Skip to main content

No project description provided

Project description

Marian Client

A client for interacting with Marian-NMT websocket server (see MarianNMT).

Marian is a state of the art Neural Machine Translation framework written in C++. Eventually, someone should write Python bindings (not it!). Until then, the recommended way of communicating with Marian is via WebSockets. The authors of Marian provide an example of doing this with this script. If you need Python2 support, you should look at that script.

What this project contributes beyond the above script:

  • Persistent connection - This keeps a connection open between Marian Server and Python. This saves a few hundred ms per call, which is significant
  • Encapsulation - just import a class, instantiate, and call. Don't think about websockets at all
  • Timeout, retries, and error handling - websockets are not the most reliable. Connections fail, timeouts happen. Just pass a value for timeout and retries when you instantiate MarianClient and this will just be handled for you.


pip install marian-client


from marian_client import MarianClient

# These are the default values:
host = "localhost"
port = "8080"

# or give the fully qualified URL
url = "ws://my.marian.server.ip/translate"

# Default values
timeout = 30  # measured in seconds - you may want to make this much lower
retries = 3  # amount of times to retry on error. backs off exponentially.

debug = False  # set to True for a little more info on errors

mc = MarianClient(PORT=port, HOST=host, timeout=timeout, retries=retries, debug=debug)
# or if you want to specify url
# mc = MarianClient(url, timeout=timeout, retries=retries, debug=debug)

# if you just want all the default values, and marian-server is running locally:
# mc = MarianClient()

tokenized_sentence = "Alice like cats ."

success, corrected_sentence, error_info = mc(tokenized_sentence)

if success:
    print(f"Call to MarianClient failed with error code {error_info[0]} and message {error_info[1]}")

# If marian-server is sert up and working, this prints
# >>> "Alice likes cats ."


  • When instantiating a MarianClient instance, if we receive a ConnectionRefusedError, we attempt to reconnect connection_retries times, with exponential backoff, maxing out at max_wait_time_between_connection_attempts.
  • This means in the default case, if Marian Server is unavailable, we will try to connect, wait 1 second, try to connect again, wait 2 seconds, try to connect again, wait 4 seconds, ... then 8, 16, 32, 64, 128, 256, 300, then actually fail, for a total wait time of 811 seconds.


Like Marian, this package is released under the MIT license.


This package was made by the NLP team at Qordoba. If you are using it, and interested in working on NLP, maybe reach out to Sam?

Thanks to Marcin Junczys-Dowmunt and the rest of the awesome authors of Marian!

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

marian-client-0.16.0.tar.gz (7.2 kB view hashes)

Uploaded source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page