# YING-VLM

We open-sourced the trained checkpoint and infernce code of [YING-VLM](https://huggingface.co/MMInstruction/YingVLM) at huggingface, which is trained on [M3IT](https://huggingface.co/datasets/MMInstruction/M3IT) dataset.


# Example of Using YING-VLM

Please install the following packages:
- torch==2.0.0
- transformers==4.31.0



Infernce example:

```python
from transformers import AutoProcessor, AutoTokenizer
from PIL import Image
import torch

from modelingYING import VLMForConditionalGeneration


# set device
device="cuda:0"

# set prompt template
prompt_template = """
<human>:
{instruction}
{input}
<bot>:
"""

# load processor and tokenizer
processor = AutoProcessor.from_pretrained("MMInstruction/YingVLM")
tokenizer = AutoTokenizer.from_pretrained("MMInstruction/YingVLM") # ziya is not available right now 


# load model
model = VLMForConditionalGeneration.from_pretrained("MMInstruction/YingVLM")
model.to(device,dtype=torch.float16)


# prepare input
image = Image.open("./imgs/night_house.jpeg")
instruction = "Scrutinize the given image and answer the connected question."
input = "What is the color of the couch?"
prompt = prompt_template.format(instruction=instruction, input=input)


# inference
inputs = processor(images=image,  return_tensors="pt").to(device, torch.float16)
text_inputs = tokenizer(prompt, return_tensors="pt")
inputs.update(text_inputs)



generated_ids = model.generate(**{k: v.to(device) for k, v in inputs.items()}, img_num=1, max_new_tokens=128, do_sample=False)
generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0].split("\n")[0] # \n is the end token

print(generated_text)
# The couch in the living room is green.





```



# Refernce

If you find our work useful, please kindly cite
```bib
@article{li2023m3it,
  title={M$^3$IT: A Large-Scale Dataset towards Multi-Modal Multilingual Instruction Tuning},
  author={Lei Li and Yuwei Yin and Shicheng Li and Liang Chen and Peiyi Wang and Shuhuai Ren and Mukai Li and Yazheng Yang and Jingjing Xu and Xu Sun and Lingpeng Kong and Qi Liu},
  journal={arXiv preprint arXiv:2306.04387},
  year={2023}
}
```