Skip to content
/ trapper Public

State-of-the-art NLP through transformer models in a modular design and consistent APIs.

License

Notifications You must be signed in to change notification settings

obss/trapper

Repository files navigation

Trapper (Transformers wRAPPER)

Python versions PyPI version Latest Release Open in Colab
Build status Dependencies Code style: black License: MIT

Trapper is an NLP library that aims to make it easier to train transformer based models on downstream tasks. It wraps huggingface/transformers to provide the transformer model implementations and training mechanisms. It defines abstractions with base classes for common tasks encountered while using transformer models. Additionally, it provides a dependency-injection mechanism and allows defining training and/or evaluation experiments via configuration files. By this way, you can replicate your experiment with different models, optimizers etc by only changing their values inside the configuration file without writing any new code or changing the existing code. These features foster code reuse, less boiler-plate code, as well as repeatable and better documented training experiments which is crucial in machine learning.

Why You Should Use Trapper

  • You have been a Transformers user for quite some time now. However, you started to feel that some computation steps could be standardized through new abstractions. You wish to reuse the scripts you write for data processing, post-processing etc with different models/tokenizers easily. You would like to separate the code from the experiment details, mix and match components through configuration files while keeping your codebase clean and free of duplication.

  • You are an AllenNLP user who is really happy with the dependency-injection system, well-defined abstractions and smooth workflow. However, you would like to use the latest transformer models without having to wait for the core developers to integrate them. Moreover, the Transformers community is scaling up rapidly, and you would like to join the party while still enjoying an AllenNLP touch.

  • You are an NLP researcher / practitioner, and you would like to give a shot to a library aiming to support state-of-the-art models along with datasets, metrics and more in unified APIs.

To see more, check the official Trapper blog post.

Key Features

Compatibility with HuggingFace Transformers

Trapper extends Transformers!

While implementing the components of trapper, we try to reuse the classes from the Transformers library as much as we can. For example, trapper uses the models, and the trainer as they are in Transformers. This makes it easy to use the models trained with trapper on other projects or libraries that depend on Transformers (or pytorch in general).

We strive to keep trapper fully compatible with Transformers, so you can always use some of our components to write a script for your own needs while not using the full pipeline (e.g. for training).

Dependency Injection and Training Based on Configuration Files

We use the registry mechanism of AllenNLP to provide dependency injection and enable reading the experiment details from the configuration files in json or jsonnet format. You can look at the AllenNLP guide on dependency injection to learn more about how the registry system and dependency injection works as well as how to write configuration files. In addition, we strongly recommend reading the remaining parts of the AllenNLP guide to learn more about its design philosophy, the importance of abstractions etc. (especially Part2: Abstraction, Design and Testing). As a warning, please note that we do not use AllenNLP's abstractions and base classes in general, which means you can not mix and match the trapper's and AllenNLP's components. Instead, we just use the class registry and dependency injection mechanisms and only adapt its very limited set of components, first by wrapping and registering them as trapper components. For example, we use the optimizers from AllenNLP since we can conveniently do so without hindering our full compatibility with Transformers.

Full Integration with HuggingFace Datasets

In trapper, we officially use the format of the datasets from datasets and provide full integration with it. You can directly use all datasets published in datasets hub without doing any extra work. You can write the dataset name and extra loading arguments (if there are any) in your training config file, and trapper will automatically download the dataset and pass it to the trainer. If you have a local or private dataset, you can still use it after converting it to the HuggingFace datasets format by writing a dataset loading script as explained here.

Support for Metrics through Jury

Trapper supports the common NLP metrics through jury. Jury is an NLP library dedicated to provide metric implementations by adopting and extending the datasets library. For metric computation during training you can use jury style metric instantiation/configuration to set up on your trapper configuration file to compute metrics on the fly on eval dataset with a specified eval_steps value. If your desired metric is not yet available on jury or datasets, you can still create your own by extending trapper.Metric and utilizing either jury.Metric or datasets.Metric for handling larger set of cases on predictions.

