Troubleshooting

Welcome to the ban-vqa troubleshooting wiki!

Preprocessing checklist

data
├── annotations
│   ├── captions_train2017.json
│   └── captions_val2017.json
├── cache
│   ├── train_target.pkl
│   ├── trainval_ans2label.pkl
│   ├── trainval_label2ans.pkl
│   ├── val_target.pkl
│   ├── vg_train_adaptive_target.pkl
│   ├── vg_train_target.pkl
│   ├── vg_val_adaptive_target.pkl
│   └── vg_val_target.pkl
├── dictionary.pkl
├── glove
│   ├── glove.6B.100d.txt
│   ├── glove.6B.200d.txt
│   ├── glove.6B.300d.txt
│   └── glove.6B.50d.txt
├── glove6b_init_300d.npy
├── image_data.json
├── question_answers.json
├── test2015.hdf5
├── test2015_imgid2idx.pkl
├── train.hdf5
├── train_imgid2idx.pkl
├── v2_OpenEnded_mscoco_test2015_questions.json
├── v2_OpenEnded_mscoco_train2014_questions.json
├── v2_OpenEnded_mscoco_val2014_questions.json
├── val.hdf5
└── val_imgid2idx.pkl

Out of memory issue

The hyperparameters of this repository demand you should have 4 GPUs with 12 GB memory each. If you have only 1 or 2 GPUs and want to run the code to train (if you want to evaluate the downloaded pretrained model, reducing batch_size in test.py does not affect the results), the first viable option might be to reduce batch_size from 256 to smaller one. However this change can cause a minor performance drop, so you may want to adjust learning rate schedule in this function of train.py. Unfortunately, we have not yet explored the hyperparameters for this scenario. If you find any good one, please share your shining settings.