Prerequisites

1. Install MMF following the installation docs.

2. Convert data into MMF format. Following the prerequisites part of Hateful Memes Dataset.
   There will be a path that shows where is data of .jsonl and images. Copy it! It will be used in step 4.

3. Install requirement.txt

pip install -r requirements.txt

4. Download the pre-trained models.
   Address: https://pan.baidu.com/s/1KzRPwRo9BQORBdYLhoed_A Password: 1122

5. Modify path
   In triad_tuples.py, change annotations_fold and images_fold to the path of data of .jsonl and images.
   In predict_training.sh, change annotations_fold to the path of data of .jsonl.
   The path is copied in step 2.

Run

If you want to predict with given models (fast)

1. Modify predict_models.sh
   In predict_models.sh, change predict_cuda_num to index of cuda that you want to predict.
2. Grant run permission

chmod a+x predict_models.sh

or

sudo chmod a+x predict_models.sh

3. Run predict_models.sh

./predict_models.sh

---

If you want to predict with training (slow)

1. Modify predict_training.sh
   In predict_training.sh, change training_cuda_num and predict_cuda_num to index of cuda that you want to train and predict.
2. Grant run permission

chmod a+x predict_training.sh

or

sudo chmod a+x predict_training.sh

3. Run predict_training.sh

./predict_training.sh

Result

final_result.csv in the current directory is the final submission.

P.S.

If it is interrupted when run predict_training.sh, change train.jsonl.bak and dev_unseen.json.bak to train.jsonl and dev_unseen.jsonl before run it again or run any other MMF models.
Because I use KFold technique, train.jsonl and dev_unseen.jsonl are changed during runing.