本页尚未翻译

本页内容仍为英文原文，源自官方文档 integrations/ai/huggingface。中文翻译正在分批进行中。

HuggingFace Hub

Use LanceDB directly on Lance datasets hosted on the HuggingFace Hub for multimodal search and retrieval.

Applies to: Embedded ✅, Enterprise ✅

HuggingFace Hub is a popular platform for sharing machine learning datasets, models, and other resources.

LanceDB can directly scan Lance datasets hosted on the HuggingFace Hub with hf:// URIs. This is enabled under the hood by the lance-huggingface integration that allows users to stream Lance datasets directly from HuggingFace without needing to download them first.

For ML and AI engineers working in LanceDB, this capability is incredibly useful for quickly exploring multimodal datasets and reusing Lance datasets shared by others, without writing custom data loaders or preprocessing pipelines.

The snippets below use the lance-format/laion-1m dataset published in Lance format. The dataset includes a million image-caption pairs, and the Lance dataset can package image embeddings alongside the metadata. This makes it useful for demonstrating LanceDB's multimodal search capabilities in combination with easy sharing via the HuggingFace Hub.

The LAION table includes multimodal columns such as:

• image (inline JPEG bytes)
• caption (text)
• img_emb (image embedding vector)
• metadata fields such as url and similarity

Install dependencies

pip install lancedb pillow

Open the dataset with LanceDB

LanceDB can open the dataset directly from the Hub, without needing to download it first. Note that in LanceDB, you need to specify the table name when opening a Lance dataset, and the HuggingFace convention is to use train and test splits for datasets. The LAION dataset is uploaded as a single split named train, so we specify the table name that contains the *.lance files when opening the dataset.

db = lancedb.connect("hf://datasets/lance-format/laion-1m/data")
table = db.open_table("train")

print(f"Opened table: {table.name}")
print(f"Rows: {len(table)}")

Inspect schema and available indexes

print(table.schema)

This prints the schema of the LAION table. Note that there's an image embedding column that's a fixed-size list of 768-dimensional floats, and a binary column containing the raw JPEG bytes of the image.

image_path: string
caption: string
NSFW: string
similarity: double
LICENSE: string
url: string
key: string
status: string
error_message: null
width: int64
height: int64
original_width: int64
original_height: int64
exif: string
md5: string
img_emb: fixed_size_list<item: float>[768]
  child 0, item: float
image: binary

When inspecting Lance datasets from HuggingFace, it's also a good idea to check whether the dataset author included any pre-built indexes that you can use for search. You can check the available indexes with:

print(table.list_indices())

[
    Index(IvfPq, columns=["img_emb"], name="img_emb_idx"),
    Index(FTS, columns=["caption"], name="caption_idx")
]

In this case, we see that we have an IVF_PQ vector index on the img_emb column, and an FTS index on the caption column, which means we can directly do vector search on the image embeddings and keyword search on the captions without needing to build the indexes ourselves!

说明

If you see an empty list, it may be because the dataset author did not include the index files when uploading to HuggingFace. You can download the dataset locally, and build the indexes yourself. See the indexing guide for instructions on building different types of indexes with LanceDB.

Projection scan

Run a simple scan by projecting relevant columns to get a feel for the dataset. For example, we can run a search without any filters or input parameters to get a small subset of the data:

rows = (
    table.search()
    .select(["caption", "url", "similarity"])
    .limit(3)
    .to_list()
)

for i, row in enumerate(rows, start=1):
    print(f"{i}. {row['caption']}")
    print(f"   url={row['url']}")
    print(f"   similarity={row['similarity']}")

We get the first three rows and their metadata printed out, which look like this:

1. Cordelia and Dudley on their wedding  day last year
   url=https://i.dailymail.co.uk/i/pix/2012/01/05/article-2082728-0EF8956600000578-53_233x315.jpg
   similarity=0.2926466464996338
2. Statistics on challenges for automation in 2021
   url=https://verloop.io/wp-content/uploads/2021/02/Challenges.jpg
   similarity=0.30174341797828674
3. Teacher Gifts / Great gifts for your child's teacher.  Don't know what to get?  Take a look at these gifts that the teacher in your life will love!
   url=https://i.pinimg.com/custom_covers/216x146/550494823141083777_1487893945.jpg
   similarity=0.3362061381340027

Scan and filter data

Filtered search is a common pattern to narrow down interesting subsets of the data during early exploration. Here's an example:

filtered = (
    table.search()
    .where("height > 600")
    .select(["caption", "url", "width", "height"])
    .limit(3)
    .to_list()
)

for row in filtered:
    print(row["caption"], row["url"], row["width"], row["height"])

This prints out the metadata for large images with height greater than 600 pixels:

Luca Trousers, mustard stripe https://cdn.shopify.com/s/files/1/0151/5333/products/IMG_0791_1024x1024.jpg?v=1585142190 384 766
Baby Blue Fitted Short Sleeve T Shirt 3 https://cdn-img.prettylittlething.com/a/d/d/1/add198cab3ec30a61102437275573f4963642528_cmf6022_3.jpg 384 612
pattern cutting made easy pdf https://i.pinimg.com/736x/7c/6c/a7/7c6ca7361815a8929b3dd6ad34a03ab9.jpg 384 1045

Export image bytes to local files

To work with a subset of the data locally, you can export the image bytes from the table and save them as JPEG files.

from pathlib import Path

sample = (
    table.search()
    .select(["image", "caption"])
    .limit(3)
    .to_list()
)

out_dir = Path("samples")
out_dir.mkdir(exist_ok=True)

for i, row in enumerate(sample):
    out_path = out_dir / f"laion_{i}.jpg"
    with open(out_path, "wb") as f:
        f.write(row["image"])
    print(f"Saved {out_path} | caption={row['caption']}")

You can now preview the images you just exported on your local machine to get a better sense of the data.

