向量数据库:usearch的简单使用+实现图片检索应用

usearch的简单使用

• usearch是快速开源搜索和聚类引擎×，用于C++、C、Python、JavaScript、Rust、Java、Objective-C、Swift、C#、GoLang和Wolfram 🔍中的向量和🔜字符串×

// https://github.com/unum-cloud/usearch/blob/main/python/README.md
$ pip install usearch
Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
Collecting usearch
  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/ba/f4/24124f65ea3e940e54af29d55204ddfbeafa86d6b94b63c2e99baff2f7d6/usearch-2.8.14-cp38-cp38-manylinux_2_28_x86_64.whl (1.5 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.5/1.5 MB 17.0 MB/s eta 0:00:00
Requirement already satisfied: numpy in /home/ubuntu/anaconda3/envs/vglm2/lib/python3.8/site-packages (from usearch) (1.23.1)
Requirement already satisfied: tqdm in /home/ubuntu/anaconda3/envs/vglm2/lib/python3.8/site-packages (from usearch) (4.66.1)
Installing collected packages: usearch
Successfully installed usearch-2.8.14

• 一个简单的例子（注：本例子在运行时向index中不断添加项目，并将最后的index持久化为一个文件，在运行时由于添加项目内存占用会不断增加）

import numpy as np
from usearch.index import Index, MetricKind, Matches

ndim = 131072
index_path =  "test.usearch"

index = Index(
    ndim=ndim, # Define the number of dimensions in input vectors
    metric='cos', # Choose 'l2sq', 'haversine' or other metric, default = 'ip'
    dtype='f32', # Quantize to 'f16' or 'i8' if needed, default = 'f32'
    connectivity=16, # How frequent should the connections in the graph be, optional
    expansion_add=128, # Control the recall of indexing, optional
    expansion_search=64, # Control the quality of search, optional
)# index = Index(ndim=ndim, metric=MetricKind.Cos)



for i in range(1,10):
    vector =  np.random.random((1000, ndim)).astype('float32')
    index.add(None, vector, log=True)

index.save(index_path)
vector =  np.random.random((1, ndim)).astype('float32')
matches: Matches = index.search(vector, 10)
ids = matches.keys.flatten()

print(matches)

# test.usearch大小： 10*1000*131072 =>2.2G  (如果dtype='f32'=>4G+)

usearch-images

• https://github.com/ashvardanian/usearch-images

运行效果

数据获取

• https://huggingface.co/datasets/unum-cloud/ann-unsplash-25k/tree/main
  

依赖 ucall

• Requires: Python >=3.9
• https://pypi.org/project/ucall/#files
  

OSError: [Errno 28] inotify watch limit reached

  File "/home/ubuntu/anaconda3/envs/usearch/lib/python3.10/site-packages/watchdog/observers/inotify_c.py", line 428, in _raise_error
    raise OSError(errno.ENOSPC, "inotify watch limit reached")
OSError: [Errno 28] inotify watch limit reached

这个错误表明在使用 watchdog 库时超过了 Linux 系统对 inotify 监视的文件数或目录数的限制。Linux 对于每个进程的 inotify 能够监视的文件和目录有一个限制，当达到这个限制时，会出现像上面的错误一样的问题。可以尝试增加系统对 inotify 的资源限制。可以通过修改 /etc/sysctl.conf 文件来增加 fs.inotify.max_user_watches 参数的值。例如：

```bash
sudo sysctl -w fs.inotify.max_user_watches=65536
```

然后，运行下面的命令使更改生效：

```bash
sudo sysctl -p
```

注意：增加监视数可能会对系统资源产生一些影响，因此请根据实际情况慎重调整。

模型加载

• https://huggingface.co/unum-cloud/uform-vl-multilingual-v2/tree/main
  

///home/ubuntu/anaconda3/envs/usearch/lib/python3.10/site-packages/uform/__init__.py
def get_checkpoint(model_name, token) -> Tuple[str, Mapping, str]:
    model_path = snapshot_download(repo_id=model_name, token=token)
    config_path = f"{model_path}/torch_config.json"
    state = torch.load(f"{model_path}/torch_weight.pt")

    return config_path, state, f"{model_path}/tokenizer.json"


def get_model(model_name: str, token: Optional[str] = None) -> VLM:
    config_path, state, tokenizer_path = get_checkpoint(model_name, token)

    with open(config_path, "r") as f:
        model = VLM(load(f), tokenizer_path)

    model.image_encoder.load_state_dict(state["image_encoder"])
    model.text_encoder.load_state_dict(state["text_encoder"])

    return model.eval()

• 修改成如下，调用时使用_model = get_model("你的下载路径")

def get_checkpoint(model_name, token) -> Tuple[str, Mapping, str]:
    model_path = model_name#snapshot_download(repo_id=model_name, token=token)
    config_path = f"{model_path}/torch_config.json"
    state = torch.load(f"{model_path}/torch_weight.pt")

    return config_path, state, f"{model_path}/tokenizer.json"


def get_model(model_name: str, token: Optional[str] = None) -> VLM:
    config_path, state, tokenizer_path = get_checkpoint(model_name, token)

    with open(config_path, "r") as f:
        model = VLM(load(f), tokenizer_path)

    model.image_encoder.load_state_dict(state["image_encoder"])
    model.text_encoder.load_state_dict(state["text_encoder"])

    return model.eval()

其他细微的修改

数据源的修改

_datasets = {
    name: _open_dataset(os.path.join("/home/ubuntu/userfile/***/Usearch/usearch-images-main/data", name))
    for name in (
        "unsplash-25k",
        # "cc-3m",
        # "laion-4m",
    )
}

dataset_names: str = st.multiselect(
    "Datasets",
    [
        dataset_unsplash_name,
        # dataset_cc_name,
        # dataset_laion_name,
    ],
    [dataset_unsplash_name],#, dataset_cc_name],
    format_func=lambda x: x.split(":")[0],
)

• 也可下载cc-3m数据：
  

数据读取的修改

    # uris: Strs = File(os.path.join(dir, "images.txt")).splitlines()
    file_path = os.path.join(dir, "images.txt")
    with open(file_path, 'r') as file:
        uris = file.read().splitlines()

CG

• “usearch” 通常指的是一个生物信息学工具，用于对DNA和蛋白质序列进行搜索和比对。具体来说，它是由Qiime软件包提供的一个用于序列分析的工具，主要用于对微生物群落的高通量测序数据进行处理和分析。Qiime（Quantitative Insights Into Microbial Ecology）是一个用于分析和解释微生物群落结构的开源软件包。在Qiime中，usearch被用于处理和比对DNA序列，以便进行物种注释、多样性分析等。USEARCH —— 最简单易学的扩增子分析流程