Task-based datasets, preprocessing, and evaluation for sequence models
Go to SeqIO ReadTheDocs Documentation Page.
SeqIO is a library for processing sequential data to be fed into downstream
sequence models. It uses
tf.data.Dataset
to create scalable data pipelines but requires minimal use of TensorFlow. In
particular, with one line of code, the returned dataset can be transformed to a
numpy iterator and hence it is fully compatible with other frameworks such as
JAX or
PyTorch.
SeqIO assumes that the dataset is a sequence. Modalities such as text or audio are naturally supported. Images are supported as long as they are represented as sequences (e.g., Image GPT).
SeqIO is a refactor of the
t5.data
library used (in conjunction with the
Mesh Tensorflow Transformer
implementation) to train the T5 models introduced in Exploring the Limits of
Transfer Learning with a Unified Text-to-Text
Transformer.
If you have used t5.data in the past and want to know how SeqIO differs,
please read this section.
pip install seqiogit clone https://github.com/google/seqio.git
cd seqio
pip install -e .At a high level, we use SeqIO with the following steps.
-
Define a
Task(and optionally aMixture). -
Define (or use an existing) a
FeatureConverterbased on the model architecture. -
Use the top-level function
seqio.get_datasetto obtain thetf.data.Datasetinstance.
We will look at each of these steps in detail.
The most important class in SeqIO is the Task. It is an abstraction that
combines:
- a raw data source
- one or more preprocessing steps
- a vocabulary to tokenize/detokenize each preprocessed feature for the model
- a postprocessor to convert detokenized model outputs into a format for evaluation
- one or more metrics to evaluate with
Oftentimes a Task lines up with a common benchmark. In this tutorial, we use
WMT 19 English-German
machine translation task. In the end, our Task will look like this:
seqio.TaskRegistry.add(
"wmt19_ende",
seqio.TfdsDataSource(tfds_name="wmt19_translate/de-en:1.0.0"),
preprocessors=[
functools.partial(
translate, source_language='en', target_language='de'),
seqio.preprocessors.tokenize, seqio.preprocessors.append_eos
],
output_features={
'inputs':
seqio.Feature(
seqio.SentencePieceVocabulary('/path/to/inputs/vocab'),
add_eos=False,
dtype=tf.int32),
'targets':
seqio.Feature(
seqio.SentencePieceVocabulary('/path/to/targets/vocab'),
add_eos=True,
dtype=tf.int32),
},
metric_fns=[bleu])We typically add the Task to the global registry when we define it (as shown
above) to make it easier to use with model configs and flags. Thus, it must
have a unique string name ("wmt19_ende" in this case). Note, however, that
you may also instantiate a seqio.Task directly without adding it to the
registry, if desired.
We'll now break down each part of the task definition.
Data sources are the first step in your pipeline, providing a way to load raw
data in many formats as a tf.data.Dataset.
All data sources are subclasses of the DataSource base class and are defined
in
dataset_providers.
Existing implementations include:
TfdsDataSourcefor loading examples from TensorFlow Datasets.TextLineDataSourcefor loading examples from text files (e.g., tsv).TFExampleDataSourcefor loadingtf.train.Exampleprotos from a file (e.g. aTFRecordfile.)FunctionDataSourcefor providing an custom function that returns atf.data.Dataset.
In our example, we are using the TfdsDataSource. We specify the name of the
WMT dataset in TFDS
("wmt19_translate"),
the specific config for the language pair that excludes the context for the open
domain setting ("de-en"), and the version number ("1.0.0").
The output_features field expects a dictionary that maps string feature names
to seqio.Feature objects. This defines what the Task is expected to produce
in its output examples. The output examples may contain additional fields, but
they must contain these fields in the specified format or exceptions will be
raised.
Each Feature includes:
- A
vocabulary, which must subclassseqio.Vocabulary, to specify how the feature can be tokenized and detokenized. You may useseqio.PassThroughVocabularyif tokenization is not necessary. add_eos, which specifies whether the feature should end with the vocabulary's EOS token.- The output
dtypewhich must be atf.dtypes.DType.
Note: specifying these options on Feature does not by itself ensure the
proper transformations are applied -- you must also include the necessary
preprocessors.
The tasks used in T5 all produce "inputs" and "targets" features to be
consumed by the text-to-text model. For a decoder-only language model, only a
single feature (e.g., "targets") would be necessary. Nevertheless, SeqIO is
flexible enough to generate arbitrary output features what will be converted
into model features by the FeatureConverter later in the
pipeline.
Preprocessors are functions that transform one tf.data.Dataset into a new
tf.data.Dataset. Typically this involves executing a map over the given
dataset. The preprocessors provided to the Task will be executed sequentially.
As an example, let's look at the previously undefined translate from the
"wmt19_ende" example above.
def translate(dataset: tf.data.Dataset,
source_language: str,
target_language: str) -> tf.data.Dataset:
def _translate(ex: Mapping[str, tf.Tensor]) -> Mapping[str, tf.Tensor]:
"""Convert a translation example to a text2text pair.
