Cogflorence 2.2 Large

🚀 microsoft/Florence-2-large在Ejafa/ye-pop数据集上微调并使用CogVLM2添加字幕

本仓库包含 microsoft/Florence-2-large 模型的微调版本。该模型在 Ejafa/ye-pop 数据集的40,000张图像子集上进行了微调，其字幕使用 THUDM/cogvlm2-llama3-chat-19B 生成。

🚀 快速开始

要使用此模型，你可以直接从Hugging Face模型中心加载它：

from transformers import AutoModelForCausalLM, AutoProcessor, AutoConfig
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = AutoModelForCausalLM.from_pretrained("thwri/CogFlorence-2.2-Large", trust_remote_code=True).to(device).eval()
processor = AutoProcessor.from_pretrained("thwri/CogFlorence-2.2-Large", trust_remote_code=True)
# Function to run the model on an example
def run_example(task_prompt, image):
    prompt = task_prompt
    # Ensure the image is in RGB mode
    if image.mode != "RGB":
        image = image.convert("RGB")
    inputs = processor(text=prompt, images=image, return_tensors="pt").to(device)
    generated_ids = model.generate(
        input_ids=inputs["input_ids"],
        pixel_values=inputs["pixel_values"],
        max_new_tokens=1024,
        num_beams=3,
        do_sample=True
    )
    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=False)[0]
    parsed_answer = processor.post_process_generation(generated_text, task=task_prompt, image_size=(image.width, image.height))
    return parsed_answer
from PIL import Image
import requests
import copy
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)
result = run_example("<MORE_DETAILED_CAPTION>" , image)
print(result)
# {'<MORE_DETAILED_CAPTION>': 'A vivid portrayal of a classic Volkswagen Beetle parked on a cobblestone street. The car is painted a vibrant turquoise, contrasting with the muted yellow of the building behind it. The building has two wooden doors, one with a white frame and the other with a dark brown finish. The sky is clear, and the sun casts a warm glow on the scene, highlighting the car's details. The image evokes a nostalgic and nostalgic mood, capturing a moment in time without posed elements.'}

✨ 主要特性

• 基于 microsoft/Florence-2-large 模型进行微调，提升了图像字幕生成能力。
• 使用 THUDM/cogvlm2-llama3-chat-19B 生成字幕，并使用 google/gemma-2-9b 进行后处理，去除模糊表述。

📚 详细文档

训练详情

• 视觉编码器：训练期间视觉编码器被冻结。
• 批量大小：64
• 梯度累积步数：16
• 学习率：5.12e-05
• 优化器：AdamW
• 调度器：多项式
• 训练轮数：8.36

数据集

微调过程使用了 Ejafa/ye-pop 数据集的40,000张图像子集。该数据集包含各种不同主题的图像，为提高模型的字幕生成能力提供了强大的训练基础。

字幕生成

字幕使用 THUDM/cogvlm2-llama3-chat-19B 生成，然后使用 google/gemma-2-9b 进行后处理以去除模糊表述。

📄 许可证

本项目采用MIT许可证。