README.md

# AToMiC Prebuilt Indexes

## Example Usage:

### Reproduction

Toolkits:
https://github.com/TREC-AToMiC/AToMiC/tree/main/examples/dense_retriever_baselines

```bash
# Skip the encode and index steps, search with the prebuilt indexes and topics directly

python search.py \
    --topics topics/openai.clip-vit-base-patch32.text.validation \
    --index indexes/openai.clip-vit-base-patch32.image.faiss.flat \
    --hits 1000 \
    --output runs/run.openai.clip-vit-base-patch32.validation.t2i.large.trec

python search.py \
    --topics topics/openai.clip-vit-base-patch32.image.validation \
    --index indexes/openai.clip-vit-base-patch32.text.faiss.flat \
    --hits 1000 \
    --output runs/run.openai.clip-vit-base-patch32.validation.i2t.large.trec
```

### Explore AToMiC datasets

```python
import torch
from pathlib import Path
from datasets import load_dataset
from transformers import AutoModel, AutoProcessor

INDEX_DIR='indexes'
INDEX_NAME='openai.clip-vit-base-patch32.image.faiss.flat'
QUERY = 'Elizabeth II'

images = load_dataset('TREC-AToMiC/AToMiC-Images-v0.2', split='train') 
images.load_faiss_index(index_name=INDEX_NAME, file=Path(INDEX_DIR, INDEX_NAME, 'index'))

model = AutoModel.from_pretrained('openai/clip-vit-base-patch32')
processor = AutoProcessor.from_pretrained('openai/clip-vit-base-patch32')

# prebuilt indexes contain L2-normalized vectors
with torch.no_grad():
    q_embedding = model.get_text_features(**processor(text=query, return_tensors="pt"))
    q_embedding = torch.nn.functional.normalize(q_embedding, dim=-1).detach().numpy()

scores, retrieved = images.get_nearest_examples(index_name, q_embedding, k=10)
```