Skip to main content

Laion-400M dataset

The Laion-400M dataset contains 400 million images with English image captions. Laion nowadays provides an even larger dataset but working with it will be similar.

The dataset contains the image URL, embeddings for both the image and the image caption, a similarity score between the image and the image caption, as well as metadata, e.g. the image width/height, the licence and a NSFW flag. We can use the dataset to demonstrate approximate nearest neighbor search in ClickHouse.

Data preparation

The embeddings and the metadata are stored in separate files in the raw data. A data preparation step downloads the data, merges the files, converts them to CSV and imports them into ClickHouse. You can use the following download.sh script for that:

number=${1}
if [[ $number == '' ]]; then
number=1
fi;
wget --tries=100 https://deploy.laion.ai/8f83b608504d46bb81708ec86e912220/embeddings/img_emb/img_emb_${number}.npy # download image embedding
wget --tries=100 https://deploy.laion.ai/8f83b608504d46bb81708ec86e912220/embeddings/text_emb/text_emb_${number}.npy # download text embedding
wget --tries=100 https://deploy.laion.ai/8f83b608504d46bb81708ec86e912220/embeddings/metadata/metadata_${number}.parquet # download metadata
python3 process.py $number # merge files and convert to CSV

Script process.py is defined as follows:

import pandas as pd
import numpy as np
import os
import sys

str_i = str(sys.argv[1])
npy_file = "img_emb_" + str_i + '.npy'
metadata_file = "metadata_" + str_i + '.parquet'
text_npy = "text_emb_" + str_i + '.npy'

# load all files
im_emb = np.load(npy_file)
text_emb = np.load(text_npy)
data = pd.read_parquet(metadata_file)

# combine files
data = pd.concat([data, pd.DataFrame({"image_embedding" : [*im_emb]}), pd.DataFrame({"text_embedding" : [*text_emb]})], axis=1, copy=False)

# columns to be imported into ClickHouse
data = data[['url', 'caption', 'NSFW', 'similarity', "image_embedding", "text_embedding"]]

# transform np.arrays to lists
data['image_embedding'] = data['image_embedding'].apply(lambda x: list(x))
data['text_embedding'] = data['text_embedding'].apply(lambda x: list(x))

# this small hack is needed becase caption sometimes contains all kind of quotes
data['caption'] = data['caption'].apply(lambda x: x.replace("'", " ").replace('"', " "))

# export data as CSV file
data.to_csv(str_i + '.csv', header=False)

# removed raw data files
os.system(f"rm {npy_file} {metadata_file} {text_npy}")

To start the data preparation pipeline, run:

seq 0 409 | xargs -P1 -I{} bash -c './download.sh {}'

The dataset is split into 410 files, each file contains ca. 1 million rows. If you like to work with a smaller subset of the data, simply adjust the limits, e.g. seq 0 9 | ....

(The python script above is very slow (~2-10 minutes per file), takes a lot of memory (41 GB per file), and the resulting csv files are big (10 GB each), so be careful. If you have enough RAM, increase the -P1 number for more parallelism. If this is still too slow, consider coming up with a better ingestion procedure - maybe converting the .npy files to parquet, then doing all the other processing with clickhouse.)

Create table

To create a table without indexes, run:

CREATE TABLE laion
(
`id` Int64,
`url` String,
`caption` String,
`NSFW` String,
`similarity` Float32,
`image_embedding` Array(Float32),
`text_embedding` Array(Float32)
)
ENGINE = MergeTree
ORDER BY id
SETTINGS index_granularity = 8192

To import the CSV files into ClickHouse:

INSERT INTO laion FROM INFILE '{path_to_csv_files}/*.csv'

Run a brute-force ANN search (without ANN index)

To run a brute-force approximate nearest neighbor search, run:

SELECT url, caption FROM laion ORDER BY L2Distance(image_embedding, {target:Array(Float32)}) LIMIT 30

target is an array of 512 elements and a client parameter. A convenient way to obtain such arrays will be presented at the end of the article. For now, we can run the embedding of a random cat picture as target.

Result

┌─url───────────────────────────────────────────────────────────────────────────────────────────────────────────┬─caption────────────────────────────────────────────────────────────────┐
│ https://s3.amazonaws.com/filestore.rescuegroups.org/6685/pictures/animals/13884/13884995/63318230_463x463.jpg │ Adoptable Female Domestic Short Hair │
│ https://s3.amazonaws.com/pet-uploads.adoptapet.com/8/b/6/239905226.jpg │ Adopt A Pet :: Marzipan - New York, NY │
│ http://d1n3ar4lqtlydb.cloudfront.net/9/2/4/248407625.jpg │ Adopt A Pet :: Butterscotch - New Castle, DE │
│ https://s3.amazonaws.com/pet-uploads.adoptapet.com/e/e/c/245615237.jpg │ Adopt A Pet :: Tiggy - Chicago, IL │
│ http://pawsofcoronado.org/wp-content/uploads/2012/12/rsz_pumpkin.jpg │ Pumpkin an orange tabby kitten for adoption │
│ https://s3.amazonaws.com/pet-uploads.adoptapet.com/7/8/3/188700997.jpg │ Adopt A Pet :: Brian the Brad Pitt of cats - Frankfort, IL │
│ https://s3.amazonaws.com/pet-uploads.adoptapet.com/8/b/d/191533561.jpg │ Domestic Shorthair Cat for adoption in Mesa, Arizona - Charlie │
│ https://s3.amazonaws.com/pet-uploads.adoptapet.com/0/1/2/221698235.jpg │ Domestic Shorthair Cat for adoption in Marietta, Ohio - Daisy (Spayed) │
└───────────────────────────────────────────────────────────────────────────────────────────────────────────────┴────────────────────────────────────────────────────────────────────────┘

8 rows in set. Elapsed: 6.432 sec. Processed 19.65 million rows, 43.96 GB (3.06 million rows/s., 6.84 GB/s.)

