clip-retrieval

Easily compute clip embeddings and build a clip retrieval system with them. 100M text+image embeddings can be processed in 20h using a 3080.

• clip client allows remote querying of backend via python. clip-client notebook
• clip inference allows you to quickly (1500 sample/s on a 3080) compute image and text embeddings
• clip index builds efficient indices out of the embeddings
• clip filter allows you to filter out the data using the clip index
• clip back hosts the indices with a simple flask service
• clip front is a simple ui querying the back. Check it out at clip-retrieval ui
• clip end2end runs img2dataset, inference, index then back and front to make all of this easier to begin with

End to end this make it possible to build a simple semantic search system. Interested to learn about semantic search in general ? You can read my medium post on the topic.

Also see laion5B and semantic search at billions scale to read more on how to make this scale to billion of samples.

If you believe in making reusable tools to make data easy to use for ML and you would like to contribute, please join the DataToML chat.

Related projects

• all_clip to load any clip model
• img2dataset to download images from urls
• open_clip to train clip models
• CLIP_benchmark to evaluate clip models

Who is using clip retrieval ?

• cah-prepro preprocess the 400M image+text crawling at home dataset. clip-retrieval is used to compute 400M clip embeddings and the indices
• autofaiss uses clip-retrieval to display an example of use (see the multimodal notebook example there)
• afiaka87 openai demo shows how to look among the 1M example released by openai for their DALL-E demo
• antarctic-captions by dzryk uses autofaiss and clip inference as a way to generate anchors for the image to text task with great success

Install

pip install clip-retrieval

If your interest it to run the laion5B index, see this doc

Clip client

ClipClient allows remote querying of a clip-retrieval backend via python.

See ClipClient - Getting Started Notebook for a jupyter notebook example.

API Initialization

During initialization you can specify a few parameters:

• backend_url: the url of the backend. (required)
• indice_name: specify the name of the index you want to use. (required)
• aesthetic_score: the aesthetic score as rated by aesthetic-predictor. Default is 9.
• use_mclip: whether to use a multi-lingual version of CLIP. Default is False.
• aesthetic_weight: the weight of the aesthetic score. Default is 0.5
• modality: search over image or text in the index, one of Multimodal.IMAGE or Multimodal.TEXT. Default is Multimodal.IMAGE.
• num_images: the number of images to return from the API. Default is 40.
• deduplicate: Whether to deduplicate the result by image embedding. Default is true.
• use_safety_model: Whether to remove unsafe images. Default is true.
• use_violence_detector: Whether to remove images with violence. Default is true.

For instance, to query the hosted backend for Laion5B with the default parameters:

from clip_retrieval.clip_client import ClipClient, Modality

client = ClipClient(url="https://knn.laion.ai/knn-service", indice_name="laion5B-L-14")

Query by text

You can find captioned images similar to the text you provide.

results = client.query(text="an image of a cat")
results[0]
> {'url': 'https://example.com/kitten.jpg', 'caption': 'an image of a kitten', 'id': 14, 'similarity': 0.2367108941078186}

Query by image

You can also find captioned images similar to the image you provide. Images can be passed via local path or url.

cat_results = client.query(image="cat.jpg")
dog_results = client.query(image="https://example.com/dog.jpg")

Query by embedding

You can also find captioned images similar to a clip embedding you provide.

cat_results = client.query(embedding_input=cat_embedding)

Query a directory of images

To enhance an existing dataset with similar text/image pairs, you can query a directory of images and combine the results.

all_results = [result for result in [client.query(image=image) for image in os.listdir("my-images")]]
with open("search-results.json", "w") as f:
    json.dump(all_results, f)

Create a dataset

You can create a dataset using the saved json results and the tool img2dataset.

img2dataset "search-results.json" \
    --input_format="json" \
    --output_folder="knn_search_dataset" \
    --caption_col="caption"

clip end2end

First pick a dataset of image urls and captions (examples) then run:

You may want to run export CUDA_VISIBLE_DEVICES= to avoid using your GPU if it doesn't have enough VRAM.

