Client API#

CLIP-as-service is designed in a client-server architecture. You can use clip_client to send images and texts to the server and receive the responses from the server. Right now, clip_client provides encoding, ranking, indexing, and searching functionalities. Additionally, it has many nice designs for speeding up the processing of a large amount of data:

  • Streaming: request sending is not blocked by the response receiving. Sending and receiving are two separate streams that run in parallel. Both are independent and each have separate internal buffer.

  • Batching: large requests are segmented into small batches and send in a stream.

  • Low memory footprint: only load data when needed.

  • Sync/async interface: provide async interface that can be easily integrated into other asynchronous system.

  • Auto-detect images and text input.

  • Support gRPC, HTTP, Websocket protocols with their TLS counterparts.

Tip

You will need to install clip_client first in Python 3.7+: pip install clip-client.

Construct client#

To use clip_client, you need to first construct a Client object, e.g.:

from clip_client import Client

c = Client('grpc://0.0.0.0:23456')

The URL-like scheme grpc://0.0.0.0:23456 is what you get after running the server. The scheme follows the format scheme://netloc:port:

Field

Description

Example

scheme

The protocol of the server, must be one of grpc, websocket, http, grpcs, websockets, https. Protocols end with s are TLS encrypted. This must match with the server protocol.

grpc

netloc

The server’s IP address or hostname

192.168.0.3

port

The public port of the server

51234

Encoding#

clip_client provides encode() function that allows you to send sentences, images to the server in a streaming and sync/async manner. Encoding here means getting the fixed-length vector representation of a text or image.

encode() supports two basic input types:

  • An iterable of str, e.g. List[str], Tuple[str], Generator[str] are all acceptable.

  • An iterable of Document, e.g. List[Document], DocumentArray, Generator[Document] are all acceptable.

Depending on the input, the output of encode() is different:

  • If the input is an iterable of str, then the output will be a numpy.ndarray.

  • If the input is an iterable of Document, then the output will be a DocumentArray.

Now let’s look at these two cases in details.

Input as iterable of strings#

  • Input: each string element is auto-detected as a sentence or an image.

  • Output: a [N, D] shape numpy.ndarray, where N is the length of the input and D is the CLIP embedding size. Each row corresponds to the embedding of the input object.

Any URI-like string, including relative, absolute file path, http/https path, data URI string will be considered as an image. Otherwise, it will be considered as a sentence.

For example,

from clip_client import Client

c = Client('grpc://0.0.0.0:23456')

r = c.encode(
    [
        'she smiled, with pain',
        'apple.png',
        'https://clip-as-service.jina.ai/_static/favicon.png',
        'data:image/gif;base64,R0lGODlhEAAQAMQAAORHHOVSKudfOulrSOp3WOyDZu6QdvCchPGolfO0o/XBs/fNwfjZ0frl3/zy7////wAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACH5BAkAABAALAAAAAAQABAAAAVVICSOZGlCQAosJ6mu7fiyZeKqNKToQGDsM8hBADgUXoGAiqhSvp5QAnQKGIgUhwFUYLCVDFCrKUE1lBavAViFIDlTImbKC5Gm2hB0SlBCBMQiB0UjIQA7',
    ]
)
print(r)

gives you

[[-0.09136295  0.42720157 -0.05784469 ... -0.42873043  0.04472527
   0.4437953 ]
 [ 0.43152636  0.1563695  -0.09363698 ... -0.11514216  0.1865044
   0.15025651]
 [ 0.42862126  0.17757078  0.08584607 ...  0.23284511 -0.00929402
   0.10993651]
 [ 0.4706376  -0.01384148  0.3877237  ...  0.1995864  -0.22621225
  -0.4837676 ]]

Input as iterable of Documents#

Tip

This feature uses DocArray, which is installed together with clip_client as an upstream dependency. You do not need to install DocArray separately.

If auto-detection on a list of raw string is too “sci-fi” to you, then you may use docarray.Document to make the input more explicit and organized. Document can be used as a container to easily represent a sentence or an image.

  • Input: each Document must be filled with .text or .uri or .blob or .tensor attribute.

    • Document filled with .text is considered as sentence;

    • Document filled with .uri or .blob or .tensor is considered as image. If .tensor is filled, then its shape must be in [H, W, C] format.

  • Output: a DocumentArray of the same input length. Each Document object in it is the same one from the input and is now filled with .embedding attribute. The order of the output is the same as the input.

Note

If the input Document is filled with both .text and .uri, then .text will be used.

