Skip to content

Latest commit

 

History

History

api

README.md

External Interfaces

This package contains the interfaces that can be used by external tools to access and use fairseq models.

Simple IO Server

This server consists in a simple, single-thread server that serves a speech translation model by accepting requests by reading one request per line from the stdin (standard input) and answering the request in the standard output.

API

Each request in the stdin should be formulated with a JSON data structure that must be put on a single, self-contained line. Two types of requests are supported: commands, and inputs. The only command supported so far is shutdown i.e., the server can be shutdown by the client by sending the following request:

{"command": "shutdown"}

The input requests instead contain three field: wav_path, src_lang, and tgt_lang. An example of a valid request is:

{"wav_path": "/this/is/the/path/to/a/wav/file.wav", "src_lang":  "en", "tgt_lang":  "it"}

Different processors may add other fields. Please refer to the specific processor for more details. Currently, the processors are st, and st_triangle that both use only those fields. The server produces in the stdout (standard output) a different response according to the processor used, but it is always formatted as a single line JSON. In the case of st_triangle, the response fields are score, translation, and transcript that respectively represent i) the model confidence on the generated output(s) expressed as log-probability, ii) the most likely translation, and iii) the most likely transcript, e.g.

{"score": -2.4353, "translation":  "Il contenuto audio", "transcript":  "The audio content"}

Usage

The server is meant to serve a specific model that is loaded at server start time. To be started, the server accepts the same arguments of the stardard fairseq generate, allowing to control all the aspects of the generation in the same way. Namely, it requires a checkpoint of the model, the dictionaries, a YAML configuration file, in addition to the generation arguments. An example of the command to start the server is:

dir_with_dictionaries_and_yaml=$1
model_path=$2
conf_yaml=$3

python api/simple_io_server_st.py $data_bin \
    --user-dir examples/speech_to_text \
    --config-yaml $conf_yaml \
    --max-tokens 10000 \
    --scoring sacrebleu --beam 10 \
    --path $model_path \
    --max-source-positions 10000 --max-target-positions 1000 \
    --task speech_to_text_tagged_dual \
    --server-processor st_triangle

Limitations

  • The server is single-thread and accepts only ONE request per time.
  • The server can serve only one model. To serve more models, start more servers.