wget https://github.com/rom1504/img2dataset/raw/main/tests/test_files/test_1000.parquet
clip-retrieval end2end test_1000.parquet /tmp/my_output

Then go to http://localhost:1234 and enjoy searching among your pictures

Use --run_back False if you don't want to run the backend

clip inference

Get some images in an example_folder, for example by doing:

pip install img2dataset
echo 'https://placekitten.com/200/305' >> myimglist.txt
echo 'https://placekitten.com/200/304' >> myimglist.txt
echo 'https://placekitten.com/200/303' >> myimglist.txt
img2dataset --url_list=myimglist.txt --output_folder=image_folder --thread_count=64 --image_size=256

You can also put text files with the same names as the images in that folder, to get the text embeddings.

Then run clip-retrieval inference --input_dataset image_folder --output_folder embeddings_folder

Output folder will contain:

• img_emb/
  • img_emb_0.npy containing the image embeddings as numpy
• text_emb/
  • text_emb_0.npy containing the text embeddings as numpy
• metadata/
  • metadata_0.parquet containing the image paths, captions and metadata

This scales to million of samples. At 1400 sample/s of a 3080, 10M samples can be processed in 2h.

API

clip_inference turn a set of text+image into clip embeddings

• input_dataset Path to input dataset. Folder if input_format is files. Bash brace pattern such as "{000..150}.tar" (see https://pypi.org/project/braceexpand/) if webdataset (required)
• output_folder Folder where the clip embeddings will be saved, as well as metadata (required)
• input_format files or webdataset (default files)
• cache_path cache path for webdataset (default None)
• batch_size Number of items to do the inference on at once (default 256)
• num_prepro_workers Number of processes to do the preprocessing (default 8)
• enable_text Enable text processing (default True)
• enable_image Enable image processing (default True)
• enable_metadata Enable metadata processing (default False)
• write_batch_size Write batch size (default 10**6)
• wds_image_key Key to use for images in webdataset. (default jpg)
• wds_caption_key Key to use for captions in webdataset. (default txt)
• clip_model CLIP model to load (default ViT-B/32). Specify it as "open_clip:ViT-B-32/laion2b_s34b_b79k" to use the open_clip or "hf_clip:patrickjohncyh/fashion-clip" to use the hugging face clip model.
• mclip_model MCLIP model to load (default sentence-transformers/clip-ViT-B-32-multilingual-v1)
• use_mclip If False it performs the inference using CLIP; MCLIP otherwise (default False)
• use_jit uses jit for the clip model (default True)
• distribution_strategy choose how to distribute the job, see distribution section for details (default sequential)
• wds_number_file_per_input_file estimation of the number of sample per tar if using wds and not specifying output_partition_count (default 10000)
• output_partition_count number of output partitions (default None)
• wandb_project wandb project to use (default clip_retrieval)
• enable_wandb whether to use wandb (default False)
• clip_cache_path cache path for clip (default None)
• slurm_job_name, the job name to use in slurm. (default None)
• slurm_partition (default None), the slurm partition to create a job in.
• slurm_jobs, the number of jobs to create in slurm. (default None)
• slurm_job_comment, the job comment to use. (default None)
• slurm_nodelist, a list of specific nodes to use .(default None)
• slurm_exclude, a list of nodes to exclude when creating jobs. (default None)
• slurm_job_timeout, if not supplied it will default to 2 weeks. (default None)
• slurm_cache_path, cache path to use for slurm-related tasks. (default None)
• slurm_verbose_wait=False, wether to print the status of your slurm job (default False)

DeepSparse Backend

DeepSparse is an inference runtime for fast sparse model inference on CPUs. There is a backend available within clip-retrieval by installing it with pip install deepsparse-nightly[clip], and specifying a clip_model with a prepended "nm:", such as "nm:neuralmagic/CLIP-ViT-B-32-256x256-DataComp-s34B-b86K-quant-ds" or "nm:mgoin/CLIP-ViT-B-32-laion2b_s34b_b79k-ds".