For example, say the dataset returns examples of this format:
{'de': 'Das ist gut.', 'en': 'That is good.'}
If source_language = 'de', target_language = 'en', then the outputs will
have the format:
{'inputs': 'translate de to en: Das ist gut.',
'targets': 'That is good.'}
Args:
ex: an example to process.
source_language: source language code (e.g. 'en') to translate from.
target_language: target language code (e.g. 'de') to translate to.
Returns:
A preprocessed example with the format listed above.
"""
src_str = f'translate {source_language}'
tgt_str = f' to {target_language}: '
return {
'inputs': tf.strings.join([src_str, tgt_str, ex[source_language]]),
'targets': ex[target_language],
}
return dataset.map(_translate,
num_parallel_calls=tf.data.experimental.AUTOTUNE)The TFDS dataset provides the dataset where each example has the form: {'de': 'Das ist gut.', 'en': 'That is good.'}. We convert this to "inputs" and
"targets" with the appropriate prompt to inform the model of the task.
A few important notes:
-
When instantiating a
Task, the preprocessor functions can have the following arguments:dataset,output_features, andsequence_length. The first (positional) dataset argument is always required. If an argument namedoutput_featuresis provided, the output feature mapping will be passed to the preprocessor. Ifsequence_lengthis provided, a mapping from feature name to its maximum final sequence length (provided by the caller) will be passed -- any sequences that are too long after preprocessing will be automatically truncated. If a preprocessor function does have other arguments, they must have default values or be bound (e.g., withfunctools.partialas used intranslate) before instantiating theTask. -
Mapping functions operate on and return
tf.Tensors using TensorFlow operations. This is more flexible than it may sound:- Automatic AutoGraph conversion allow you to write python control flow in your transformations.
- tf.experimental.numpy provides a numpy interface.
tf.py_functionallows you to wrap arbitrary Python code. Note:tf.datapipelines using this function can only be run in the python process where they were defined, and performance is limited by the python GIL.
See
tf.data.Datasetdocumentation for more details. -
When calling
map, it is important to always setnum_parallel_calls=tf.data.experimental.AUTOTUNEto avoid creating a bottleneck. Theseqio.map_over_datasetdecorator helps enforce this as follows.@seqio.map_over_dataset def translate(ex: Mapping[str, tf.Tensor], source_language: str, target_language: str) -> Mapping[str, tf.Tensor]: """Convert a translation dataset to a text2text pair. For example, say the dataset returns examples of this format: {'de': 'Das ist gut.', 'en': 'That is good.'} If source_language = 'de', target_language = 'en', then the outputs will have the format: {'inputs': 'translate German to English: Das ist gut.', 'targets': 'That is good.'} Args: ex: an example to process. source_language: source language code (e.g. 'en') to translate from. target_language: target language code (e.g. 'de') to translate to. Returns: A preprocessed example with the format listed above. """ src_str = f'translate {source_language}' tgt_str = f' to {target_language}: ' return { 'inputs': tf.strings.join([src_str, tgt_str, ex[source_language]]), 'targets': ex[target_language], }
Note that
translatetakes as input an individual example. Thenseqio.map_over_datasetdecorates it to a function that takes in atf.data.Datasetinstance. -
Stochastic operations must be stateless if deterministic pipelines are needed. To get (optionally deterministic) seeds for these operations, use the
seqio.map_over_dataset(num_seeds=n)decorator. For example:def random_chunk( dataset: tf.data.Dataset, sequence_length: Mapping[str, int] ) -> tf.data.Dataset: """Takes a random chunk out of each feature with size `sequence_length`.""" @seqio.map_over_dataset(num_seeds=1) def take_chunk( ex: Mapping[str, tf.Tensor], seed ) -> Mapping[str, tf.Tensor]: new_ex = {} for k, v in ex.items(): if k in sequence_length: length = sequence_length[k] start_idx = tf.random.stateless_uniform( (), seed, 0, tf.size(v) - (length + 1)) new_ex[k] = v[start_idx:start_idx+length] else: new_ex[k] = v return new_ex return take_chunk(dataset)
If
num_seeds > 1, the arg will instead be calledseedsand will contain a sequence of seeds.
In our "wmt_19_ende" task, we also use the predefined preprocessors
seqio.preprocessors.tokenize and seqio.preprocessors.append_eos. The former
uses each Feature.vocabulary to tokenize it, and the the latter appends
Feature.vocabulary.eos_id to the feature if the Feature.add_eos is True. See
preprocessors.py for
their implementations and other useful preprocessors.
During evaluation, the model outputs are first detokenized using the output
feature vocabulary. Before passing these predictions to the metric functions,
they can be run through a Python postprocessing function, alongside the full
input example. Similarly, the raw targets are run through this function before
being passed to the metrics. Since the postprocess function is used on both the
model output and the targets, it is passed an is_target boolean in case the
behavior should be different. It is also passed the fully preprocessed example,
including fields that were excluded from output_features.
For the "wmt19_ende", we don't need any postprocessors. See "trivia_qa_open"
task in the Advanced Postprocessing Task for
an example postprocessor.