Run a ANN with an ANN index

Create a new table with an ANN index and insert the data from the existing table:

CREATE TABLE laion_annoy
(
`id` Int64,
`url` String,
`caption` String,
`NSFW` String,
`similarity` Float32,
`image_embedding` Array(Float32),
`text_embedding` Array(Float32),
INDEX annoy_image image_embedding TYPE annoy(),
INDEX annoy_text text_embedding TYPE annoy()
)
ENGINE = MergeTree
ORDER BY id
SETTINGS index_granularity = 8192;

INSERT INTO laion_annoy SELECT * FROM laion;

By default, Annoy indexes use the L2 distance as metric. Further tuning knobs for index creation and search are described in the Annoy index documentation. Let's check now again with the same query:

SELECT url, caption FROM laion_annoy ORDER BY l2Distance(image_embedding, {target:Array(Float32)}) LIMIT 8

Result

┌─url──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┬─caption──────────────────────────────────────────────────────────────┐
│ http://tse1.mm.bing.net/th?id=OIP.R1CUoYp_4hbeFSHBaaB5-gHaFj │ bed bugs and pets can cats carry bed bugs pets adviser │
│ http://pet-uploads.adoptapet.com/1/9/c/1963194.jpg?336w │ Domestic Longhair Cat for adoption in Quincy, Massachusetts - Ashley │
│ https://thumbs.dreamstime.com/t/cat-bed-12591021.jpg │ Cat on bed Stock Image │
│ https://us.123rf.com/450wm/penta/penta1105/penta110500004/9658511-portrait-of-british-short-hair-kitten-lieing-at-sofa-on-sun.jpg │ Portrait of british short hair kitten lieing at sofa on sun. │
│ https://www.easypetmd.com/sites/default/files/Wirehaired%20Vizsla%20(2).jpg │ Vizsla (Wirehaired) image 3 │
│ https://images.ctfassets.net/yixw23k2v6vo/0000000200009b8800000000/7950f4e1c1db335ef91bb2bc34428de9/dog-cat-flickr-Impatience_1.jpg?w=600&h=400&fm=jpg&fit=thumb&q=65&fl=progressive │ dog and cat image │
│ https://i1.wallbox.ru/wallpapers/small/201523/eaa582ee76a31fd.jpg │ cats, kittens, faces, tonkinese │
│ https://www.baxterboo.com/images/breeds/medium/cairn-terrier.jpg │ Cairn Terrier Photo │
└──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┴──────────────────────────────────────────────────────────────────────┘

8 rows in set. Elapsed: 0.641 sec. Processed 22.06 thousand rows, 49.36 MB (91.53 thousand rows/s., 204.81 MB/s.)

The speed increased significantly at the cost of less accurate results. This is because the ANN index only provide approximate search results. Note the example searched for similar image embeddings, yet it is also possible to search for positive image caption embeddings.

Creating embeddings with UDFs

One usually wants to create embeddings for new images or new image captions and search for similar image / image caption pairs in the data. We can use UDF to create the target vector without leaving the client. It is important to use the same model to create the data and new embeddings for searches. The following scripts utilize the ViT-B/32 model which also underlies the dataset.

Text embeddings

First, store the following Python script in the user_scripts/ directory of your ClickHouse data path and make it executable (chmod +x encode_text.py).

encode_text.py:

#!/usr/bin/python3
import clip
import torch
import numpy as np
import sys

if __name__ == '__main__':
device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device)
for text in sys.stdin:
inputs = clip.tokenize(text)
with torch.no_grad():
text_features = model.encode_text(inputs)[0].tolist()
print(text_features)
sys.stdout.flush()

Then create encode_text_function.xml in a location referenced by <user_defined_executable_functions_config>/path/to/*_function.xml</user_defined_executable_functions_config> in your ClickHouse server configuration file.

<functions>
<function>
<type>executable</type>
<name>encode_text</name>
<return_type>Array(Float32)</return_type>
<argument>
<type>String</type>
<name>text</name>
</argument>
<format>TabSeparated</format>
<command>encode_text.py</command>
<command_read_timeout>1000000</command_read_timeout>
</function>
</functions>

You can now simply use:

SELECT encode_text('cat');

The first run will be slow because it loads the model, but repeated runs will be fast. We can then copy the output to SET param_target=... and can easily write queries.

Image embeddings

Image embeddings can be created similarly but we will provide the Python script the path to a local image instead of the image caption text.

encode_image.py

#!/usr/bin/python3
import clip
import torch
import numpy as np
from PIL import Image
import sys

if __name__ == '__main__':
device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device)
for text in sys.stdin:
image = preprocess(Image.open(text.strip())).unsqueeze(0).to(device)
with torch.no_grad():
image_features = model.encode_image(image)[0].tolist()
print(image_features)
sys.stdout.flush()

encode_image_function.xml

<functions>
<function>
<type>executable_pool</type>
<name>encode_image</name>
<return_type>Array(Float32)</return_type>
<argument>
<type>String</type>
<name>path</name>
</argument>
<format>TabSeparated</format>
<command>encode_image.py</command>
<command_read_timeout>1000000</command_read_timeout>
</function>
</functions>

Then run this query:

SELECT encode_image('/path/to/your/image');