README.md

metadata

license: gpl-3.0
language:
  - zh
  - en
pipeline_tag: visual-question-answering
tags:
  - ziya
  - fengshenbang
  - LVLM
  - visual question answering

Ziya-Visual-14B-Chat

• Main Page:Fengshenbang
• Github: Fengshenbang-LM

姜子牙系列模型

• Ziya-LLaMA-13B-v1.1
• Ziya-LLaMA-13B-v1
• Ziya-LLaMA-7B-Reward
• Ziya-LLaMA-13B-Pretrain-v1

软件依赖

pip install torch==1.12.1 tokenizers==0.13.3 git+https://github.com/huggingface/transformers

模型分类 Model Taxonomy

需求 Demand	任务 Task	系列 Series	模型 Model	参数 Parameter	额外 Extra
多模态 Multi-Modal	通用 General	姜子牙-多模态 Ziya-Visual	InstructBLIP LLaMA	14B	English&Chinese

使用 Usage

import gradio as gr
from PIL import Image
import torch
import random
from fengshen.models.instruct_ditto.modeling_instruct_ditto import InstructDittoLMForConditionalGeneration, DittoQFromerForPretrain, DittoLMForConditionalGeneration
from torchvision.transforms import Compose, ToTensor, Resize, Normalize
from transformers import LlamaTokenizer, BertTokenizer, GenerationConfig
from torchvision.transforms import Normalize, Compose, RandomResizedCrop, InterpolationMode, ToTensor, RandomHorizontalFlip

OPENAI_DATASET_MEAN = (0.48145466, 0.4578275, 0.40821073)
OPENAI_DATASET_STD = (0.26862954, 0.26130258, 0.27577711)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
_MODEL_PATH = "your model path"

transforms = Compose([
    RandomResizedCrop(
        224,
        scale=(0.5, 1.0),
        interpolation=InterpolationMode.BICUBIC,
    ),
    RandomHorizontalFlip(),
    ToTensor(),
    Normalize(mean=OPENAI_DATASET_MEAN, std=OPENAI_DATASET_STD),
])

model = InstructDittoLMForConditionalGeneration.from_pretrained(_MODEL_PATH).to(device).eval()
instruct_tokenizer = BertTokenizer.from_pretrained(os.path.join(_MODEL_PATH, "qformer_tokenizer"))
tokenizer = LlamaTokenizer.from_pretrained(_MODEL_PATH, use_fast = False)

qformer_prompt = "{prompt}"
qformer_prompt_list = []
prompt_prefix = ''
llm_prompt = "<human>: {prompt}\n<bot>:"
llm_prompt_list = []

prompt = ["your prompt"]

for i in prompt:
    qformer_prompt_list.append(qformer_prompt.format_map({"prompt":i}))
    llm_prompt_list.append(llm_prompt.format_map({"prompt":i}))

image_url = ["your image"]

imgs = []
for img_url in image_url:
    imgs.append(transforms(Image.open(img_url).convert('RGB')))

config = GenerationConfig(
    # do_sample=True, #False
    # num_beams=3, # 3
    # min_length=4,
    max_new_tokens=128, 
    repetition_penalty=1.18, 
    # length_penalty=1, 
    temperature=0.7,
    top_p=0.1,
    bos_token_id=1,
    eos_token_id=2, 
    pad_token_id=39410, 
)

imgs = torch.stack(imgs)

instruct_tokenizer.padding_side = 'right'
tokenizer.padding_side = 'left'

for i in range(imgs.shape[0]):
    prompt_prefix_ids = tokenizer(prompt_prefix, return_tensors="pt").input_ids
    qformer_instruct_ids = instruct_tokenizer(qformer_prompt_list[i], return_tensors="pt").input_ids
    llm_instruct_ids = tokenizer(llm_prompt_list[i], return_tensors="pt", add_special_tokens=False).input_ids
    qformer_instruct_atts = instruct_tokenizer(qformer_prompt_list[i], return_tensors="pt").attention_mask
    llm_instruct_atts = tokenizer(llm_prompt_list[i], return_tensors="pt", add_special_tokens=False).attention_mask
    captions = model.generate(
                            imgs[i].unsqueeze(0).to('cuda'),
                            qformer_instruct_ids=qformer_instruct_ids.to('cuda'),
                            prompt_prefix_ids = prompt_prefix_ids.to('cuda'),
                            llm_instruct_ids=llm_instruct_ids.to('cuda'),
                            generation_config=config
                            )
    caption = tokenizer.decode(captions[0])
    print("问: " + prompt[i] + "\n" + "答: " + caption)

引用 Citation

如果您在您的工作中使用了我们的模型，可以引用我们的论文，论文：

If you are using the resource for your work, please cite the our paper, paper:

@article{fengshenbang,
  author    = {Jiaxing Zhang and Ruyi Gan and Junjie Wang and Yuxiang Zhang and Lin Zhang and Ping Yang and Xinyu Gao and Ziwei Wu and Xiaoqun Dong and Junqing He and Jianheng Zhuo and Qi Yang and Yongfeng Huang and Xiayu Li and Yanghan Wu and Junyu Lu and Xinyu Zhu and Weifeng Chen and Ting Han and Kunhao Pan and Rui Wang and Hao Wang and Xiaojun Wu and Zhongshen Zeng and Chongpei Chen},
  title     = {Fengshenbang 1.0: Being the Foundation of Chinese Cognitive Intelligence},
  journal   = {CoRR},
  volume    = {abs/2209.02970},
  year      = {2022}
}

@article{lu2023ziya,
  title={Ziya-VL: Bilingual Large Vision-Language Model via Multi-Task Instruction Tuning},
  author={Lu, Junyu and Zhang, Dixiang and Wu, Xiaojun and Gao, Xinyu and Gan, Ruyi and Zhang, Jiaxing and Song, Yan and Zhang, Pingjian},
  journal={arXiv preprint arXiv:2310.08166},
  year={2023}
}

You can also cite our website:

欢迎引用我们的网站:

@misc{Fengshenbang-LM,
  title={Fengshenbang-LM},
  author={IDEA-CCNL},
  year={2021},
  howpublished={\url{https://github.com/IDEA-CCNL/Fengshenbang-LM}},
}