Metrics are functions that are passed (by the Evaluator) the
fully-materialized list of postprocessed model outputs (or scores) and targets
and return a mapping from string names to MetricValue objects containing their
values. These are most commonly floating-point scalars, but may also be text,
images, audio, histograms, etc (see
metrics.py for the full list).
The first argument of a metric function must always be called targets. If the
second argument of a metric function is called predictions, it will be passed
the decoded and detokenized model prediction. If it is called scores, it will
be passed a list of log-likelihood scores for each example.
If multiple metric functions are provided, they will all be used and their returned mappings merged.
Prediction metrics are computed using the postprocessed targets and model
outputs (predictions). The args must be named targets and predictions.
Let's look at the metric function used for "wmt19_ende" task. A standard metric
for the translation task is BLEU and we use sacrebleu implementation.
def bleu(targets: Sequence[str], predictions: Sequence[str]):
"""Computes BLEU score.
Args:
targets: list of strings or list of list of strings if multiple references
are present.
predictions: list of strings
Returns:
bleu_score across all targets and predictions
"""
if isinstance(targets[0], list):
targets = [[x for x in target] for target in targets]
else:
# Need to wrap targets in another list for corpus_bleu.
targets = [targets]
bleu_score = sacrebleu.corpus_bleu(predictions, targets,
smooth_method="exp",
smooth_value=0.0,
force=False,
lowercase=False,
tokenize="intl",
use_effective_order=False)
return {"bleu": bleu_score.score}Score metrics are computed using the postprocessed targets and their
log-likelihood scores according to the model. The args must be named targets
and scores.
def perplexity(targets: Sequence[str], scores: Sequence[float]):
return {
"perplexity": seqio.metrics.Scalar(np.exp(np.mean(scores)))
}Once you have multiple Tasks added to the TaskRegistry, you can define
Mixtures that will combine the examples from them according to some specified
rate. Examples will then be sampled from each task in proportion to its rate.
As an example, Multilingual T5 uses a Mixture of
per-language Tasks with tail languages up-weighted in the mixture.
There are 3 ways to specify the tasks and their rates:
-
Provide a rate along with each task's name (rates are normalized before sampling). In this example, the rates provided are units of the final mixture that come from the component tasks. Here, 1/(1+7) of the final mixture will come from "task1".
seqio.MixtureRegistry.add( "mix1", [("task1", 1), ("task2", 7)] )
-
Provide a constant default rate for some or all tasks, which will be used when only the name is provided. The example below will produce identical mixing rates as the previous one.
seqio.MixtureRegistry.add( "mix1", [("task1", 0.5), "task2"], default_rate=3.5 )
-
Provide a function that generates the rate for each task at runtime. The example below uses the provided
seqio.mixing_rate_num_examples, which uses the number of examples (computed during offline caching) as the rate for each task.seqio.MixtureRegistry.add( "mix2", ["task1", "task2"], default_rate=seqio.mixing_rate_num_examples )
You can also include Mixtures in your Mixture! For example, the following
task would contain 1/24 (from "mix1") + 1/3 "task1", 7/24 (from "mix1") of
"task2", and 1/3 "task3".
seqio.MixtureRegistry.add(
"mix3",
["mix1", "task1", "task3"],
default_rate=1
)If sampling without replacement is important for your task, you can achieve that by using either deterministic tasks or using dataset checkpointing (and not running more than an epoch) for a non-deterministic task. Otherwise, the mixture may sample with replacement.
Now that your Task (and/or Mixture) is defined, its primary functionality is
to use it to generate a dataset.
You may first need to use seqio.get_mixture_or_task(mixture_or_task_name) to
access your dataset provider from the registry.
After that, you can call get_dataset to build the tf.data.Dataset. For
example:
dataset = seqio.get_mixture_or_task("mix1").get_dataset(
sequence_length={"inputs": 256, "targets": 128},
split="train",
shuffle=True,
num_epochs=1,
shard_info=seqio.ShardInfo(index=0, num_shards=10),
use_cached=False,
seed=42
)
# Print the first 5 examples.
for _, ex in zip(range(5), dataset.as_numpy_iterator()):
print(ex)Some notes on a few of the arguments:
sequence_length: An optional mapping from feature name to maximum length. Will be passed to the preprocessors with asequence_lengthargument. If notNone, the final example features will be truncated if they exceed the specified length. Note that this value may be required to be set if any of the preprocessors use thesequence_lengthargument and do not handle theNonecase.num_epochs: The number of times to repeat the source dataset. Preprocessing will be re-applied with new seeds to enable new samples from stochastic steps. Note that if theCacheDatasetPlaceholderis included (see below) preprocessing is only re-applied after that step.shard_info: An optional sharding specification for loading a deterministic subset of the dataset. Loading will be most efficient if the number of shards evenly divides the number of shards in the raw data source.use_cached: Specifies whether to load from a pre-cached task for increased performance or to do the preprocessing on-the-fly. See the following section for details on how to cache your task, which must be done before this can be set toTrue.seed: An optional seed to use for deterministic shuffling and (stateless) stochastic ops. These operations will still be pseudorandom but will be reproducible with the same seed. Set toNoneif determinism is not desired.