Caution

The correctness of result and the order of output rely on the uniqueness of id of the input Document. The id will be implicitly generated if not provided. If you set the id manually, then you must make sure the id is unique, otherwise the results will not be complete.

The explicitness comes from now you have to put the content into the Document attributes. For example, we can rewrite the above example as below:

from clip_client import Client
from docarray import Document

c = Client('grpc://0.0.0.0:23456')

da = [
    Document(text='she smiled, with pain'),
    Document(uri='apple.png'),
    Document(uri='apple.png').load_uri_to_image_tensor(),
    Document(blob=open('apple.png', 'rb').read()),
    Document(uri='https://clip-as-service.jina.ai/_static/favicon.png'),
    Document(
        uri='data:image/gif;base64,R0lGODlhEAAQAMQAAORHHOVSKudfOulrSOp3WOyDZu6QdvCchPGolfO0o/XBs/fNwfjZ0frl3/zy7////wAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACH5BAkAABAALAAAAAAQABAAAAVVICSOZGlCQAosJ6mu7fiyZeKqNKToQGDsM8hBADgUXoGAiqhSvp5QAnQKGIgUhwFUYLCVDFCrKUE1lBavAViFIDlTImbKC5Gm2hB0SlBCBMQiB0UjIQA7'
    ),
]

r = c.encode(da)

Instead of sending a list of Document, you can also wrap it with a DocumentArray and then send it:

r = c.encode(DocumentArray(da))

Now that the return result is a DocumentArray, we can get a summary of it using r.summary().

╭──────────────────────────── Documents Summary ─────────────────────────────╮
│                                                                            │
│   Length                        6                                          │
│   Homogenous Documents          False                                      │
│   4 Documents have attributes   ('id', 'mime_type', 'uri', 'embedding')    │
│   1 Document has attributes     ('id', 'mime_type', 'text', 'embedding')   │
│   1 Document has attributes     ('id', 'embedding')                        │
│                                                                            │
╰────────────────────────────────────────────────────────────────────────────╯
╭────────────────────── Attributes Summary ───────────────────────╮
│                                                                 │
│   Attribute   Data type      #Unique values   Has empty value   │
│  ─────────────────────────────────────────────────────────────  │
│   embedding   ('ndarray',)   6                False             │
│   id          ('str',)       6                False             │
│   mime_type   ('str',)       5                False             │
│   text        ('str',)       2                False             │
│   uri         ('str',)       4                False             │
│                                                                 │
╰─────────────────────────────────────────────────────────────────╯

To get the embedding of all Documents, simply call r.embeddings:

[[-0.09136295  0.42720157 -0.05784469 ... -0.42873043  0.04472527
   0.4437953 ]
 [ 0.43152636  0.1563695  -0.09363698 ... -0.11514216  0.1865044
   0.15025651]
 [ 0.43152636  0.1563695  -0.09363698 ... -0.11514216  0.1865044
   0.15025651]
 [ 0.42862126  0.17757078  0.08584607 ...  0.23284511 -0.00929402
   0.10993651]
 [ 0.4706376  -0.01384148  0.3877237  ...  0.1995864  -0.22621225
  -0.4837676 ]]

Tip

Reading an image file into bytes and put into .blob is possible as shown above. However, it is often unnecessary. Especially if you have a lot of images, loading all of them into memory is not a good idea. Rule of thumb, always use .uri and trust clip_client to handle it well.

Async encoding#

To encode Document in an asynchronous manner, one can use aencode().

Tip

Despite the sexy word “async”, many data scientists have misconceptions about asynchronous behavior. And their motivation of using async function is often wrong. Async is not a silver bullet. In a simple language, you will only need .aencode() when there is another concurrent task that is also async. Then you want to “overlap” the time spending of these two tasks.

If your system is sync by design, there is nothing wrong about it. Go with encode() until you see a clear advantage of using aencode(), or until your boss tell you to do so.

In the following example, there is another job another_heavylifting_job to represent a job like writing to database, downloading large file.

import asyncio
from clip_client import Client

c = Client('grpc://0.0.0.0:23456')


async def another_heavylifting_job():
    # can be writing to database, downloading large file
    # big IO ops
    await asyncio.sleep(3)


async def main():
    t1 = asyncio.create_task(another_heavylifting_job())
    t2 = asyncio.create_task(c.aencode(['hello world'] * 100))
    await asyncio.gather(t1, t2)


asyncio.run(main())

The final time cost will be less than 3s + time(t2).

Ranking#

Tip

This feature is only available with clip_server>=0.3.0.