Vector search

You can use LanceDB to run vector search directly on the data on the Hub, without needing to download the dataset or build your own vector index. This makes it incredibly easy to explore the dataset and iterate on your search queries before you decide to download a local copy for further experimentation on your end.

# Pick an arbitrary image embedding from the dataset
query_embedding = (
    table.search()
    .select(["img_emb"])
    .limit(1)
    .to_list()[0]["img_emb"]
)

results = (
    table.search(query_embedding, vector_column_name="img_emb")
    .select(["caption", "url", "_distance"])
    .limit(3)
    .to_list()
)

for row in results:
    print(row["_distance"], row["caption"])

distance	caption
0.17765313386917114	Cordelia and Dudley on their wedding day last year
0.17765313386917114	Cordelia and Dudley on their wedding day last year
0.17765313386917114	Cordelia and Dudley on their wedding day last year

注意

Note that the LAION dataset is known to contain a lot of duplicate images, so you may see the same image showing up multiple times in the search results.

Full-text search

Run an FTS search query that uses BM25 ranking on the caption column (on which we already have an FTS index):

fts_results = (
    table.search("dog running on beach", query_type="fts")
    .select(["caption", "url", "_score"])
    .limit(3)
    .to_list()
)

caption	url	_score
running with dog	https://www.doggytastic.com/wp…	15.73168
Dog Running in Water	https://static.wixstatic.com/m…	14.756516
Dogs on the run by heidiannemo…	http://ih2.redbubble.net/image…	14.756516

Download the full dataset

注意

You may hit HuggingFace rate limits when streaming large samples from hf://, despite using a HuggingFace token. For repeated queries or queries that operate on the full dataset, it's recommended to download the dataset locally and query from disk.

Here's how to download the entire dataset via the HuggingFace CLI:

huggingface-cli download lance-format/laion-1m --repo-type dataset --local-dir ./laion-1m

Upload your own datasets to HuggingFace in Lance format

This section shows how you can upload your own Lance datasets to the HuggingFace Hub to share with the community.

First, install the HuggingFace CLI and export both OPENAI_API_KEY and HF_TOKEN. Then, create a Lance dataset using LanceDB on a local machine, and then proceed to upload it to the Hub via a CLI command.

export OPENAI_API_KEY=...
export HF_TOKEN=hf_...
hf auth login --token "$HF_TOKEN"

A typical sequence of steps is given below.

1. Upload your local directory to the Hub

Upload the full local directory to a specified repository on the HuggingFace Hub. The command below uploads the contents of your local LanceDB directory at /path/to/your_local_dir to a new repository named your_hf_org/repo_name under your HuggingFace account.

hf upload-large-folder /path/to/your_local_dir your_hf_org/repo_name \
  --repo-type dataset \
  --revision main

说明

The upload-large-folder command is designed for uploading large datasets (potentially terabytes in size) and will handle multipart uploads, retries, and resuming interrupted uploads.

2. Inspect dataset versions

Because you can query your remote dataset directly from HuggingFace with hf:// URIs in LanceDB, you can easily inspect the dataset versions and updates on the Hub without needing to download the data locally. This is very useful to keep track of changes to the dataset and iterate on your data collection and curation process.

db = lancedb.connect("hf://datasets/your_hf_org/repo_name")
table = db.open_table("table_name")

versions = table.list_versions()
print(versions)

This will print out the list of versions available for the dataset on the Hub, along with their metadata such as creation date and description.

3. Add a dataset card

The Hub dataset card allows you to communicate the schema and usage of the dataset to other developers. It sits at the repo's root in a file named README.md on the Hub. This project keeps the source card text in HF_DATASET_CARD.md, so you can publish updates to the dataset there and upload it as README.md using the following command on the HF CLI: this requires a regular hf upload because it is a single-file upload to a specific target path (a custom commit message can be added if you wish).

hf upload lancedb/magical_kingdom HF_DATASET_CARD.md README.md \
  --repo-type dataset \
  --commit-message "Update dataset card"

4. Update the dataset

Over time, you may want to add new rows (append) or columns (backfill) to your dataset as your needs evolve. You can make the necessary updates to your local dataset using LanceDB, and then upload the updated version back to the Hub with the same hf upload-large-folder command.

hf upload-large-folder /path/to/your_local_dir your_hf_org/repo_name \
  --repo-type dataset \
  --revision main

The CLI will only upload the new data that has changed since the last upload, avoiding wasted I/O while making it easy to keep your dataset up-to-date on the Hub.

That's it! Your dataset is now updated on the Hub with the new data and schema changes, and other users can query the latest version of the dataset directly from HuggingFace with hf:// URIs in LanceDB.

Explore more Lance datasets on HuggingFace

The LanceDB team is actively uploading useful and interesting datasets in Lance format to the HuggingFace Hub under the lance-format organization. We actively encourage the HuggingFace and LanceDB communities to upload their own Lance datasets to the Hub to share with others!

In the meantime, feel free to check out the HuggingFace Hub to discover more Lance datasets uploaded by the community.

[card]

Click here to explore the latest trending Lance datasets on 🤗 HuggingFace!

Open