DA library will help you to easily adapt a language model to your own data domain or task.
Both domain adaptation (e.g. Beltagy, 2019) and task adaptation (e.g. Gururangan, 2020) are reported to consistently improve quality of the language models on end tasks, and improve model's comprehension on more niche domains, suggesting that it's might always be a good idea to adapt before the final fine-tuning. However, it is still not a common practice, maybe because it is still a tedious thing to do. A multi-step, or multi-objective training requires a separate configuration of every training step due to the differences in the models' architectures specific to the chosen training objective and data set.
Domain Adaptation (DA) framework abstracts the term of Objective away from the model.
With DA, Any objective can be applied to any model, for as long as the trained model has some head of compatible shape.
The ordering in which the Objectives are applied is determined by the given Schedule instance.
Conventionally, the objectives are applied sequentially (that's what SequentialSchedule does),
but they might as well be applied in a rotation (StridedSchedule), or balanced dynamically,
e.g. according to its objectives` losses (coming soon).
In the DA framework, instead of providing the Trainer with a model and a dataset,
a user constructs a Schedule composed of the initialised Objectives, where each Objective performs its
dataset sampling and objective-specific feature alignment (compliant with objective.compatible_head).
Additionally, all objectives share the three common model Head types (Sequence classification, Token classification and Language modelling),
which minimises the number of different data formats that the user has to maintain.
DA introduces objective-centric, instead of model-centric approach to the training process, that makes it easier to experiment with trainings on multiple objectives. Thanks to that, you can do some things, that are difficult, or impossible in other NLP frameworks (like HF Transformers, FairSeq or SciPy). For example:
- Domain adaptation or Task adaptation: thanks to DA, you do not have to handle the model between different training scripts, minimising a chance of error and improving reproducibility
- Seamlessly experiment with different schedule strategies, allowing you, e.g. to backpropagate based on multiple objectives in every training step
- Track the progress of the model, concurrently on each relevant objective, allowing you to easier recognise weak points of your model
- Easily perform Multi-task learning, which reportedly improves model robustness
- Although DA aims primarily for adapting the models of the transformer family, the library is designed to work with any PyTorch model
Built upon the well-established and maintained 🤗 Transformers library, DA library will automatically support future new NLP models out-of-box. The adaptation of DA to a different version of Hugging Face Transformers library should not take longer than a few minutes.
First, install the library. If you clone it, you can also use the provided example scripts.
git clone {this repo}
cd DA
python -m pip install -e .You can find and run the full examples from below in tests/end2end_usecases_test.py folder.
Say you have nicely annotated entities in a set of news articles, but eventually, you want to use the language model to detect entities in invoices, for example. You can either train the NER model on news articles, hoping that it will not lose much accuracy on other domains. Or you can concurrently train on both data sets:
# 1. pick the models - randomly pre-initialize the appropriate heads
lang_module = LangModule(test_base_models["token_classification"])
# 2. pick objectives
# Objectives take either List[str] for in-memory iteration, or a source file path for streamed iteration
objectives = [MaskedLanguageModeling(lang_module,
batch_size=1,
texts_or_path="tests/mock_data/domain_unsup.txt"),
TokenClassification(lang_module,
batch_size=1,
texts_or_path="tests/mock_data/ner_texts_sup.txt",
labels_or_path="tests/mock_data/ner_texts_sup_labels.txt")]
# 3. pick a schedule of the selected objectives
# This one will initially fit the first objective to the convergence on its eval set, fit the second
schedule = SequentialSchedule(objectives, training_arguments)
# 4. Run the training using Adapter, similarly to running HF.Trainer, only adding `schedule`
adapter = Adapter(lang_module, schedule, training_arguments)
adapter.train()
# 5. save the trained lang_module (with all heads)
adapter.save_model("entity_detector_model")
# 6. reload and use it like any other Hugging Face model
ner_model = AutoModelForTokenClassification.from_pretrained("entity_detector_model")
tokenizer = AutoTokenizer.from_pretrained("entity_detector_model")
inputs = tokenizer("Is there any Abraham Lincoln here?")
outputs = ner_model(**inputs)
print(tokenizer.batch_decode(outputs))Say you have a lot of clean parallel texts for news articles (just like you can find on OPUS), but eventually, you need to translate a different domain, for example chats with a lot of typos, or medicine texts with a lot of latin expressions.
