Florence-2 Recap-DataComp LoRA Adapter
This repository contains a LoRA adapter trained on the UCSC-VLAA/Recap-DataComp-1B dataset for the Florence-2-base-FT model. It's designed to enhance the model's captioning capabilities, providing more detailed and descriptive image captions.
Usage
To use this LoRA adapter, you'll need to load it along with the Florence-2-base model using the PEFT library. Here's an example of how to use it:
from PIL import Image
from transformers import AutoProcessor, AutoModelForCausalLM
from peft import PeftModel, PeftConfig
import requests
def caption(image):
base_model = AutoModelForCausalLM.from_pretrained("microsoft/Florence-2-base-ft", trust_remote_code=True)
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base-ft", trust_remote_code=True)
prompt = "<MORE_DETAILED_CAPTION>"
adapter_name = "NikshepShetty/Florence-2-Recap-DataComp"
model = PeftModel.from_pretrained(base_model, adapter_name, trust_remote_code=True)
inputs = processor(text=prompt, images=image, return_tensors="pt")
generated_ids = model.generate(
input_ids=inputs["input_ids"],
pixel_values=inputs["pixel_values"],
max_new_tokens=1024,
do_sample=False,
num_beams=3
)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
parsed_answer = processor.post_process_generation(generated_text, task="<MORE_DETAILED_CAPTION>", image_size=(image.width, image.height))
print(parsed_answer)
url = "https://huggingface.co./datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg?download=true"
image = Image.open(requests.get(url, stream=True).raw)
caption(image)
This code demonstrates how to:
- Load the base Florence-2 model
- Load the LoRA adapter
- Process an image and generate a detailed caption
Note: Make sure you have the required libraries installed: transformers, peft, einops, flash_attn, timm, Pillow, and requests.
Evaluation results
Our LoRA adapter shows improvements over the base Florence-2 model across all metrics for MORE_DETAILED_CAPTION tag for 1000 images on the foundation-multimodal-models/DetailCaps-4870 dataset:
Metric | Base Model | Adapted Model | Improvement |
---|---|---|---|
CAPTURE | 0.546 | 0.553 | +1.3% |
METEOR | 0.213 | 0.240 | +12.7% |
BLEU | 0.110 | 0.150 | +36.4% |
CIDEr | 0.031 | 0.035 | +12.9% |
ROUGE-L | 0.275 | 0.294 | +6.9% |
These results demonstrate that our LoRA adapter enhances the image captioning capabilities of the Florence-2 base model, particularly in generating more detailed and accurate captions.
- Downloads last month
- 2
Dataset used to train NikshepShetty/Florence-2-Recap-DataComp
Evaluation results
- meteor on foundation-multimodal-models/DetailCaps-4870self-reported0.240
- bleu on foundation-multimodal-models/DetailCaps-4870self-reported0.150
- cider on foundation-multimodal-models/DetailCaps-4870self-reported0.035
- capture on foundation-multimodal-models/DetailCaps-4870self-reported0.553
- rouge-l on foundation-multimodal-models/DetailCaps-4870self-reported0.294