REILX/llava-Qwen2-7B-Instruct-Chinese-CLIP

模型 llava-Qwen2-7B-Instruct-Chinese-CLIP 增强中文文字识别能力和表情包内涵识别能力，接近gpt4o、claude-3.5-sonnet的识别水平！

<img src="./images/llavaqwen2b_chinese-clip.jpg" alt="logo" style="display: block; margin: 0 auto;" width="300" height="300">

1. 1.模型结构：</br> llava-Qwen2-7B-Instruct-Chinese-CLIP = Qwen/Qwen2-7B-Instruct + multimodalprojector + OFA-Sys/chinese-clip-vit-large-patch14-336px</br>

1. 1.微调模块
2. 2.visiontower和languagemodel的qproj, kproj, vproj, oproj, gateproj, upproj, down_proj模块进行lora训练</br>
3. 3.mmp层全量训练</br>

1. 1.微调参数
2. 2.lorar=32,loraalpha=64,numtrainepochs=5,perdevicetrainbatchsize=1,gradientaccumulationsteps=8,highlr=1e-3,lowlr=2e-5,modelmaxlength=2048.</br>
3. 3.训练时长：5小时12分钟

1. 1.数据集</br> 使用gemini-1.5-pro, gemini-1.5-flash, yi-vision, gpt4o,claude-3.5-sonnet模型描述emo-visual-data和ChineseBQB数据集。</br> 文本描述信息通过REILX/chinese-meme-description-dataset 下载</br> 图像可通过emo-visual-data, ChineseBQB下载</br> 图片数据总量1.8G，约10835张中文表情包图片。文字总量42Mb，约24332个图像文本对描述信息。

1. 1.效果展示</br> 以下测试结果显示模型能识别图像中的文字信息，且能正确识别表情包想要表达的内涵。对比REILX/llava-1.5-7b-hf-meme-lora模型中也测试了原始llava-1.5-7b-hf模型的输出，模型无法正确识别图像中的文本信息。</br> <img src="./images/llava-qwen-2-7b-OFA-Syschinese-clip-memechinesebqbmerged0708fp16/llava-qwen2-7b-OFA-Syschinese-clip-fp16-01.PNG" width="800" height="400"> <img src="./images/llava-qwen-2-7b-OFA-Syschinese-clip-memechinesebqbmerged0708fp16/llava-qwen2-7b-OFA-Syschinese-clip-fp16-02.PNG" width="800" height="400"> <img src="./images/llava-qwen-2-7b-OFA-Syschinese-clip-memechinesebqbmerged0708fp16/llava-qwen2-7b-OFA-Syschinese-clip-fp16-03.PNG" width="800" height="400"> <img src="./images/llava-qwen-2-7b-OFA-Syschinese-clip-memechinesebqbmerged0708fp16/llava-qwen2-7b-OFA-Syschinese-clip-fp16-04.PNG" width="800" height="400"> <img src="./images/llava-qwen-2-7b-OFA-Syschinese-clip-memechinesebqbmerged0708fp16/llava-qwen2-7b-OFA-Syschinese-clip-fp16-05.PNG" width="800" height="400"> <img src="./images/llava-qwen-2-7b-OFA-Syschinese-clip-memechinesebqbmerged0708fp16/llava-qwen2-7b-OFA-Syschinese-clip-fp16-06.PNG" width="800" height="400"> </br> 以下三张图为gpt4o的识别效果</br> <img src="./images/gpt4o-01.JPG" width="600" height="400"> <img src="./images/gpt4o-02.JPG" width="600" height="400"> <img src="./images/gpt4o-03.JPG" width="600" height="400">

1. 1.代码</br> 推理代码

from transformers import LlavaForConditionalGeneration, AutoProcessor
import torch
from PIL import Image

raw_model_name_or_path = "/保存的完整模型路径"
model = LlavaForConditionalGeneration.from_pretrained(raw_model_name_or_path, device_map="cuda:0", torch_dtype=torch.bfloat16)
processor = AutoProcessor.from_pretrained(raw_model_name_or_path)
model.eval()

def build_model_input(model, processor):
    messages = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "<image>\n 你是一位有深度的网络图片解读者，擅长解读和描述网络图片。你能洞察图片中的细微之处，对图中的人物面部表情、文字信息、情绪流露和背景寓意具有超强的理解力，描述信息需要详细。"}
    ]
    prompt = processor.tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    image = Image.open("01.PNG")
    inputs = processor(text=prompt, images=image, return_tensors="pt", return_token_type_ids=False)
    
    for tk in inputs.keys():
        inputs[tk] = inputs[tk].to(model.device)
    generate_ids = model.generate(**inputs, max_new_tokens=200)
    
    generate_ids = [
        oid[len(iids):] for oid, iids in zip(generate_ids, inputs.input_ids)
    ]
    gen_text = processor.batch_decode(generate_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False)[0]
    return gen_text
build_model_input(model, processor)