For improved performance at load time and to avoid redundant computations for
commonly used tasks, you can pre-cache your Task with all or part of the
preprocessing done in advance of training; this partial preprocessing is
especially useful if the Task is stochastic and one wishes to cache the
deterministic operations while running the stochastic ones on the fly. Caching
stochastic SeqIO Mixtures in this way is not supported.
The first step to doing so is to add a
seqio.CacheDatasetPlaceholder(required=False) as one of the steps in your
preprocessing pipeline. All steps before the placeholder will be cached offline
and all steps after will be executed on the fly at load time. You may set
required=True if you want get_dataset to fail unless use_cached=True.
Caveats:
- Any stochastic operations that you wish to be re-run when
num_epochs > 1or with a differentseedshould go after the placeholder since only a single sample will be cached. - Any preprocessing steps that use the
sequence_lengthargument must come after theseqio.CacheDatasetPlaceholderpreprocessor since this is only known at runtime, or an exception will be raised. If you wish to cache for a specific sequence length, you can useseqio.experimental.add_fully_cached_task.
Once your Task is registered, you can run
cache_tasks_main
to execute the offline preprocessing, providing it with the module containing
your task definitions via the --module_import flag. For very large datasets,
it's recommended you run this Apache Beam script on
a distributed framework like
Google Cloud DataFlow.
Finally, you are ready to load the cached version of your Task (or Mixture)
containing it. You will need to add the path to the directory you passed to
--output_cache_dir via seqio.add_global_cache_dirs(["/my/cache/dir"]). Now
when you call task_or_mixture.get_dataset(..., use_cached=True), the data will
be loaded from the cache directory instead of the raw data source.
The role of Task is to provide the dataset object with as little
model-specific features (e.g., generic "inputs" and "targets") while the Feature
Converters transform the model-agnostic features to model-specific features
(e.g., "encoder_input_tokens"). We refer to the former as "task features" and
the latter as "model features".
Let's use machine translation (English to German) as a running example.
The raw data consists of sentence pairs such as
"That is good\tDas ist gut."
A task registered to Task (e.g.,
wmt_t2t_ende_v003)
reads these sentence pairs from the data source and applies a series of
preprocessors.
One of the internal representations looks like
{"inputs": "translate English to German: That is good.",
"targets": "Das ist gut."}The final output from the Task is a tokenized version of the parallel
sentences. In the following toy example (the token ids do not correspond to the
above string example), the dataset consists of 2 examples.
dataset = [{"inputs": [7, 8, 5], "targets": [3, 9]},
{"inputs": [8, 4, 9, 3], "targets": [4]}]The format is in the tf.data.Dataset (i.e., each example is a dictionary with
"inputs" and "targets" fields.
The FeatureConverter then takes this as an input and converts to the
model-specific features. In addition, the feature converter performs padding and
optionally packing (for model implementations that support it) for efficiency.
For example, let's assume that we are using the standard Transformer
architecture with an encoder and a decoder. The output of the feature converter
is
converted_dataset = [{
"encoder_input_tokens": [7, 8, 5, 1, 8, 4, 9, 3, 1, 0],
"encoder_segment_ids": [1, 1, 1, 1, 2, 2, 2, 2, 2, 0],
"encoder_positions": [0, 1, 2, 3, 0, 1, 2, 3, 4, 0],
"decoder_target_tokens": [3, 9, 1, 4, 1, 0, 0],
"decoder_input_tokens": [0, 3, 9, 0, 4, 0, 0],
"decoder_loss_weights": [1, 1, 1, 1, 1, 0, 0],
"decoder_positions": [0, 1, 2, 0, 1, 0, 0],
"decoder_segment_ids": [1, 1, 1, 2, 2, 0, 0],
}]In this case, two task examples are packed into one. *_segment_id and
*_position are the fields used to denote the membership and position of packed
token in the original sequence. The EOS ids (i.e., 1) are appended. In addition,
each fields is padded to the specified length.
We will look at the details of this example in Encoder-decoder architecture:
seqio.EncDecFeatureConverter section.
We provide feature converters for three common architectures: encoder-decoder, decoder-only and encoder-only. Here we describe how users can use the feature converters for each of these architectures out of the box as a part of the SeqIO library.
In the SeqIO library, each architecture has a class defining how the task
features are converted to model features. Since these feature converters are
already implemented, it is straightforward to use them by providing the class as
a feature_converter argument of the seqio.get_dataset function. The
following sections show example usage of seqio.get_dataset.
This is the architecture of the original Transformer paper. For the
English-to-German translation task, the following function call retrieves the
tf.data.Dataset object with the model features.
dataset: tf.data.Dataset = seqio.get_dataset(
mixture_or_task_name="wmt_t2t_ende_v003",
task_feature_lengths={"inputs": 32, "targets": 32},
dataset_split="train",
shuffle=True,
feature_converter=seqio.EncDecFeatureConverter(pack=True)
)The resulting dataset object has the following 7 fields
| Feature name | Explanation |
|---|---|
encoder_input_tokens |
Input tokens to the encoder. |
encoder_positions |
Position index in the sequence before packing. |
encoder_segment_ids |
Sequence membership before packing. Two positions with |
| the same positive integer mean that they belong to the same sequence before | |
| packing. | |
decoder_input_tokens |
Input tokens to the decoder. |
decoder_target_tokens |
Output tokens from the decoder. |
decoder_loss_weights |
A weight on each position that can be used as a mask. |
decoder_positions |
Position index in the sequence before packing. |
decoder_segment_ids |
Same as encoder_segment_ids but for decoder. |
This architecture consists of a single autoregressive stack, which we denote as a "decoder".
