Multiple TranslationModels Implementation #210

jerinphilip · 2021-07-30T12:37:56Z

Fixes #209. Please note that accompanying changes in BRT (browsermt/bergamot-translator-tests#43) will have to be reviewed in context of this PR.

Separate TranslationModel, Threaded vs Non-Threaded workflows operating with TranslationModel
WASM Bindings temporary editing to get green on CI. Let's pretend this works (it most likely does, FWIW).
Marian's config-parsing doesn't support multiple configs to be used in non-trivial workflows (pivoting, multiple models; marian --configs is nested env); CLI Rework. Bergamot has it's own CLI now. Individual TranslationModel configs are supplied as --model-configs <path> for testing purposes. Library code can load from std::string &config or Ptr<Options> options = TranslationModel::Config as before.
Tests in BRT for Outbound translation, demonstrating the modifications work for Outbound Translation Requirements #78.

… be destructed

…n't have sentence level anymore

…ing> is not cool

…ased signalling mechanism, not batchers

Should work for both BatchingPool and AggregateBatchingPool. This ensures that what previously worked with BatchingPool should work with AggregateBatchingPool as well.

Two different state-machines are in separate classes now: * BlockingService: Blocking and operations happening sequentially. This class is Service for WASM. * AsyncService: Asynchronous and operates through callbacks. Multithreading used to achieve async operations. The thread-safe wrapper for batching-pool has been reset to wrap around the underlying container, unifying API in both while maintaining differenet arguments. Access to the unsafe underlying storage is forbidden in AsyncService through making the container private in the wrapped class.

src/translator/translation_model.cpp

src/translator/service.h

src/translator/translation_model.cpp

src/translator/translation_model.h

src/translator/aggregate_batching_pool.cpp

src/translator/aggregate_batching_pool.h

src/translator/threadsafe_batcher.cpp

src/translator/translation_model.cpp

kpu

Unbounded memory consumption.

src/translator/translation_model.h

…h) remain

…slationModel and Service

…nslator into multiple-models

… BRT updated accordingly

…default entered (config)

…or consistency

Jerin Philip added 16 commits July 30, 2021 11:47

Adding a TranslationModel implementation

47ddedf

Projected service methods to take arg: TranslationModel

db2b8db

Prettify: CallbackType alias, some ifdef fixes

c715b48

Code runs without SIGSEGV now, transfer to TranslationModel functional

7492f8b

BatchTranslator class is removed, long live TranslationModel?!

7a6d3b5

loadBackend(...) separated, should improve diff view

33ff241

Removing default arguments and making illegal values aborts

f2e50be

Translation& -> shared_ptr<...>; If it's in the process, it shouldn't…

7481579

… be destructed

BatchingPool (previously Batcher) aggregation with AggregateBatchingPool

c90ecd7

Still carrying WASM/blocking workflow

b9e3b62

We *really* should get rid of these multiple ifdef paths

99d3e47

WASM_COMPATIBLE_SOURCE=on|off double build exercise

68f13a3

[empty-commit] Breaking API after this

ab44bee

translate(string, opts) -> translate(model, string, opts)

59f8c58

Batcher -> BatchingPool; addWholeRequest -> addRequest, because we do…

5363998

…n't have sentence level anymore

Fixing semantics; typedef <thread-unsafe-something> Threadsafe<someth…

22a591b

…ing> is not cool

jerinphilip changed the title ~~Multiple TranslationModels~~ Multiple TranslationModels Implementation Aug 3, 2021

This was referenced Aug 4, 2021

First class pivot translation capability #212

Closed

Provision of new model(s) for testing multiple model workflows browsermt/students#43

Closed

Jerin Philip added 6 commits August 5, 2021 12:08

Renaming files to reflect classes included and reorg

cb11583

Adding renamed files and fixing includes

979501c

Empty comment line removal

fa1cfc5

Isolate properly: Explicit shutdown is only required for the thread-b…

07bb9d5

…ased signalling mechanism, not batchers

More reorg + doc for translation_model.h

bdb610e

ThreadsafeBatcher templated for BatchingPool type

48a28c2

Should work for both BatchingPool and AggregateBatchingPool. This ensures that what previously worked with BatchingPool should work with AggregateBatchingPool as well.

jerinphilip force-pushed the multiple-models branch from 68ae377 to 48a28c2 Compare August 5, 2021 15:17

Jerin Philip added 4 commits August 6, 2021 11:25

Adding comments to explain the template based generalization

a53c44c

Service comments marking Async/Blocking common code

dc3f1ab

Removing possibly unnecessary WASM_COMPATIBLE_SOURCE

b78a2f5