metadata

license: apache-2.0
language:
  - zh
  - en
pipeline_tag: image-text-to-text

从0到1训练一个定制版的llava模型

基于openai/clip-vit-large-patch14-336 和Qwen1.5-4B-Chat模型，构建一个llava模型
使用数据liuhaotian/LLaVA-CC3M-Pretrain-595K
训练方式是deepspeed-zero2、lora进行微调。

关联的github

https://github.com/yuanzhoulvpi2017/zero_nlp/tree/main/train_llava

关联的b站学习视频

待填充

推理代码


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


raw_model_name_or_path = "yuanzhoulvpi/llava_qwen15-4b-chat_openai-clip-vit-large-patch14-336"
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()
print('ok')

testdata = (
'<image>\nRelay a brief, clear account of the picture shown.', # 提问
 'large kitchen island with an overhang and dining space next to it', # 真实答案
 'data/liuhaotian/LLaVA-CC3M-Pretrain-595K/images_dl/GCC_train_001899387.jpg' # 图片路径
)

def build_model_input(model, processor, testdata:tuple):
    messages = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": testdata[0]},
    ]
    prompt = processor.tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    # print(prompt)
    # print("*"*20)
    image = Image.open(testdata[2])
    inputs = processor(text=prompt, images=image, return_tensors="pt")
    
    for tk in inputs.keys():
        inputs[tk] = inputs[tk].to(model.device)
    generate_ids = model.generate(**inputs, max_new_tokens=20)
    
    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, testdata)

# 'the kitchen is a bright yellow with a glass top island and a large window that looks out to the'