A3PRVR（2026AAAI）

December 21, 2025 · View on GitHub

Environment Setup

Strict version replication is not required. If version conflicts occur, install the versions supported by your local machine.

# Create and activate conda environment
conda create -n A3PRVR python=3.9
conda activate A3PRVR

# Install dependencies
cd root_path  # Replace with your project root directory path
pip install -r requirements.txt

Data Dependencies

Only the Charades-STA dataset needs to be downloaded for reproduction (smaller in size than ActivityNet and TVR). Complete dependency data download:
Charades-STA Data Download Link

Activitynet Data Download Link

TVR Data Download Link

Pretrained Model Preparation (Optional)

Only when you need to enable the Hard negative sample generation module (the loss function corresponding to Figure 4 in the paper), you must first download the RoBERTa pretrained weights:

cd root_path  # Project root directory
python3 utils/download_roberta.py

Training

Each dataset has a separate training script. Execute the following commands to start training:

Charades-STA

cd root_path
bash do_charades.sh

ActivityNet

cd root_path
bash do_activitynet.sh

TVR

cd root_path
bash do_tvr.sh

Training Parameter Explanation

  • If not enabling Hard negative sample generation: Do not pass the neg_query_loss parameter, and you can delete all related parameters in the script:
    • neg_query_loss
    • neg_query_loss_weight
    • neg_action_num
    • neg_object_num
    • neg_query_loss_branch
  • If enabling: Retain the above parameters and adjust their values as needed.

Testing

The testing process reuses the training scripts, and only two additional parameters need to be added:

Testing Steps

  1. Download the pretrained weights (ckpt):
    Charades-STA Model Checkpoint Download Link

    ActivityNet Model Checkpoint Download Link

    TVR Model Checkpoint Download Link

  2. Modify the training script and add the following parameters:

    • only_eval: Enable evaluation-only mode
    • eval_ckpt="your_ckpt_path": Replace with the storage path of the downloaded ckpt file

Testing Example (Taking Charades-STA as an Example)

cd root_path
# Assume the modified script already includes the only_eval and eval_ckpt parameters
bash do_charades.sh