# 1. pick the models - randomly pre-initialize the appropriate heads
lang_module = LangModule(test_base_models["translation"])
# (optional) pick train and validation evaluators for the objectives
seq2seq_evaluators = [BLEU(decides_convergence=True)]
# 2. pick objectives - we use BART's objective for adaptation and mBART's seq2seq objective for fine-tuning
objectives = [DenoisingObjective(lang_module,
batch_size=1,
texts_or_path="tests/mock_data/domain_unsup.txt",
val_evaluators=seq2seq_evaluators),
DecoderSequence2Sequence(lang_module, batch_size=1,
texts_or_path=paths["texts"]["target_domain"]["translation"],
val_evaluators=seq2seq_evaluators,
labels_or_path=paths["labels"]["target_domain"]["translation"],
source_lang_id="en", target_lang_id="cs")]
# 2. pick objectives - we use BART's objective for adaptation and mBART's seq2seq objective for fine-tuning
objectives = [DenoisingObjective(lang_module,
batch_size=1,
texts_or_path="tests/mock_data/domain_unsup.txt"),
DecoderSequence2Sequence(lang_module,
batch_size=1,
texts_or_path="tests/mock_data/seq2seq_sources.txt",
labels_or_path="tests/mock_data/seq2seq_targets.txt",
val_evaluators=seq2seq_evaluators,
source_lang_id="en", target_lang_id="cs")]
# 3. pick a schedule of the selected objectives
# this one will shuffle the batches of both objectives
schedule = StridedSchedule(objectives, training_arguments)
# 4. train using Adapter
adapter = Adapter(lang_module, schedule, training_arguments)
adapter.train()
# 5. save the trained (multi-headed) lang_module
adapter.save_model("translator_model")
# 6. reload and use it like any other Hugging Face model
translator_model = AutoModelForSeq2SeqLM.from_pretrained("translator_model/DecoderSequence2Sequence")
tokenizer = AutoTokenizer.from_pretrained("translator_model/DecoderSequence2Sequence")
tokenizer.src_lang, tokenizer.tgt_lang = "en", "cs"
# 7. use the model anyhow you like, e.g. as a translator with iterative generation
inputs = tokenizer("A piece of text to translate.", return_tensors="pt")
output_ids = translator_model.generate(**inputs)
output_text = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
print(output_text)Try this example with training resources resolution from OPUS in examples/machine_translation/train_wiki_adapt_bible.py
but contributions are welcome :) (see How can you contribute below)
We've seen that transformers can outstandingly model very complicated tasks, which makes us think that experimenting with more complex training objectives can also improve their desperately-needed generalisation abilities (many studies report transformers inability to generalise the end task, e.g. on language inference, paraphrase detection, or machine translation).
This way, we're also hoping to enable the easy use of the most accurate deep language models for more specialised domains of application, where a little supervised data is available, but much more unsupervised sources can be found (a typical Domain adaptation case). Such applications include for instance machine translation of non-canonical domains (chats or expert texts) or personal names recognition in texts of a domain with none of its own labeled names, but the use-cases are limitless.
New and exciting objectives appear in NLP papers every day, and the DA library aims to make it as simple as possible to add them! If you'd like to add a new Objective in DA follow these steps:
- Implement it: pick the logically-best-matching abstract objective from
objectives/objective_base.py, and implement the remaining abstract methods. - Test it: add a simple test for your objective to
tests/objectives_test.py, that will passassert_module_objective_ok. - End-to-end-test it: add a test to
end2end_usecases_test.pyto show the others the complete demonstration on how to use the objective in a meaningful way - (optional): Create an example that will apply the objective in a real training process, on real data.
See other examples in
examplesfolder. - Share! Create a PR or issue here in GitHub with a link to your fork, and we'll happily take a look!
If you have any other question(s), feel free to ping a message to xxx@xxx.xx