# Laion-400M dataset The [Laion-400M dataset](https://laion.ai/blog/laion-400-open-dataset/) contains 400 million images with English image captions. Laion nowadays provides [an even larger dataset](https://laion.ai/blog/laion-5b/) 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](../../engines/table-engines/mergetree-family/annindexes.md) 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: ```bash wget --tries=100 https://deploy.laion.ai/8f83b608504d46bb81708ec86e912220/embeddings/img_emb/img_emb_${1}.npy # download image embedding wget --tries=100 https://deploy.laion.ai/8f83b608504d46bb81708ec86e912220/embeddings/text_emb/text_emb_${1}.npy # download text embedding wget --tries=100 https://deploy.laion.ai/8f83b608504d46bb81708ec86e912220/embeddings/metadata/metadata_${1}.parquet # download metadata python3 process.py ${1} # merge files and convert to CSV ``` Script `process.py` is defined as follows: ```python 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: ```bash 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: ```sql 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: ```sql 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: ```sql 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: ```sql 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](../../engines/table-engines/mergetree-family/annindexes.md). Let's check now again with the same query: ```sql 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](../../sql-reference/functions/index.md#sql-user-defined-functions) 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`: ```python #!/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 `/path/to/*_function.xml` in your ClickHouse server configuration file. ```xml executable encode_text Array(Float32) String text TabSeparated encode_text.py 1000000 ``` You can now simply use: ```sql 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` ```python #!/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` ```xml executable_pool encode_image Array(Float32) String path TabSeparated encode_image.py 1000000 ``` Then run this query: ```sql SELECT encode_image('/path/to/your/image'); ```