Server

Models

class mammoth.translate.translation_server.ServerModel(opts, model_id, preprocess_opt=None, tokenizer_opt=None, postprocess_opt=None, custom_opt=None, load=False, timeout=-1, on_timeout='to_cpu', model_root='./', ct2_model=None, ct2_translator_args=None, ct2_translate_batch_args=None)

Bases: object

Wrap a model with server functionality.

Parameters:

• opts (dict) – Options for the Translator

• model_id (int) – Model ID

• preprocess_opt (list) – Options for preprocess processus or None

• tokenizer_opt (dict) – Options for the tokenizer or None

• postprocess_opt (list) – Options for postprocess processus or None

• custom_opt (dict) – Custom options, can be used within preprocess or postprocess, default None

• load (bool) – whether to load the model during __init__()

• timeout (int) – Seconds before running do_timeout() Negative values means no timeout

• on_timeout (str) – Options are [“to_cpu”, “unload”]. Set what to do on timeout (see do_timeout().)

• model_root (str) – Path to the model directory it must contain the model and tokenizer file

build_tokenizer(tokenizer_opt)

Build tokenizer described by tokenizer_opt.

detokenize(sequence, side='tgt')

Detokenize a single sequence

Same args/returns as tokenize()

do_timeout()

Timeout function that frees GPU memory.

Moves the model to CPU or unloads it; depending on attr`self.on_timemout` value

maybe_convert_align(src, tgt, align)

Convert alignment to match detokenized src/tgt (or not).

Parameters:

• src (str) – The tokenized source sequence.

• tgt (str) – The tokenized target sequence.

• align (str) – The alignment correspand to src/tgt pair.

Returns:

The alignment correspand to detokenized src/tgt.

Return type:

align (str)

maybe_detokenize(sequence, side='tgt')

De-tokenize the sequence (or not)

Same args/returns as tokenize()

maybe_detokenize_with_align(sequence, src, side='tgt')

De-tokenize (or not) the sequence (with alignment).

Parameters:

• sequence (str) – The sequence to detokenize, possibly with alignment separated by |||.

• src – Source sequence.

• side (str) – Side to detokenize ('tgt' or 'src').

Returns:

A tuple of (sequence, align) where sequence is the detokenized string and align is the alignment string or None.

maybe_postprocess(sequence)

Postprocess the sequence (or not)

maybe_preprocess(sequence)

Preprocess the sequence (or not)

maybe_tokenize(sequence, side='src')

Tokenize the sequence (or not).

Same args/returns as tokenize

parse_opt(opts)

Parse the option set passed by the user using mammoth.opts

Parameters:

opts (dict) – Options passed by the user

Returns:

full set of options for the Translator

Return type:

opts (argparse.Namespace)

postprocess(sequence)

Preprocess a single sequence.

Parameters:

sequence (str) – The sequence to process.

Returns:

The postprocessed sequence.

Return type:

sequence (str)

preprocess(sequence)

Preprocess a single sequence.

Parameters:

sequence (str) – The sequence to preprocess.

Returns:

The preprocessed sequence.

Return type:

sequence (str)

rebuild_seg_packages(all_preprocessed, results, scores, aligns, n_best)

Rebuild proper segment packages based on initial n_seg.

to_gpu()

Move the model to GPU.

tokenize(sequence, side='src')

Tokenize a single sequence.

Parameters:

sequence (str) – The sequence to tokenize.

Returns:

The tokenized sequence.

Return type:

tok (str)

tokenizer_marker(side='src')

Return marker used in side tokenizer.

Core Server

exception mammoth.translate.translation_server.ServerModelError

Bases: Exception

class mammoth.translate.translation_server.Timer(start=False)

Bases: object

class mammoth.translate.translation_server.TranslationServer

Bases: object

clone_model(model_id, opts, timeout=-1)

Clone a model model_id.

Different options may be passed. If opts is None, it will use the same set of options

list_models()

Return the list of available models

load_model(opts, model_id=None, **model_kwargs)

Load a model given a set of options

preload_model(opts, model_id=None, **model_kwargs)

Preloading the model: updating internal datastructure

It will effectively load the model if load is set

run(inputs)

Translate inputs

We keep the same format as the Lua version i.e. [{"id": model_id, "src": "sequence to translate"},{ ...}]

We use inputs[0][“id”] as the model id

start(config_file)

Read the config file and pre-/load the models.

unload_model(model_id)

Manually unload a model.

It will free the memory and cancel the timer