A demo colab notebook is available here
Download the zip file from here
Unzip files to hateful-memes/datasets
It supports 2 type of models, ViLBERT CC and Visual BERT COCO.
Methods include original, text augmentation (back translation), image captioning, and object labels.
Model Key | Method | Config |
---|---|---|
vilbert | original | configs/vilbert_original.yaml |
vilbert | text augmentation | configs/vilbert_back_translation.yaml |
vilbert | object labels | configs/vilbert_objects.yaml |
vilbert | object labels + text augmentation | configs/vilbert_objects_back_translation.yaml |
vilbert | image captioning | configs/vilbert_caption.yaml |
vilbert | image captioning + text augmentation | configs/vilbert_caption_back_translation.yaml |
vilbert | image captioning + object labels | configs/vilbert_caption_objects.yaml |
vilbert | image captioning + object labels + text augmentation | configs/vilbert_caption_objects_back_translation.yaml |
visual_bert | original | configs/visual_bert_original.yaml |
visual_bert | text augmentation | configs/visual_bert_back_translation.yaml |
visual_bert | object labels | configs/visual_bert_objects.yaml |
visual_bert | object labels + text augmentation | configs/visual_bert_objects_back_translation.yaml |
visual_bert | image captioning | configs/visual_bert_caption.yaml |
visual_bert | image captioning + text augmentation | configs/visual_bert_caption_back_translation.yaml |
visual_bert | image captioning + object labels | configs/visual_bert_caption_objects.yaml |
visual_bert | image captioning + object labels + text augmentation | configs/visual_bert_caption_objects_back_translation.yaml |
$ python tools/run.py config=<CONFIG> model=<MODEL KEY> dataset=hateful_memes
This will save the training outputs to an experiment folder under ./save
directory.