Abstractions and Base Classes

Following AllenNLP, we implement our own registrable base classes to abstract away the common operations for data processing and model training.

  • Data reading and preprocessing base classes including

    • The classes to be used directly: DatasetReader, DatasetLoader and DataCollator.

    • The classes that you may need to extend: LabelMapper,DataProcessor , DataAdapter and TokenizerWrapper.

    • TokenizerWrapper classes utilizing AutoTokenizer from Transformers are used as factories to instantiate wrapped tokenizers into which task-specific special tokens are registered automatically.

  • ModelWrapper classes utilizing the AutoModelFor... classes from Transformers are used as factories to instantiate the actual task-specific models from the configuration files dynamically.

  • Optimizers from AllenNLP: Implemented as children of the base Optimizer class.

  • Metric computation is supported through jury. In order to make the metrics flexible enough to work with the trainer in a common interface, we introduced metric handlers. You may need to extend these classes accordingly

    • For conversion of predictions and references to a suitable form for a particular metric or metric set: MetricInputHandler.
    • For manipulating resulting score object containing the metric results: MetricOutputHandler.

Usage

To use trapper, you need to select the common NLP formulation of the problem you are tackling as well as decide on its input representation, including the special tokens.

Modeling the Problem

The first step in using trapper is to decide on how to model the problem. First, you need to model your problem as one of the common modeling tasks in NLP such as seq-to-seq, sequence classification etc. We stick with the Transformers' way of dividing the tasks into common categories as it does in its AutoModelFor... classes. To be compatible with Transformers and reuse its model factories, trapper formalizes the tasks by wrapping the AutoModelFor... classes and matching them to a name that represents a common task in NLP. For example, the natural choice for POS tagging is to model it as a token classification (i.e. sequence labeling) task. On the other hand, for question answering task, you can directly use the question answering formulation since Transformers already has a support for that task.

Modeling the Input

You need to decide on how to represent the input including the common special tokens such as BOS, EOS. This formulation is directly used while creating the input_ids value of the input instances. As a concrete example, you can represent a sequence classification input with BOS ... actual_input_tokens ... EOS format. Moreover, some tasks require extra task-specific special tokens as well. For example, in conditional text generation, you may need to prompt the generation with a special signaling token. In tasks that utilizes multiple sequences, you may need to use segment embeddings (via token_type_ids) to label the tokens according to their sequence.

Class Reference

trapper_components

The above diagram shows the basic components in Trapper. To use trapper on training, evaluation on a task that is not readily supported in Transformers, you need to extend the provided base classes according to your own needs. These are as follows:

For Training & Evaluation: LabelMapper, DataProcessor, DataAdapter, TokenizerWrapper, MetricInputHandler, MetricOutputHandler.

For Inference: In addition to the ones listed above, you may need to implement a transformers.Pipeline or directly use one from Transformers if they already implemented one that matches your need.

Typically Extended Classes

  1. LabelMapper: Used in tasks that require mapping between categorical labels and integer ids such as token classification.

  2. DataProcessor: This class is responsible for taking a single instance in dict format, typically coming from a datasets.Dataset, extracting the information fields suitable for the task and hand, and converting their values to integers or collections of integers. This includes, tokenizing the string fields, and getting the token ids, converting the categoric labels to integer ids and so on.

  3. DataAdapter: This is responsible for converting the information fields inside an instance dict that was previously processed by a DataProcessor to a format suitable for feeding into a transformer model. This also includes handling the special tokens signaling the start or end of a sequence, the separation of tho sequence for a sequence-pair task as well as chopping excess tokens etc.

