Attention-LLaVA

A hot-pluggable tool for visualizing LLaVA's attention.

Example

Dialog

Heatmap

Attention

The attention on the image of the first 5 words which are "The image features a woman".

Usage

1. Install LLaVA from Link.
2. Put the attention.py into:

LLaVA-main/llava/eval

3. Run by this command:

cd LLaVA-main
python -m llava.eval.attention \
  --checkpoint path/to/llava/checkpoint \
  --image path/to/image \
  --layer 32 \
  --output path/to/output/result \
  --max-length 64

https://github.com/junyangwang0410/Attention-LLaVA/blob/main/Mobile-Agent.mp4