One can also rank cross-modal matches via rank() or arank(). First construct a cross-modal Document where the root contains an image and .matches contain sentences to rerank. One can also construct text-to-image rerank as below:

from docarray import Document

d = Document(
    uri='.github/README-img/rerank.png',
    matches=[
        Document(text=f'a photo of a {p}')
        for p in (
            'control room',
            'lecture room',
            'conference room',
            'podium indoor',
            'television studio',
        )
    ],
)
from docarray import Document

d = Document(
    text='a photo of conference room',
    matches=[
        Document(uri='.github/README-img/4.png'),
        Document(uri='.github/README-img/9.png'),
        Document(uri='https://clip-as-service.jina.ai/_static/favicon.png'),
    ],
)

Then call rank, you can feed it with multiple Documents as a list:

from clip_client import Client

c = Client(server='grpc://0.0.0.0:23456')
r = c.rank([d])

print(r['@m', ['text', 'scores__clip_score__value']])

Finally, in the return you can observe the matches are re-ranked according to .scores['clip_score']:

[['a photo of a television studio', 'a photo of a conference room', 'a photo of a lecture room', 'a photo of a control room', 'a photo of a podium indoor'], 
[0.9920725226402283, 0.006038925610482693, 0.0009973491542041302, 0.00078492151806131, 0.00010626466246321797]]

Indexing#

Tip

This feature is only available with clip_client>=0.7.0, and the server is running with a FLOW consisting of encoder and indexer.

You can index Documents via index() or aindex().

from clip_client import Client
from docarray import Document

c = Client('grpc://0.0.0.0:23456')

da = [
    Document(text='she smiled, with pain'),
    Document(uri='apple.png'),
    Document(uri='apple.png').load_uri_to_image_tensor(),
    Document(blob=open('apple.png', 'rb').read()),
    Document(uri='https://clip-as-service.jina.ai/_static/favicon.png'),
    Document(
        uri='data:image/gif;base64,R0lGODlhEAAQAMQAAORHHOVSKudfOulrSOp3WOyDZu6QdvCchPGolfO0o/XBs/fNwfjZ0frl3/zy7////wAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAACH5BAkAABAALAAAAAAQABAAAAVVICSOZGlCQAosJ6mu7fiyZeKqNKToQGDsM8hBADgUXoGAiqhSvp5QAnQKGIgUhwFUYLCVDFCrKUE1lBavAViFIDlTImbKC5Gm2hB0SlBCBMQiB0UjIQA7'
    ),
]

r = c.index(da)

Now that the return result is a DocumentArray, we can get a summary of it.

╭──────────────────────────── Documents Summary ─────────────────────────────╮
│                                                                            │
│   Length                        6                                          │
│   Homogenous Documents          False                                      │
│   4 Documents have attributes   ('id', 'mime_type', 'uri', 'embedding')    │
│   1 Document has attributes     ('id', 'mime_type', 'text', 'embedding')   │
│   1 Document has attributes     ('id', 'embedding')                        │
│                                                                            │
╰────────────────────────────────────────────────────────────────────────────╯
╭────────────────────── Attributes Summary ───────────────────────╮
│                                                                 │
│   Attribute   Data type      #Unique values   Has empty value   │
│  ─────────────────────────────────────────────────────────────  │
│   embedding   ('ndarray',)   6                False             │
│   id          ('str',)       6                False             │
│   mime_type   ('str',)       5                False             │
│   text        ('str',)       2                False             │
│   uri         ('str',)       4                False             │
│                                                                 │
╰─────────────────────────────────────────────────────────────────╯

The embedding is the output of the encoder, which is a 512-dim vector. Now we can use the indexer to search for the indexed Documents.

Searching#

Tip

This feature is only available with clip_client>=0.7.0, and the server is running with a FLOW consisting of encoder and indexer.

You can use search() or asearch() to search for relevant Documents in the index for a given query.

from clip_client import Client

c = Client('grpc://0.0.0.0:23456')

result = c.search(['smile'], limit=2)

print(result['@m', ['text', 'scores__cosine']])

The results will look like this, the most relevant doc is “she smiled, with pain” with the cosine distance of 0.096. And the apple image has the cosine distance of 0.799.

[['she smiled, with pain', ''], [{'value': 0.09604918956756592}, {'value': 0.7994111776351929}]]

You can set the limit parameter (default is 10) to control the number of the most similar documents to be retrieved.

