mccastles / bowie

VQA BoWIE: Bag of Words Image Explanation

VQA BoWIE: Bag of Words Image Explanation

Data missing in repo

• VQA_image_features.h5 - pretrained CNN features (you don't need actual images)
• glove.6B.300d.txt - word embeddings
• v2_Annotations_Train_mscoco.zip - answers data & more
• v2_Questions_Train_mscoco.zip - questions data, images ids & more

Project Structure

.
├── Data
│   ├── Additional
│   │   ├── VQA_image_features.h5 (absent)
│   │   ├── image_ids_vqa.json
│   │   ├── imgid2imginfo.json
│   │   └── VQA_img_features2id.json
│   ├── GloVe
│   │   └── glove.6B.300d.txt (absent)
│   ├── Subset
│   │   └── (empty)
│   └── VQA_Train
│       ├── v2_Annotations_Train_mscoco.zip (absent)
│       └── v2_Questions_Train_mscoco.zip (absent)
├── Gallery
│   └── (some pics)
├── Notebooks
│   ├── scripts
│   │   └── (empty)
│   ├── DataPreparation.ipynb
│   ├── EmbeddingPreparation.ipynb
│   ├── Preprocessing.ipynb
│   └── TrainBOWIMG.ipynb
├── README.md
└── Weights
    └── 443700_emb_ep_3_34.0.pt

Demo

( more examples: Demo section )

Todo

• [Done] Pure BoW vs W2V comarison
• [Done] Linear Layer matrix splitting into textual and visual parts
• [Done] Different metrics instead of a silly accuracy
• Data statistics
• Make sure the same pics ids don't end up in the same train/valid/test split
• Different questions for the same picture, vise versa
• Split questions by type, score for each type (yes/no, number, other)
• Good README with references