---
datasets:
- liuhaotian/LLaVA-Pretrain
- liuhaotian/LLaVA-Instruct-150K
language:
- en
tags:
- llava
- phi
license: mit
library_name: transformers
widget:
- text: "What animal is it?"
  src: "https://huggingface.co./datasets/mishig/sample_images/resolve/main/tiger.jpg"
- text: "Where is it?"
  src: "https://huggingface.co./datasets/mishig/sample_images/resolve/main/palace.jpg"
---

# Multi-crop LLaVA-3b

<a target="_blank" href="https://colab.research.google.com/drive/1W7JQrFXwFunAY1XvS31mwC7mrXBgGD_M">
  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
</a>

## Model details

The core idea behind multi-crop LLaVA is that instead of N visual token embeddings per image, I generate one token embedding per N parts of the image. 
Having high-quality embeddings for smaller parts of the image helps to extract more details and understand the scene better.

For every crop of the image, I generate an embedding from the full SigLIP encoder (size [1, 1152]) and then push all N embeddings through the LLaVA adapter, which 
gives the token embedding of size [N, 2560]. Right now, the tokens do not contain explicit information about their position in the original image. I plan to add it later.

MC-LLaVA-3b was fine-tuned from [Dolphin 2.6 Phi](https://huggingface.co./cognitivecomputations/dolphin-2_6-phi-2) using vision tower from
[SigLIP 400M](https://huggingface.co./timm/ViT-SO400M-14-SigLIP-384). 

The context length during training was 1200 tokens, as the L4 GPUs I used didn't allow me to get more.

As Dolphin 2.6 Phi, LLaVA-3b uses ChatML prompt format:

```
<|im_start|>system
You are Dolphin, a helpful AI assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
```

## How to use

**Install dependencies**

```bash
!pip install -q open_clip_torch timm einops
```

**Download modeling files**

```python
from huggingface_hub import hf_hub_download

hf_hub_download(repo_id="visheratin/LLaVA-3b", filename="configuration_llava.py", local_dir="./", force_download=True)
hf_hub_download(repo_id="visheratin/LLaVA-3b", filename="configuration_phi.py", local_dir="./", force_download=True)
hf_hub_download(repo_id="visheratin/LLaVA-3b", filename="modeling_llava.py", local_dir="./", force_download=True)
hf_hub_download(repo_id="visheratin/LLaVA-3b", filename="modeling_phi.py", local_dir="./", force_download=True)
hf_hub_download(repo_id="visheratin/LLaVA-3b", filename="processing_llava.py", local_dir="./", force_download=True)
```

**Create a model**

```python
from modeling_llava import LlavaForConditionalGeneration
import torch

model = LlavaForConditionalGeneration.from_pretrained("visheratin/LLaVA-3b", torch_dtype=torch.float16)
model = model.to("cuda")
```

**Create processors**

```python
from transformers import AutoTokenizer
from processing_llava import LlavaProcessor, OpenCLIPImageProcessor

tokenizer = AutoTokenizer.from_pretrained("visheratin/LLaVA-3b")
image_processor = OpenCLIPImageProcessor(model.config.preprocess_config)
processor = LlavaProcessor(image_processor, tokenizer)
```

**Set image and text**

```python
from PIL import Image
import requests

image_file = "https://images.unsplash.com/photo-1439246854758-f686a415d9da"
raw_image = Image.open(requests.get(image_file, stream=True).raw)

prompt = """<|im_start|>system
A chat between a curious human and an artificial intelligence assistant.
The assistant gives helpful, detailed, and polite answers to the human's questions.
The assistant does not hallucinate and pays very close attention to the details.<|im_end|>
<|im_start|>user
<image>
Describe the image.<|im_end|>
<|im_start|>assistant
"""
```

**Process inputs**

```python
inputs = processor(prompt, raw_image, model, return_tensors='pt')

inputs['input_ids'] = inputs['input_ids'].to(model.device)
inputs['attention_mask'] = inputs['attention_mask'].to(model.device)
```

**Generate the data**

```python
import torch

with torch.inference_mode():
  output = model.generate(**inputs, max_new_tokens=200, do_sample=True, temperature=0.4, pad_token_id=tokenizer.eos_token_id, eos_token_id=tokenizer.eos_token_id)
```

## Benchmarks

- TextVQA - 38.59%
- GQA - 49.6%
- VQAv2 - 64.24%
- VizWiz - 24.88%
- POPE - 80.59%
- V*-bench - 52.25% (OCR - 46.66%, GPT4V-hard - 41.17%, direct attributes - 43.48%, relative position - 65.79%)

## License

The model is licensed under MIT license, but since the data used for model training is largely synthetic, you should also follow OpenAI and Google Gemini terms of service. 
Which means don't create competitor models for them.

## Acknowledgments

Thanks to [ML Collective](https://mlcollective.org/) for providing credits for computing resources.