原文：
Shikra: Unleashing Multimodal LLM's Referential Dialogue Magic
代码：
https://github.com/shikras/shikra
模型：
https://huggingface.co/shikras/shikra-7b-delta-v1
https://huggingface.co/shikras/shikra7b-delta-v1-0708
第一个是论文用的，第二个会有迭代。
本人的shikra论文解读，逐行解读，非常详细！
部署：
-
下载GitHub工程，和shikras的模型参数，注意，还要下载LLaMA-7b的模型；
-
创建环境：
conda create -n shikra python=3.10
conda activate shikra
pip install -r requirements.txt
后面我运行的时候报缺包了，又pip install了以下包，不过每个人情况不同：
pip install uvicorn
pip install mmengine然后还会报错：
File "/usr/local/lib/python3.10/dist-packages/cv2/typing/__init__.py", line 171, in <module>
LayerId = cv2.dnn.DictValue
AttributeError: module 'cv2.dnn' has no attribute 'DictValue'解决方案:
修改/usr/local/lib/python3.10/dist-packages/cv2/typing/init .py
注释掉LayerId = cv2.dnn.DictValue这行即可。
-
权重下载和合并
shikra官方提供的模型权重需要和llama1-7b合并之后才能用，然而llama1需要申请，比较麻烦，在hf上找到了平替（这一步我走了好久QwQ）：
https://huggingface.co/huggyllama/llama-7b
大家自己下载，然后运行官方提供的合并代码：python mllm/models/shikra/apply_delta.py
--base /path/to/llama-7b
--target /output/path/to/shikra-7b-merge
--delta shikras/shikra-7b-delta-v1
得到了可用的模型参数shikra-7b-merge。
注意要把参数文件夹里config里的模型路径改成merge版的。
此外还需要下载clip 模型参数：
https://huggingface.co/openai/clip-vit-large-patch14
代码和配置文件中有多处调用/openai/clip-vit-large-patch14，要改成本地版本。如果不预先下载，应该会在运行时自动下载，大家看网络情况自行选择。
- 我写的demo文件，用于在命令行测试模型效果，主要是为了不用gradio 和fastapi这些东西。
python
import argparse
import os
import sys
import base64
import logging
import time
from pathlib import Path
from io import BytesIO
import torch
import uvicorn
import transformers
from PIL import Image
from mmengine import Config
from transformers import BitsAndBytesConfig
sys.path.append(str(Path(__file__).parent.parent.parent))
from mllm.dataset.process_function import PlainBoxFormatter
from mllm.dataset.builder import prepare_interactive
from mllm.models.builder.build_shikra import load_pretrained_shikra
from mllm.dataset.utils.transform import expand2square, box_xyxy_expand2square
# Set up logging
log_level = logging.DEBUG
transformers.logging.set_verbosity(log_level)
transformers.logging.enable_default_handler()
transformers.logging.enable_explicit_format()
# prompt for coco
# Argument parsing
parser = argparse.ArgumentParser("Shikra Local Demo")
parser.add_argument('--model_path', default = "xxx/shikra-merge", help="Path to the model")
parser.add_argument('--load_in_8bit', action='store_true', help="Load model in 8-bit precision")
parser.add_argument('--image_path', default = "xxx/shikra-main/mllm/demo/assets/ball.jpg", help="Path to the image file")
parser.add_argument('--text', default="What do you see in this image? Please mention the objects and their locations using the format [x1,y1,x2,y2].", help="Text prompt")
parser.add_argument('--boxes_value', nargs='+', type=int, default=[], help="Bounding box values (x1, y1, x2, y2)")
parser.add_argument('--boxes_seq', nargs='+', type=int, default=[], help="Sequence of bounding boxes")
parser.add_argument('--do_sample', action='store_true', help="Use sampling during generation")
parser.add_argument('--max_length', type=int, default=512, help="Maximum length of the output")
parser.add_argument('--top_p', type=float, default=1.0, help="Top-p value for sampling")
parser.add_argument('--temperature', type=float, default=1.0, help="Temperature for sampling")
args = parser.parse_args()
model_name_or_path = args.model_path
# Model initialization
model_args = Config(dict(
type='shikra',
version='v1',
# checkpoint config
cache_dir=None,
model_name_or_path=model_name_or_path,
vision_tower=r'xxx/clip-vit-large-patch14',
pretrain_mm_mlp_adapter=None,
# model config
mm_vision_select_layer=-2,
model_max_length=2048,
# finetune config
freeze_backbone=False,
tune_mm_mlp_adapter=False,
freeze_mm_mlp_adapter=False,
# data process config
is_multimodal=True,
sep_image_conv_front=False,
image_token_len=256,
mm_use_im_start_end=True,
target_processor=dict(
boxes=dict(type='PlainBoxFormatter'),
),
process_func_args=dict(
conv=dict(type='ShikraConvProcess'),
target=dict(type='BoxFormatProcess'),
text=dict(type='ShikraTextProcess'),
image=dict(type='ShikraImageProcessor'),
),
conv_args=dict(
conv_template='vicuna_v1.1',
transforms=dict(type='Expand2square'),
tokenize_kwargs=dict(truncation_size=None),
),
gen_kwargs_set_pad_token_id=True,
gen_kwargs_set_bos_token_id=True,
gen_kwargs_set_eos_token_id=True,
))
training_args = Config(dict(
bf16=False,
fp16=True,
device='cuda',
fsdp=None,
))
quantization_kwargs = dict(
quantization_config=BitsAndBytesConfig(
load_in_8bit=args.load_in_8bit,
)
) if args.load_in_8bit else dict()
model, preprocessor = load_pretrained_shikra(model_args, training_args, **quantization_kwargs)
# Convert the model and vision tower to float16
if not getattr(model, 'is_quantized', False):
model.to(dtype=torch.float16, device=torch.device('cuda'))
if not getattr(model.model.vision_tower[0], 'is_quantized', False):
model.model.vision_tower[0].to(dtype=torch.float16, device=torch.device('cuda'))
preprocessor['target'] = {'boxes': PlainBoxFormatter()}
tokenizer = preprocessor['text']
# Load and preprocess the image
pil_image = Image.open(args.image_path).convert("RGB")
ds = prepare_interactive(model_args, preprocessor)
image = expand2square(pil_image)
boxes_value = [box_xyxy_expand2square(box, w=pil_image.width, h=pil_image.height) for box in zip(args.boxes_value[::2], args.boxes_value[1::2], args.boxes_value[2::2], args.boxes_value[3::2])]
ds.set_image(image)
ds.append_message(role=ds.roles[0], message=args.text, boxes=boxes_value, boxes_seq=args.boxes_seq)
model_inputs = ds.to_model_input()
model_inputs['images'] = model_inputs['images'].to(torch.float16)
# Generate
gen_kwargs = dict(
use_cache=True,
do_sample=args.do_sample,
pad_token_id=tokenizer.pad_token_id,
bos_token_id=tokenizer.bos_token_id,
eos_token_id=tokenizer.eos_token_id,
max_new_tokens=args.max_length,
top_p=args.top_p,
temperature=args.temperature,
)
input_ids = model_inputs['input_ids']
st_time = time.time()
with torch.inference_mode():
with torch.autocast(device_type='cuda', dtype=torch.float16):
output_ids = model.generate(**model_inputs, **gen_kwargs)
print(f"Generated in {time.time() - st_time} seconds")
input_token_len = input_ids.shape[-1]
response = tokenizer.batch_decode(output_ids[:, input_token_len:])[0]
print(f"Response: {response}")这么良心，点个关注吧，会持续更新多模态大模型相关内容。