  4. TokenizerWrapper: This class wraps a pretrained tokenizer from the Transformers while also recording the special tokens needed for the task to the wrapped tokenizer. It also stores the missing values from BOS - CLS, EOS - SEP token pairs for the tokenizers that only support one of them. This means you can model your input sequence by using the bos_token for start and eos_token for end without thinking which model you are working with. If your task and input modeling needs extra special tokens e.g. the <CONTEXT> for a context dependent task, you can store these tokens by setting the _TASK_SPECIFIC_SPECIAL_TOKENS class variable in your TokenizerWrapper subclass. Otherwise, you can directly use TokenizerWrapper.

  5. MetricInputHandler: This class is mainly responsible for preprocessing applied to predictions and labels (references). This is performed for transforming the predictions and labels to a suitable format to be fed in metrics for computation. For example, while using BLEU in a language generation task, the predictions and labels need to be converted to a string or list of strings. However, for extractive question answering task in which the predictions are returned as start and end indices pointing the answer within the context, additional information (e.g context in such case) may be needed, so directly returning the start and end indices in this case does not help, and additional operation is needed to be done by converting predictions to actual answers extracted from the context. You are able to do this kind of operations through MetricInputHandler, storing additional information, converting predictions and labels to a suitable format, manipulating resulting score. Furthermore, in child classes helper classes can also be implemented (e.g TokenizerWrapper, LabelMapper) for required tasks. In this class, we provide three main functionality:

    • _extract_metadata(): This method allows user to extract metadata from dataset instances to be later used for preprocessing predictions and labels in preprocess() method.
    • __call__(): This method allows converting predictions and labels into a suitable form for metric computation. The default behavior is defined as directly returning predictions and labels without manipulation, but only applying argmax() to predictions to convert the model predictions to predictions input for metrics.
  6. MetricOutputHandler: The intention of this class is to support for manipulating the score object returned by the metric computation phase. Jury returns a well-constructed dictionary output for all metrics; however, to shorten dictionary items, manipulate the information within the output or to add additional information to score dictionary, this class can be extended as desired.

  7. transformers.Pipeline: The pipeline mechanism from Transformers have not been fully integrated yet. For now, you should check Transformers to find a pipeline that is suitable for your needs and does the same pre-processing. If you could not find one, you may need to write your own Pipeline by extending transformers.Pipeline or one of its subclasses and add it to transformers.pipelines.SUPPORTED_TASKS map. To enable instantiation of the pipelines from the checkpoint folders, we provide a factory, create_pipeline_from_checkpoint function. It accepts a checkpoint directory of a completed experiment, the path to the config file (already saved in that directory), as well as the task name that you used while adding the pipeline to SUPPORTED_TASKS. It re-creates the objects you used while training such as model wrapper, label mapper etc and provides them as keyword arguments to constructor of the pipeline you implemented.

Registering classes from custom modules to the library

We support both file based and command line argument based approaches to register the external modules written by the users.

Option 1 - File based

You should list the packages or modules (for stand-alone modules not residing inside any package) containing the classes to be registered as plugins to a local file named .trapper_plugins. This file must reside in the same directory where you run the trapper run command. Moreover, it is recommended to put the plugins file where the modules to be registered resides (e.g. the project root) for convenience since that directory will be added to the PYTHONPATH. Otherwise, you need to add the plugin module/packages to the PYTHONPATH manually. Another reminder is that each listed package must have an __init__.py file that imports the modules containing the custom classes to be registered.

E.g., let's say our project's root directory is project_root and the experiment config file is inside the root with a name test_experiment.jsonnet. To run the experiment, you should run the following commands:

cd project_root
trapper run test_experiment.jsonnet

Below output shows the content of the project_root directory.

ls project_root

  ner
  tests
  datasets
  .trapper_plugins
  test_experiment.jsonnet

Additionally, here is the content of the project_root/.trapper_plugins.

cat project_root/.trapper_plugins

  ner.core.models
  ner.data.dataset_readers
Option 2 - Using the command line argument

You can specify the packages and/or modules you want to be registered using the --include-package argument. However, note that you need to repeat the argument for each package/module to be registered.