A decoder autoregressively produces an output sequence. Therefore, it can be used as a standard language model if the task dataset has only "targets" features, i.e., self-supervised. If the task dataset also has an "inputs" field, e.g., supervised machine translation, the decoder can still be used by concatenating the inputs and targets fields. See Raffel et al. (2020), Section 3.2.1 for more detailed take on this topic.
We support both uses cases and refer to the former as standard language model and the latter as prefix language model. Each of these models is described separately below.
Note that we do not provide special features to denote how the dataset should be consumed. For example, a Transformer-based fully autoregressive decoder has a fully-causal self-attention layer. Since there are many ways of implementing the masking pattern for such attention layer and, more importantly, SeqIO is not limited to attention-based models, we leave it up to the model implementations to apply the masking pattern. There is one exception, and we cover this in the Prefix LM section below.
A common use pattern is to pretrain a decoder model with the left-to-right
language modeling objective (unsupervised) using seqio.LMFeatureConverter and
then fine-tune (supervised) using seqio.PrefixLMFeatureConverter.
For the standard language model, the task dataset only has "targets" field. Therefore, the sequence length specification only needs to specify targets.
dataset: tf.data.Dataset = seqio.get_dataset(
mixture_or_task_name="standard_lm",
task_feature_lengths={"targets": 32},
dataset_split="train",
shuffle=True,
feature_converter=seqio.LMFeatureConverter(pack=True)
)Note that "standard_lm" is not a registered task in the codebase. It is the left-to-right language modeling task, i.e., predict the next token given the previous tokens on some language corpus (e.g., C4).
The output dataset has the following model features.
| Feature name | Explanation |
|---|---|
decoder_target_tokens |
Output tokens from the decoder |
decoder_input_tokens |
Input tokens to the decoder |
decoder_loss_weights |
Binary mask to indicate where the loss should be taken |
decoder_positions |
Position index in the sequence before packing |
decoder_segment_ids |
Sequence membership before packing. Two positions with |
| the same positive integer mean that they belong to the same sequence before | |
| packing. |
The decoder_target_tokens is a shifted version of decoder_input_tokens for
the standard teacher-forced autoregressive training.
If the input dataset has a notion of "inputs" and "targets", we can concatenate them so that we can still use a single stack decoder. Therefore, the output only contains "targets" just like standard LM case.
We use the same toy example for English-to-German translation task as a running example:
{"inputs": "translate English to German: That is good.",
"targets": "Das ist gut."}
To be consumed by the decoder-only stack, seqio.PrefixLMFeatureConverter
concatenates them form the new "targets". Consider 2-layer decoder architecture
whose activations are shown below
That is good <EOS> Das ist gut <EOS>
| | | | | | | |
u1 u2 u3 u4 u5 u6 u7 u8
| | | | | | | |
v1 v2 v3 v4 v5 v6 v7 v8
| | | | | | | |
<BOS> That is good <EOS> Das ist gut
Let us denote the first layer's activation in the ith position as vi.
Similarly, let ui denote the activation of the second layer in the ith
position.
For attention-based sequence models such as Transformer decoders, the self-attention layer is used to encode contextualized representation of the sequence. At a given layer, each position's representation is computed as a function of the representations of the tokens before its position in the previous layer.
Referring to the toy example, when computing u2 with fully-causal masking, we
do not use v3. This results in a representation u2 of the word "is" that
does not take into account the word "good", which is unnecessarily limiting.
For Prefix LM, this issue is resolved by having the fully visible masking
pattern for the inputs portion only. For example, when computing u2, v1,
v2, v3, v4 and v5 are all visible and taken into account. For the tokens
in the "targets" of the Task dataset, we use the causal masking. For example,
when computing u6, all vi for i <= 6 are taken into account but not v7.
Why is `v5` included in the inputs attention pattern?
In the same translation example, we note that when computing `u2`, the activation corresponding to the position where \ token was input (i.e., `v5`) was visible. This doesn't count as "cheating" because the model doesn't see the next word "Das". This can provide additional context in building the representation for "good". In this case, `u4` has the context that "good" is the last word in the sentence.seqio.PrefixLMFeatureConverter provides a feature decoder_causal_attention
to encode this information. For the above example, we have
decoder_causal_attention = [1, 1, 1, 1, 1, 0, 0, 0]
indicating that the non-causal attention can be applied to the first five positions. Note that this feature seems trivial, but for a packed dataset the inputs and targets boundary are more nuanced.
A final consideration for the prefix LM is that because we concatenate "inputs"
and "targets", which tokens are used for the loss computation is a modeling
decision. For example, we can penalize the models only for the "targets" tokens
or we may choose to penalize building the representation for "inputs" tokens.