Inference Worker

If you wish to have more control over how inference is run, you can create and call workers directly using clip-retrieval inference.worker

Example Usage:

clip-retrieval inference.worker \
--tasks="[0]" \
--input_dataset="input/folder/{000000..000100}.tar" \
--output_folder="example/path" \
--input_format="webdataset" \
--output_partition_count="1"

Doing so will invoke a single worker that can be instructed to focus on a specific subset of the input_dataset. That worker will sequentially process the tasks passed to it. Here, tasks is a lists of partition_id's that this worker will be responsible for.

To manually compute the number of tasks, use the following formula: number_samples / wds_number_file_per_input_file.

The API is very similar to clip-retrieval inference with some minor changes:

• tasks A list of integers representing the partition_id's that this worker is responsible for computing. (required)
• input_dataset Path to input dataset. Folder if input_format is files. Bash brace pattern such as "{000..150}.tar" (see https://pypi.org/project/braceexpand/) if webdataset (required)
• output_folder Folder where the clip embeddings will be saved, as well as metadata (required)
• output_partition_count number of output partitions (required)
• input_format files or webdataset (default files)
• cache_path cache path for webdataset (default None)
• batch_size Number of items to do the inference on at once (default 256)
• num_prepro_workers Number of processes to do the preprocessing (default 8)
• enable_text Enable text processing (default True)
• enable_image Enable image processing (default True)
• enable_metadata Enable metadata processing (default False)
• wds_image_key Key to use for images in webdataset. (default jpg)
• wds_caption_key Key to use for captions in webdataset. (default txt)
• clip_model CLIP model to load (default ViT-B/32). Specify it as "open_clip:ViT-B-32-quickgelu" to use the open_clip or "hf_clip:patrickjohncyh/fashion-clip" to use the hugging face clip model.
• mclip_model MCLIP model to load (default sentence-transformers/clip-ViT-B-32-multilingual-v1)
• use_mclip If False it performs the inference using CLIP; MCLIP otherwise (default False)
• use_jit uses jit for the clip model (default True)
• wandb_project wandb project to use (default clip_retrieval)
• enable_wandb whether to use wandb (default False)
• clip_cache_path cache path for clip (default None)

> Note: The worker does not accept the following arguments > * write_batch_size Write batch size (default 10**6) > * distribution_strategy choose how to distribute the job, see distribution section for details (default sequential) > * wds_number_file_per_input_file estimation of the number of sample per tar if using wds and not specifying output_partition_count (default 10000) > * any of the SLURM arguments

Loading/writing files on hdfs

• To load a webdataset from a hdfs folder, set --input_dataset "pipe:hdfs dfs -cat path_on_hdfs" in the request without the "hdfs://" prefix.
• To write the output on hdfs, set --output_hdfs_folder to the path on hdfs prefixed by "hdfs://"

Example of hdfs query using webdataset format: `clip_inference --input_dataset "pipe:hdfs dfs -cat /myfolder/webdataset/{00000..00010}.tar" --output_folder "hdfs://myfolder/embeddings" --input_format webdataset

Loading/writing files on s3

`clip_inference --input_dataset "pipe:aws s3 cp --quiet s3://myfolder/webdataset/{00000..00010}.tar -" --output_folder "s3://myfolder/embeddings" --input_format webdataset

Distributed inference

To run this on multiple nodes (and multiple gpus), see tutorial at docs/distributed_clip_inference.md

Clip index

Clip index takes as input the output of clip inference and makes an index out of it using autofaiss

clip-retrieval index --embeddings_folder embeddings_folder --index_folder index_folder

• --max_index_memory_usage "16G" option allow configuring the amount of ram the index will consume. More ram, better knn recall (Default 4G).
• --current_memory_available 24G allows controlling how much ram is used during the creation process (Default 16G).
• --image_subfolder "img_emb" allows to specify a subfolder for the image embeddings which is concatenated to the --embeddings_folder option (Default img_emb).
• --text_subfolder "text_emb" allows to specify a subfolder for the text embeddings which is concatenated to the --embeddings_folder option (Default text_emb).
• --copy_metadata True makes it possible to choose whether to copy metadata or not at the end of the process (Default True).
• --nb_cores 8 allows controlling the number of threads (Default None, which will use all cores).