E.g. the running the following commands is an alternative to Option-1 to start the experiment specified in the test_experiment.jsonnet.

trapper run test_experiment.jsonnet \
--include-package ner.core.models \
--include-package ner.data.dataset_readers

Running a training and/or evaluation experiment

Config File Based Training Using the CLI

Go to your project root and execute the trapper run command with a config file specifying the details of the training experiment. E.g.

trapper run SOME_DIRECTORY/experiment.jsonnet

Don't forget to provide the args["output_dir"] and args["result_dir"] values in your experiment file. Please look at the examples/pos_tagging/README.md for a detailed example.

Script Based Training

Go to your project root and execute the trapper run command with a config file specifying the details of the training experiment. E.g.

trapper run SOME_DIRECTORY/experiment.jsonnet

Don't forget to provide the args["output_dir"] and args["result_dir"] values in your experiment file. Please look at the examples/pos_tagging/README.md for a detailed example.

Examples for Using Trapper as a Library

We created an examples folder that includes example projects to help you get started using trapper. Currently, it includes a POS tagging project using the CONLL2003 dataset, and a question answering project using the SQuAD dataset. The POS tagging example shows how to use trapper on a task that does not have a direct support from Transformers. It implements all custom components and provides a complete project structure including the tests. On the other hand, the question answering example shows using trapper on a task that Transformers already supported. We implemented it to demonstrate how trapper may still be helpful thanks to configuration file based experiments.

Training a POS Tagging Model on CONLL2003

Since transformers lacks a direct support for POS tagging, we added an example project that trains a transformer model on CONLL2003 POS tagging dataset and perform inference using it. It is a self-contained project including its own requirements file, therefore you can copy the folder into another directory to use as a template for your own project. Please follow its README.md to get started.

Training a Question Answering Model on SQuAD Dataset

You can use the notebook in the Example QA Project examples/question_answering/question_answering.ipynb to follow the steps while training a transformer model on SQuAD v1.

Currently Supported Tasks and Models From Transformers

Hypothetically, nearly all models should work on any task if it has an entry in the table of AutoModelFor... factories for that task. However, since some models require more (or less) parameters compared to most of the models in the library, you might get errors while using such models. We try to cover these edge cases them by adding the extra parameters they require. Feel free to open an issue/PR if you encounter/solve such issues in a model-task combination. We have used trapper on a limited set of model-task combinations so far. We list these combinations below to indicate that they have been tested and validated to work without problems.

Table of Model-task Combinations Tested so far

model question_answering token_classification
BERT
ALBERT
DistillBERT
ELECTRA
RoBERTa

Installation

Environment Creation

It is strongly recommended creating a virtual environment using conda or virtualenv etc. before installing this package and its dependencies. For example, the following code creates a conda environment with name trapper and python version 3.7.10, and activates it.

conda create --name trapper python=3.7.10
conda activate trapper

Regular Installation

You can install trapper and its dependencies by pip as follows.

pip install trapper

Contributing

If you would like to open a PR, please create a fresh environment as described before, clone the repo locally and install trapper in editable mode as follows.

git clone https://github.com/obss/trapper.git
cd trapper
pip install -e .[dev]

After your changes, please ensure that the tests are still passing, and do not forget to apply code style formatting.

Testing trapper

Caching the test fixtures from the HuggingFace datasets library

To speed up the data-related tests, we cache the test dataset fixtures from HuggingFace's datasets library using the following command.

python -m scripts.cache_hf_datasets_fixtures

Then, you can simply run the tests with the following command:

python -m scripts.run_tests

NOTE: To significantly speed up the tests depending on HuggingFace's Transformers and datasets libraries, you can set the following environment variables to make them work in offline mode. However, beware that you may need to run the tests once first without setting these environment variables so that the pretrained models, tokenizers etc. are downloaded and cached.

export TRANSFORMERS_OFFLINE=1 HF_DATASETS_OFFLINE=1

Code Style

To check code style,

python -m scripts.run_code_style check

To format codebase,

python -m scripts.run_code_style format

Contributors