This is controlled by loss_on_targets_only argument (defaults to True) to
seqio.PrefixLMFeatureConverter constructor. In the above example, we would get
decoder_loss_weights = [0, 0, 0, 0, 1, 1, 1, 1]
This indicates that the last 4 positions are used for the loss computation.
To get the dataset with prefix LM features, we can use
dataset: tf.data.Dataset = seqio.get_dataset(
mixture_or_task_name="wmt_t2t_ende_v003",
task_feature_lengths={"inputs": 32, "targets": 32},
dataset_split="train",
shuffle=True,
feature_converter=seqio.PrefixLMFeatureConverter(
pack=True,
loss_on_targets_only=True)
)The resulting features have length 64 because it concatenates inputs and targets each with length 32.
The output dataset has the following model features. Note that the only
additional feature is decoder_causal_attention.
| Feature name | Explanation |
|---|---|
decoder_target_tokens |
Output tokens from the decoder |
decoder_input_tokens |
Input tokens to the decoder |
decoder_loss_weights |
Binary mask to indicate where the loss should be |
| taken | |
decoder_positions |
Position index in the sequence before packing |
decoder_segment_ids |
Sequence membership before packing. Two positions with |
| the ` same positive integer mean that they belong to the same sequence before | |
| packing. | |
decoder_causal_attention |
Binary mask denoting which tokens are in the |
| non-causal masking region. |
Like decoder-only architecture, this one is a single stack, but not autoregressive.
One notable assumption is that the inputs and targets are aligned, i.e., they
have the same sequence length and ith position in the targets correspond to
the output representation of the ith token in the inputs.
A common model using encoder-only architecture is
BERT. We provide Encoder feature converter
class to support the Masked Language Modeling (MLM) objective from BERT.
We assume that a unique sentinel such as [MASK] token is used to mask some
fraction of the input text and the task is to recover the original text.
Therefore, the "targets" is naturally defined as the original text whereas
"inputs" are the masked text.
Encoder-only models are often used for classification tasks. In BERT, a special
token [CLS] is prepended to the input sequence. The last layer's activation
corresponding to this sentinel token is the contextualized representation of the
sequence. We assume that such "classification" sentinel is prepended.
Consider the following example for the MLM task. The input dataset has two
examples, which is packed to one example. We assume that mask_id = 9 and the
[CLS] token has id of 8.
dataset = [{"inputs": [8, 9, 9, 3, 4], "targets": [8, 7, 4, 3, 4]},
{"inputs": [8, 3, 9], "targets": [8, 3, 6]}]
converted_dataset = {
"encoder_input_tokens": [8, 9, 9, 3, 4, 1, 8, 3, 9, 1, 0],
"encoder_target_tokens": [8, 7, 4, 3, 4, 1, 8, 3, 6, 1, 0],
"encoder_segment_ids": [1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 0],
"encoder_positions": [0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 0],
"encoder_loss_weights": [0, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0],
}Note that the packed sequence has [CLS] token at the beginning of each
sequences. Also note that the loss is taken only on the masked position.
To use the pre-defined EncoderFeatureConverter, provide mask_id as an
argument.
dataset: tf.data.Dataset = seqio.get_dataset(
mixture_or_task_name="some mlm task",
task_feature_lengths={"inputs": 32, "targets": 32},
dataset_split="train",
shuffle=True,
feature_converter=seqio.EncoderFeatureConverter(
pack=True,
mask_id=9)
)The resulting dataset object has the following 5 fields
| Feature name | Explanation |
|---|---|
encoder_input_tokens |
Input tokens to the encoder |
encoder_positions |
Position index in the sequence before packing |
encoder_segment_ids |
Sequence membership before packing. Two positions with |
| the ` same positive integer mean that they belong to the same sequence before | |
| packing. | |
encoder_target_tokens |
Output tokens from the encoder |
encoder_loss_weights |
Binary mask to indicate where the loss should be taken |
For a custom model architecture, you need to create a subclass of
FeatureConverter and override two methods _convert_features and
get_model_feature_lengths to define how task features are mapped to the model
features, including the length relationships. The existing feature converters
(e.g., seqio.EncDecFeatureConverter) follow the same pattern, which can be a
useful starting point.
The SeqIO Evaluator class provides a way to evaluate models on SeqIO Tasks
and Mixtures. For an interactive walkthrough of SeqIO evaluation, see the
Evaluation Notebook.
The following is a deep-dive into the Evaluator class.
An Evaluator instance can be created by passing a SeqIO Task or
Mixture, and additional eval params like feature converter, split, sequence
lengths, seed, etc. The Evaluator init calls get_dataset for each Task to be
evaluated with the appropriate params, creating the task_dataset, and invokes
the model-specific feature converter on the task_dataset to create features
that can be passed to a model, called model_dataset. Both task_dataset and
model_dataset are stored in-memory so that the dataset can be reused across
multiple evaluations (e.g. on checkpoints from a training run). Both datasets
are enumerated so that even if the order of examples is changed during model
inference, the enumeration can be used to match model outputs to examples from
the task_dataset.