Profiling#

You can use profile() to give a quick test on the server to make sure everything is good.

from clip_client import Client

c = Client('grpc://0.0.0.0:23456')

c.profile()

This give you a tree-like table showing the latency and percentage.

 Roundtrip  16ms  100%                                                          
├──  Client-server network  12ms  75%                                           
└──  Server  4ms  25%                                                           
    ├──  Gateway-CLIP network  0ms  0%                                          
    └──  CLIP model  4ms  100%      

Under the hood, .profile() sends a single empty Document to the CLIP-server for encoding and calculates a summary of latency. The above tree can be read as follows:

  • From calling client.encode() to returning the results, everything counted, takes 16ms to finish.

  • Among them the time spent on the server is 4ms, the remaining 12ms is spent on the client-server communication, request packing, response unpacking.

  • During the 4ms server processing time, CLIP model takes 4ms, whereas the Gateway to CLIP communication takes no time.

.profile() can also take a string argument and asks CLIP-server to encode it. This string can be a sentence, local/remote image file URI. For example:

c.profile('hello, world')
c.profile('apple.png')
c.profile('https://docarray.jina.ai/_static/favicon.png')

Single query latency is often very fluctuated. Running .profile() multiple times may give you different results. Nonetheless, it helps you understand who to blame if CLIP-as-service is running slow for you: the network? the computation? But certainly not this software itself.

Best practices#

In this section, we will show you some best practices for using this client. We will use encoding as an example. The same applies to all other methods.

Control batch size#

You can specify .encode(..., batch_size=8) to control how many Documents are sent in each request. You can play this number and find the perfect balance between network transmission and GPU utilization.

Intuitively, setting batch_size=1024 should result in very high GPU utilization on each request. However, a large batch size like this also means sending each request would take longer. Given that clip-client is designed with request and response streaming, large batch size would not benefit from the time overlapping between request streaming and response streaming.

Control prefetch size#

To control the number of in-flight batches, you can use the .encode(..., prefetch=100) option. The way this works is that when you send a large request, the outgoing request stream will usually finish before the incoming response stream due to the asynchronous design. This is because the request handling is typically time-consuming, which can prevent the server from sending back the response and may cause it to close the connection as it thinks the incoming channel is idle. By default, the client is set to a prefetch value of 100. However, it is recommended to use a lower value for expensive operations and a higher value for faster response times.

For more information about client prefetching, please refer to Rate Limit in Jina documentation.

Show progressbar#

You can use .encode(..., show_progress=True) to turn on the progress bar.

../../_images/client-pgbar.gif

Hint

Progress bar may not show up in the PyCharm debug terminal. This is an upstream issue of rich package.

Processing large number of Documents#

Here are some suggestions when encoding a large number of Documents:

  1. Use Generator as input to load data on-demand. You can put your data into a Generator and feed to .encode:

    def data_gen():
        for _ in range(100_000):
            yield Document(uri=...)
    
    
    c = Client(...)
    c.encode(data_gen())
    

    Yield raw strings is also acceptable, e.g. to encode all images under a directory, you can simply do:

    from glob import iglob
    
    c.encode(iglob('**/*.png'))
    
  2. Adjust the batch_size parameters.

  3. Adjust the prefetch parameters.

  4. Turn on the progressbar.

Danger

In any case, avoiding the following coding:

for d in big_list:
    c.encode([d])

This is extremely slow as only one document is encoded at a time, it is a bad utilization of the network and not leveraging any duplex streaming.

Custom callback#

clip_client by default collects all the results and returns them to users. However, if you want to process the results on-the-fly, you can also pass a callback function when sending the request. For example, you can use the callback to save the results to a database, or render the results to a webpage. Specifically, you can specify any of the three callback functions: on_done, on_error, and on_always.

  • on_done is executed while streaming, after successful completion of each request

  • on_error is executed while streaming, whenever an error occurs in each request

  • on_always is always performed while streaming, no matter the success or failure of each request

Note that these callbacks only work for requests (and failures) inside the stream. For on_error, if the failure is due to an error happening outside of streaming, then it will not be triggered. For example, a SIGKILL from the client OS during the handling of the request, or a networking issue, will not trigger the callback. Learn more about handling exceptions in on_error.

Callback functions take a Response of the type DataRequest, which contains resulting Documents, parameters, and other information. Learn more about handling DataRequest in callbacks.