The output is a folder containing:

• image.index containing a faiss index for images
• text.index containing a faiss index for texts
• metadata folder containing the parquet metadata

Thanks to autofaiss and faiss, this scales to hundred of million of samples in a few hours.

You may want to carefully pick how much memory to use for your index in order to maximize the knn recall. autofaiss index selection colab can help along with autofaiss score_index command to check the recall of your index. In general indices using more memory get a better recall and hence are closer to a naive (slow) knn

Clip filter

Once the embeddings are computed, you may want to filter out the data by a specific query. For that you can run clip-retrieval filter --query "cat" --output_folder "cat/" --indice_folder "indice_folder" It will copy the 100 best images for this query in the output folder. Using the --num_results or --threshold may be helpful to refine the filter

Thanks to fast knn index, this can run in real time (<10ms) for large K values (100000), and in minutes for very large K values.

This scripts works for small datasets. For larger ones, please check [notebook/simple_filter.ipynb].

Clip back

Clip back is a simple knn service backend. If using both hdf5 and faiss memory mapping, it uses only the memory used by clip which is 4GB.

Run (output_folder is the output of clip index)

echo '{"example_index": "output_folder"}' > indices_paths.json
clip-retrieval back --port 1234 --indices-paths indices_paths.json

Options:

• --use_jit True uses jit for the clip model
• --clip_model "ViT-B/32" allows choosing the clip model to use. Prefix with "open_clip:" to use an open_clip model.
• --enable_mclip_option True loads the mclip model, making it possible to search in any language.
• --columns_to_return='["url", "image_path", "caption", "NSFW"] allows you to specify which columns should be fetched from the metadata and returned by the backend. It's useful to specify less in case of hdf5 caching to speed up the queries.
• --enable_faiss_memory_mapping=True option can be passed to use an index with memory mapping. That decreases the memory usage to zero.
• --enable_hdf5 True option can be passed to enable hdf5 caching for the metadata. HDF5 caching makes it possible to use the metadata with almost no memory usage.
• --use_arrow True allows using arrow instead of hdf5. Should be used along with clip_back_prepro for very large datasets (billions)
• --reorder_metadata_by_ivf_index True option takes advantage of the data locality property of results of a knn ivf indices: it orders the metadata collection in order of the IVF clusters. That makes it possible to have much faster metadata retrieval as the reads are then accessing a few mostly sequential parts of the metadata instead of many non sequential parts. In practice that means being able to retrieve 1M items in 1s whereas only 1000 items can be retrieved in 1s without this method. This will order the metadata using the first image index.
• --provide_safety_model True will automatically download and load a safety model. You need to pip install autokeras optional dependency for this to work.
• --provide_violence_detector True will load a violence detector, paper
• --provide_aesthetic_embeddings True will load the aesthetic embeddings and allow users to make the query move towards a nicer point of the clip space

These options can also be provided in the config file to have different options for each index. Example:

{
        "laion5B": {
                "indice_folder": "/mnt/laion5B/prepared_data",
                "provide_safety_model": true,
                "enable_faiss_memory_mapping": true,
                "use_arrow": true,
                "enable_hdf5": false,
                "reorder_metadata_by_ivf_index": false,
                "columns_to_return": ["url", "caption"],
                "clip_model": "ViT-L/14",
                "enable_mclip_option": false
        },
        "laion_400m": {
                "indice_folder": "/mnt/laion400M/index100",
                "provide_safety_model": true,
                "enable_faiss_memory_mapping": true,
                "enable_hdf5": true,
                "use_arrow": false,
                "reorder_metadata_by_ivf_index": true,
                "enable_mclip_option": true,
                "clip_model": "ViT-B/32"
        }
}

hdf5 or arrow caching is a good idea to use if:

• you do not have enough ram to load the metadata in memory
• your disk is fast (ie you have a ssd)

At this point you have a simple flask server running on port 1234 and that can answer these queries:

• /indices-list -> return a list of indices
• /knn-service that takes as input:

{
    "text": "a text query",
    "image": "a base64 image",
    "image_url": "http://some-url.com/a.jpg",
    "modality": "image", // image or text index to use
    "num_images": 4, // number of output images
    "indice_name": "example_index",
    "num_result_ids": 4 // optional, if specified fetch this number of results in total but only num_images with metadata
}

text, image and image_url are mutually exclusive and returns:

[
    {
        "image": "base 64 of an image",
        "text": "some result text",
        "id": 543
    },
    {
        "image": "base 64 of an image",
        "text": "some result text",
        "id": 782
    }
]

Each object may also contain an url field if the metadata provides it.

The id is the position of the item in the index. It may be used to query metadata with the /metadata endpoint:

{
    "indice_name": "example_index",
    "ids": [543, 782]
}

which returns:

{
    "image": "base 64 of an image",
    "text": "some result text"
    // any other key available in the metadata and specified in columns_to_return cli option
}

num_result_ids argument of /knn-service and /metadata can be used together to do large knn queries and then fetch the metadata only when needed. It makes sense to do that as knn search can be very efficient thanks to strong locality of reference of knn IVF index making it fast to do knn with a large K, whereas the current on-disk implementation of metadata (hdf5) does not have that property and hence cannot handle retrieving a large amount of random items fast. In particular this can be used to implement infinite scroll in a front end.

By default the backend will also expose a front end. That front end will by default hit this backend, however you may need to specify whether this is happening over http or https, in this case use the option --default_backend to specify the backend url. --url_column allows specifying the name of the column url for the front

Clip back: Benchmark and monitoring

This backends has a 50ms latency if using memory mapped indices and metadata. Throughput is about 20 query/s. For high throughput, using a grpc server is required as well as a GPU for fast clip inference, turning off memory mapping options can also speed up requests, at the cost of high ram usage.

This backends also exposes a prometheus /metrics endpoint as well as an human readable summary at /metrics-summary. This can (optionally) be used to setup a grafana dashboard for monitoring:

It can be seen on this dashboard that the slowest part of any call is fetching the image by its url in case of image url search, taking up to 300ms. For text queries or image queries, the latency is about 50ms. Here is an example of output in the metrics summary:

Among 20.0 calls to the knn end point with an average latency of 0.1889s per request, the step costs are (in order):
                        name                               description  calls  average proportion
0              download_time             Time spent downloading an url      6  0.3215s     170.2%
1          metadata_get_time            Time spent retrieving metadata     20  0.0415s      21.9%
2             knn_index_time       Time spent doing a knn on the index     20  0.0267s      14.1%
3  image_clip_inference_time   Time spent doing a image clip inference      6  0.0206s      10.9%
4   text_clip_inference_time    Time spent doing a text clip inference     14  0.0186s       9.8%
5          image_prepro_time  Time spent doing the image preprocessing      6  0.0097s       5.2%
6           text_prepro_time   Time spent doing the text preprocessing     14  0.0020s       1.0%

clip-front

Clip front is a simple UI that connects to clip back and display the results. You can use it at clip-retrieval ui

Or you can run it yourself with:

npm install -g clip-retrieval-front
clip-retrieval-front 3005

You can also run it with clip-retrieval front or back from the python package.

Development

For development it, go to front and run npm install then npm start.

For development

Either locally, or in gitpod (do export PIP_USER=false there)

Setup a virtualenv:

python3 -m venv .env
source .env/bin/activate
pip install -e .

to run tests:

pip install -r requirements-test.txt

then

make lint
make test

You can use make black to reformat the code

python -m pytest -x -s -v tests -k "test_runner" to run a specific test

If you want to use the front through the python backend or frontend, run

cd front
npm install
npm run build
cd ..
pip install -e .

Citation

@misc{beaumont-2022-clip-retrieval,
  author = {Romain Beaumont},
  title = {Clip Retrieval: Easily compute clip embeddings and build a clip retrieval system with them},
  year = {2022},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/rom1504/clip-retrieval}}
}