For Mixtures, each sub-Task is evaluated separately, regardless of mixing rates, because in the context of eval benchmarks, Mixtures commonly refer to a collection of Tasks belonging to that benchmark, each of which is evaluated separately, e.g. SuperGLUE mixture.
Once an Evaluator instance is created with a SeqIO Task or Mixture, a model
can be evaluated by calling evaluator.evaluate(...) and passing a predict_fn
and/or a predict_with_aux_fn and/or a score_fn to interact with the model.
predict_fn takes the model_dataset as input and outputs a Sequence[(index, token_ids)] where token_ids is the sequence of token ids generated by the
model for the input example whose index matches index. Therefore, even if
predict_fn mixes the order of the examples during prediction, the order can be
corrected as long as the correct index for each example is maintained. A common
example is the multi-host setup where the evaluation dataset is split amongst
multiple hosts that independently make predictions and combine the results
during which the ordering can be mixed. predict_with_aux_fn is similar to
predict_fn, except that it can also return a dictionary of auxiliary values
along with each sequence of token_ids, e.g. scores from the generated tokens.
The score_fn takes the model_dataset as input and returns a
Sequence[(index, score)] where score is the sequence of log likelihood
scores for the targets in the dataset. This simple interface allows users to
easily integrate the SeqIO evaluation flow with popular training frameworks in
TF and Jax.
Corresponding to the model fns, users can configure three kinds of metric fns in
their Tasks, which are differentiated by their function signature. Metrics
computed on the outputs of predict_fn (and predict_with_aux_fn) have the
signature targets and predictions (and optionally aux_values), while
metrics computed on the outputs of score_fn have the signature targets and
scores. The Evaluator takes care of calling the correct model fns and
metric fns during evaluation. Here is an example of a metric of each type.
def sequence_accuracy(targets, predictions):
seq_acc = 100 * np.mean([p == t for p, t in zip(predictions, targets)])
return {"sequence_accuracy": seq_acc}
def log_likelihood(targets, scores):
log_likelihood = np.mean([scipy.special.logsumexp(el) for el in scores])
return {"log_likelihood": log_likelihood}
There are 4 steps involved in the evaluation using predicted tokens:
- the
predict_fnorpredict_with_aux_fnreturns indices and output_tokens:Sequence[Tuple[int, Sequence[int]]], potentially with some auxiliary values. - output tokens are decoded by
vocab.decode - postprocessors configured in Tasks are applied to the decoded output. These are denoted as predictions.
- metric fns configured in Tasks are applied to the predictions and the cached targets.
There are 2 steps involved in the evaluation using scores:
- the
score_fnreturns indices and scores:Sequence[Tuple[int, Sequence[float]]] - metric fns configured in Tasks is applied to the scores and the cached targets.
Training codebases like T5X provide integration with SeqIO evaluation to allow evaluating checkpoints on SeqIO Tasks and Mixtures. See T5X Eval for instructions.
The original t5 library introduced and implemented the t5.data.Task
abstraction for specifying preprocessing and evaluation metrics for text-to-text
tasks. When creating a task, users specify a source dataset of raw text, some
preprocessing steps, a vocabulary for tokenization, and evaluation metrics. The
fully-specified Task can then be used to pre-train or fine-tune a
encoder-decoder transformer model. However, the design included many baked-in
assumptions about the types of tasks users could specify.
SeqIO removes some of the constraints of this abstraction:
- Inputs and outputs are no longer required to be strings (e.g., it may be images or audio).
- Architectures other than the original encoder-decoder are supported (e.g., decoder-only language models like GPT or encoder-only models like BERT).
- Users can control at which stage of the pipeline offline caching occurs.
- Users can control when and where EOS tokens are added.
Furthermore, SeqIO has been made more modular with respect to the Mesh TensorFlow Transformer. This allows it to be used with other model implementations with more consistency and much less code duplication.
This version of TriviaQA was introduced in Roberts et al. 2020.
seqio.TaskRegistry.add(
"trivia_qa_open",
source=seqio.TfdsDataSource(
tfds_name="trivia_qa/unfiltered.nocontext:1.1.0",
splits={
"train": "train[:90%]",
"validation": "train[90%:]",
"test": "validation"
}),
preprocessors=[
tqa_open_preprocessor,
seqio.preprocessors.tokenize,
seqio.preprocessors.append_eos,
],
output_features={
"inputs": seqio.Feature(
seqio.SentencePieceVocabulary("/path/to/inputs/vocab"),
add_eos=False, dtype=tf.int32
),
"targets": seqio.Feature(
seqio.SentencePieceVocabulary("/path/to/targets/vocab"),
add_eos=True, dtype=tf.int32
),
},
postprocess_fn=tqa_open_postprocessor,
metric_fns=[tqa_metric])In this example, we are using the TfdsDataSource. We specify the name of the
TriviaQA dataset in TFDS
("trivia_qa"), the
specific config that excludes the context for the open domain setting
("unfiltered.nocontext"), and the version number ("1.1.0"). We also override
the default splits to match what is commonly used for the open domain setting.
