# AFRAgent — Epoch 7 checkpoint

This model is a checkpoint from **AFRAgent** ([paper](https://arxiv.org/abs/2512.00846), WACV 2026): an Adaptive Feature Renormalization–based GUI agent for smartphone automation, built on InstructBLIP.

- **Architecture:** AnyResAdaIn (any-resolution adaptive feature renormalization)
- **Base model:** Salesforce/instructblip-flan-t5-xl
- **Training:** Fine-tuned on Android-in-the-Wild (AITW) — 7 epochs, `all_data_any_res_adain_finetuning` (bs128, ip512, op256, ep12 run; this is checkpoint at step 56266 ≈ epoch 7)

## How to load

Requires the [AFRAgent](https://github.com/neerajanand321/AFRAgent/tree/main) codebase for the custom `AnyResAdaIn` class.

```python
# Clone AFRAgent and add to path, then:
from models.any_res_adain_queries_fusion import AnyResAdaIn
from transformers import InstructBlipProcessor, AutoTokenizer

model = AnyResAdaIn.from_pretrained("neeraj321/AFRAgent_pure_multimodel")
processor = InstructBlipProcessor.from_pretrained("neeraj321/AFRAgent_pure_multimodel")
tokenizer = AutoTokenizer.from_pretrained("neeraj321/AFRAgent_pure_multimodel")
```

For evaluation with the AFRAgent script:

```bash
python instructblip_main.py \
  --evaluate_dir neeraj321/AFRAgent_pure_multimodel \
  --train_any_res_adain True \
  --use_high_res True \
  --data_root dataset/aitw/general/general \
  --input_len 512 --output_len 256 --eval_bs 64
```

## License

MIT

## Citation

```bibtex
@article{anand2025afragent,
  title={AFRAgent: An Adaptive Feature Renormalization Based High Resolution Aware GUI agent},
  author={Anand, Neeraj and others},
  journal={WACV},
  year={2026}
}
```