smile-datasets 0.0.6

La**S**t **mile** datasets: Use `tf.data` to solve the last mile data loading problem for tensorflow.

smile-datasets

LaSt mile Datasets: Use tf.data to solve the last mile data loading problem for tensorflow.

If you want to load public datasets, try:

• tensorflow/datasets
• huggingface/datasets

If you want to load local, personal dataset with minimized boilerplate, use Smile Dataset!

Support Matrix

task	supported	core abstractions
question answering	[x]	ExampleForQuestionAnswering, DatasetForQuestionAnswering, DatapipeForQuestionAnswering
masked language model	[x]	ExampleForMaskedLanguageModel, DatasetForMaskedLanguageModel, DatapipeForMaskedLanguageModel
sequence classification	[x]	ExampleForSequenceClassification, DatasetForSequenceClassification, DatapipeForSequenceClassification
token classification	[x]	ExampleForTokenClassification, DatasetForTokenClassification, DatapipeForTokenClassification
unsupervised simcse	[x]	ExampleForUnsupervisedSimCSE, DatasetForUnsupervisedSimCSE, DatapipeForUnsupervisedSimCSE
supervised simcse	[x]	ExampleForSupervisedSimCSE, DatasetForSupervisedSimCSE, DatapipeForSupervisedSimCSE
hard negative simcse	[x]	ExampleForHardNegativeSimCSE, DatasetForHardNegativeSimCSE, DatapipeForHardNegativeSimCSE

Usage

All datapipes for different tasks has the same interface.

Here is an example for question answering task, but you can use datapipe the same way for other tasks.

Example for Question Answering

from smile_datasets import DatasetForQuestionAnswering, DatapipeForQuestionAnswering

# each line is a JSON {"sequece": "我喜欢自然语言处理(NLP)"}
train_input_jsonl_files = ["data/train.jsonl"]
train_dataset = DatapipeForQuestionAnswering.from_jsonl_files(
    input_files=train_input_jsonl_files, 
    vocab_file="bert/vocab.txt",
    batch_size=32,
)

# check dataset
print(next(iter(train_dataset)))

# model = build_keras_model(...)
# model.compile(...)
# train model
model.fit(train_dataset, callbacks=[...])

For maximum flexibility, you can always subclass DatasetForQuestionAnswering to load your dataset, just like torch.utils.data.Dataset:

from smile_datasets import DatasetForQuestionAnswering, DatapipeForQuestionAnswering, ParserForQuestionAnswering

class DuReaderDatasetForQuestionAnswering(DatasetForQuestionAnswering):
    """Dataset reader for DuReader dataset."""

    def __init__(self, input_files, vocab_file, subset="rubost", **kwargs) -> None:
        super().__init__()
        self.parser = ParserForQuestionAnswering(tokenizer=None, vocab_file=vocab_file, **kwargs)
        if subset == "rubost":
            self.instances = list(readers.read_dureader_rubost(input_files, **kwargs))
        else:
            self.instances = list(readers.read_dureader_checklist(input_files, **kwargs))
        self.examples = []
        for instance in self.instances:
            e = self.parser.parse(instance)
            if not e:
                continue
            self.examples.append(e)

    def __len__(self):
        return len(self.examples)

    def __getitem__(self, index) -> ExampleForQuestionAnswering:
        return self.examples[index]


dataset = DuReaderDatasetForQuestionAnswering(input_files=["data/trian.jsonl"], vocab_file="bert/vocab.txt")
train_dataset = DatapipeForQuestionAnswering.from_dataset(dataset, batch_size=32)

# check dataset
print(next(iter(train_dataset)))

# model = build_keras_model(...)
# model.compile(...)
# train model
model.fit(train_dataset, callbacks=[...])

For better performance, you can convert dataset to tfrecord ahead of time, and then build datapipe from tfrecord files directly:

# save dataset in tfrecord format
dataset.save_tfrecord(output_files="data/train.tfrecord")

# build datapipe from tfrecord files
train_dataset = DatapipeForQuestionAnswering.from_tfrecord_files(input_files="data/train.tfrecord", batch_size=32)

# check dataset
print(next(iter(train_dataset)))

# model = build_keras_model(...)
# model.compile(...)
# train model
model.fit(train_dataset, callbacks=[...])