Specifically, we set our "test" split to be the TFDS "validation" split, and
create a small pseudo-"validation" set by taking examples out of the TFDS
"train" split.
The preprocessor tqa_open_preprocessor is defined as follows.
def tqa_open_preprocessor(
dataset: tf.data.Dataset,
prefix:str = "trivia_qa question: "
) -> tf.data.Dataset:
"""Convert TriviaQA dataset to open domain qa examples.
The function takes the trivia_qa TFDS dataset and emits examples of the
form:
{
"inputs": "trivia_qa question: What are the names of the Olsen Twins?"
"targets": "Mary-Kate and Ashley",
"answers": ["Mary-Kate and Ashley", "Ashley and Mary-Kate"]
}
Args:
dataset: a tf.data.Dataset to process.
prefix: str, prefix to prepend to the inputs.
Returns:
a tf.data.Dataset
"""
def tqa_map(ex):
"""Map TriviaQA example to text-to-text example."""
return {
"inputs": prefix + ex["question"],
"targets": ex["answer"]["value"],
"answers": ex["answer"]["aliases"],
}
return dataset.map(tqa_map, num_parallel_calls=tf.data.experimental.AUTOTUNE)Or with the seqio.map_overdataset decorator, we have
def tqa_open_preprocessor(
dataset: tf.data.Dataset,
prefix: str = "trivia_qa question: "
) -> tf.data.Dataset:
@seqio.map_over_dataset
def tqa_map(ex: Mapping[str, tf.Tensor]) -> Mapping[str, tf.Tensor]:
"""Map TriviaQA example to text-to-text example."""
return {
"inputs": prefix + ex["question"],
"targets": ex["answer"]["value"],
"answers": ex["answer"]["aliases"],
}
return tqa_map(dataset)Here we made a thin wrapper to emphasize that the function decorated by
seqio.map_over_dataset takes in an instance of tf.data.Dataset. In practice,
this wrapper is not necessary.
The postprocessor for this example is tqa_open_postprocessor, which is defined
as follows:
def tqa_open_postprocessor(output_or_target, example=None, is_target=False):
"""Returns output as answer, or all answers if the full example is provided."""
if is_target:
return [a.decode("utf-8") for a in example["answers"]]
else:
return output_or_target.decode("utf-8")When processing the target, we ignore output_or_target (equivalent to
example["targets"]) since it is just selecting a single answer in
trivia_qa_open. Instead, we extract the full list of answers from the example
and convert them from bytes to text. When handling the model output, we simply
convert it to text from detokenized bytes.
The metric function tqa_metric is defined as:
def tqa_metric(
targets: Sequence[Sequence[str]],
predictions: Sequence[str]
) -> Mapping[str, seqio.metrics.MetricValueValue]:
"""Computes official TriviaQA metrics.
Args:
targets: list of lists of strings
predictions: list of strings
Returns:
dict with score_key: squad score across all targets and predictions
"""
if len(targets) != len(predictions):
raise ValueError("Number of targets and predictions must match.")
def _normalize_answer(text):
"""Lower text and remove punctuation, articles and extra whitespace."""
# Remove articles.
text = re.sub(r"\b(a|an|the)\b", " ", s)
# Remove punctuation.
for punc in string.punctuation:
text = text.replace(punc, '')
# Normalize white space
text = " ".join(s.split())
return text
# Normalize answers before comparing.
targets = [[_normalize_answer(t) for t in u] for u in targets]
predictions = [_normalize_answer(p) for p in predictions]
em = np.mean([
max(pred == gt for gt in ground_truths)
for pred, ground_truths in zip(predictions, targets)
])
return {
"exact_match": seqio.metrics.Scalar(em),
}Please use the following bibtex entry to cite SeqIO.
@article{roberts2022t5x,
url = {https://arxiv.org/abs/2203.17189},
author = {Roberts, Adam and Chung, Hyung Won and Levskaya, Anselm and Mishra,
Gaurav and Bradbury, James and Andor, Daniel and Narang, Sharan and Lester,
Brian and Gaffney, Colin and Mohiuddin, Afroz and Hawthorne, Curtis and
Lewkowycz, Aitor and Salcianu, Alex and van Zee, Marc and Austin, Jacob and
Goodman, Sebastian and Soares, Livio Baldini and Hu, Haitang and
Tsvyashchenko, Sasha and Chowdhery, Aakanksha and Bastings, Jasmijn and
Bulian, Jannis and Garcia, Xavier and Ni, Jianmo and Chen, Andrew and Kenealy,
Kathleen and Clark, Jonathan H. and Lee, Stephan and Garrette, Dan and
Lee-Thorp, James and Raffel, Colin and Shazeer, Noam and Ritter, Marvin and
Bosma, Maarten and Passos, Alexandre and Maitin-Shepard, Jeremy and Fiedel,
Noah and Omernick, Mark and Saeta, Brennan and Sepassi, Ryan and Spiridonov,
Alexander and Newlan, Joshua and Gesmundo, Andrea},
title = {Scaling Up Models and Data with $\texttt{t5x}$ and $\texttt{seqio}$},
journal={arXiv preprint arXiv:2203.17189},
year = {2022},
}