In the following example, we will use on_done to save the results to a database. We use a simple dict to simulate the database. The error is saved to log file using on_error. on_always will print the number of documents processed in each request.

from clip_client import Client

db = {}


def my_on_done(resp):
    for doc in resp.docs:
        db[doc.id] = doc


def my_on_error(resp):
    with open('error.log', 'a') as f:
        f.write(resp)


def my_on_always(resp):
    print(f'{len(resp.docs)} docs processed')


c = Client('grpc://0.0.0.0:12345')
c.encode(
    ['hello', 'world'], on_done=my_on_done, on_error=my_on_error, on_always=my_on_always
)

Note

If either on_done or on_always is specified, the default behavior of returning the results is disabled. You need to handle the results yourself.

Client parallelism#

In case you instanciate a clip_client object using the grpc protocol, keep in mind that grpc clients cannot be used in a multi-threaded environment (check this gRPC issue for reference). What you should do, is to rely on asynchronous programming or multi-processing rather than multi-threading.

To use clip_client in a Flask application, you can introduce multi-processing based parallelism to your app using gunicorn:

gunicorn -w 4 -b 127.0.0.1:4000 myproject:app

To use clip_client in a FastAPI application, you have to manually restrict the thread number to 1 at the starting state of the app:

import uvicorn
from fastapi import FastAPI
from clip_client import Client
from anyio.lowlevel import RunVar
from anyio import CapacityLimiter

c = Client('grpc://0.0.0.0:51001')
app = FastAPI()

@app.on_event("startup")
def startup():
    print("start")
    RunVar("_default_thread_limiter").set(CapacityLimiter(1))

@app.post("/")
def encode():
    r =  c.encode(['Hello world', 'Hello Jina'])
    print(r)

Then it can run with multiprocessing using

gunicorn myproject:app --workers 4 --worker-class uvicorn.workers.UvicornWorker --bind 0.0.0.0:4000

Appendix: Plain HTTP request via curl#

Tip

Sending large embeddings over plain HTTP is often not the best idea. Websocket is often a better choice, allows one to call clip-server from Javascript with much better performance.

If your server is spawned with protocol: http and cors: True, then you do not need to call the server via Python client. You can simply do it via curl or Javascript by sending a JSON to http://address:port/post. Notice, the /post endpoint at the end. For example,

To encode sentences:

curl -X POST http://0.0.0.0:51000/post \ 
     -H 'Content-Type: application/json' \
     -d '{"data":[{"text": "First do it"}, {"text": "then do it right"}, {"text": "then do it better"}], "execEndpoint":"/"}'

To encode a local image, you need to load it as base64 string and put into the blob field, and be careful with the quotes there:

curl -X POST http://0.0.0.0:51000/post \ 
     -H 'Content-Type: application/json' \
     -d '{"data":[{"text": "First do it"}, {"blob":"'"$( base64 test-1.jpeg)"'" }], "execEndpoint":"/"}'

To encode a remote image, you can simply put its address into uri field:

curl -X POST http://0.0.0.0:51000/post \ 
     -H 'Content-Type: application/json' \
     -d '{"data":[{"text": "First do it"}, {"uri": "https://clip-as-service.jina.ai/_static/favicon.png"}], "execEndpoint":"/"}'

Run it, you will get:

{"header":{"requestId":"8b1f4b419bc54e95ab4b63cc086233c9","status":null,"execEndpoint":"/","targetExecutor":""},"parameters":null,"routes":[{"executor":"gateway","startTime":"2022-04-01T15:24:28.267003+00:00","endTime":"2022-04-01T15:24:28.328868+00:00","status":null},{"executor":"clip_t","startTime":"2022-04-01T15:24:28.267189+00:00","endTime":"2022-04-01T15:24:28.328748+00:00","status":null}],"data":[{"id":"b15331b8281ffde1e9fb64005af28ffd","parent_id":null,"granularity":null,"adjacency":null,"blob":null,"tensor":null,"mime_type":"text/plain","text":"hello, world!","weight":null,"uri":null,"tags":null,"offset":null,"location":null,"embedding":[-0.022064208984375,0.1044921875, ..., -0.1363525390625,-0.447509765625],"modality":null,"evaluations":null,"scores":null,"chunks":null,"matches":null}]}

The embedding is inside .data[].embedding. If you have jq installed, you can easily filter the embeddings out via:

curl -X POST http://0.0.0.0:51000/post \
     -H 'Content-Type: application/json' \
     -d '{"data":[{"text": "hello, world!"}, {"blob":"'"$( base64 test-1.jpeg)"'" }], "execEndpoint":"/"}' | \
     jq -c '.data[] | .embedding'
[-0.022064208984375,0.1044921875,...]
[-0.0750732421875,-0.166015625,...]