrain_cae_mam
export CKPT=98500
export DATA=path_to_prost_cache

CUDA_VISIBLE_DEVICES=0 nohup python eval_prost.py \
--checkpoint $CKPT \
--model_config config/hero_finetune.json \
--batch_size 64 \
--dataroot $DATA \
--output_dir $PRETRAIN_EXP

(Optional) 5. To reproduce the CAE dataset preprocessing:

A. Create the CAE dataset from scratch: refer to this repository.
B. Extract visual features: refer to this repository.
C. Collect visual features & save into lmdb:

export $PATH_TO_EXP=path_to_single_result_verb_exp ($PATH_TO_CAE/single_result_verb_exp/42)
export $PATH_TO_STORAGE=path_to_videos
export $EXP_NAME=full 

Collect visual features:

python scripts/collect_video_feature_paths.py  \
--feature_dir $PATH_TO_STORAGE/videos/visual_features \
--output $PATH_TO_EXP/video_db/ \
--eval_table $PATH_TO_EXP/eval_table/eval_table.json  \
--dataset $EXP_NAME \
--nproc 0

Convert to lmdb:

python scripts/convert_videodb.py \
--vfeat_info_file $PATH_TO_EXP/video_db/$EXP_NAME/video_feat_info.pkl \
--output $PATH_TO_EXP/video_db \
--dataset $EXP_NAME \
--frame_length 2  \
--compress \
--clip_interval -1

After running the above codes, $PATH_TO_EXP/video_db/$EXP_NAME should have the following folder structure:

    ├── id2nframe.json
    ├── resnet_slowfast_2.0_compressed
    │    ├── data.mdb
    │    ├── lock.mdb
    └── video_feat_info.pkl

D. Tokenize subtitles & save into lmdb:

python scripts/prepro_sub.py \
--annotation $PATH_TO_EXP/cae.json \
--vid2nframe $PATH_TO_EXP/video_db/$EXP_NAME/id2nframe.json \
--eval_table $PATH_TO_EXP/eval_table/eval_table.json \
--frame_length 2 \
--output $PATH_TO_EXP/txt_db/$EXP_NAME
--task cae