SMASHED is a toolkit designed to apply transformations to samples in datasets, such as fields extraction, tokenization, prompting, batching, and more. Supports datasets from Huggingface, torchdata iterables, or simple lists of dictionaries.
Project description
Sequential MAppers for Sequences of HEterogeneous Dictionaries is a set of Python interfaces designed to apply transformations to samples in datasets, which are often implemented as sequences of dictionaries. To start, run
pip install smashed
Example of Usage
Mappers are initialized and then applied sequentially. In the following example, we create a mapper that is applied to a samples, each containing a sequence of strings. The mappers are responsible for the following operations.
- Tokenize each sequence, cropping it to a maximum length if necessary.
- Stride sequences together to a maximum length or number of samples.
- Add padding symbols to sequences and attention masks.
- Concatenate all sequences from a stride into a single sequence.
import transformers
from smashed.mappers import (
TokenizerMapper,
MultiSequenceStriderMapper,
TokensSequencesPaddingMapper,
AttentionMaskSequencePaddingMapper,
SequencesConcatenateMapper,
)
tokenizer = transformers.AutoTokenizer.from_pretrained(
pretrained_model_name_or_path='bert-base-uncased',
)
mappers = [
TokenizerMapper(
input_field='sentences',
tokenizer=tokenizer,
add_special_tokens=False,
truncation=True,
max_length=80
),
MultiSequenceStriderMapper(
max_stride_count=2,
max_length=512,
tokenizer=tokenizer,
length_reference_field='input_ids'
),
TokensSequencesPaddingMapper(
tokenizer=tokenizer,
input_field='input_ids'
),
AttentionMaskSequencePaddingMapper(
tokenizer=tokenizer,
input_field='attention_mask'
),
SequencesConcatenateMapper()
]
dataset = [
{
'sentences': [
'This is a sentence.',
'This is another sentence.',
'Together, they make a paragraph.',
]
},
{
'sentences': [
'This sentence belongs to another sample',
'Overall, the dataset is made of multiple samples.',
'Each sample is made of multiple sentences.',
'Samples might have a different number of sentences.',
'And that is the story!',
]
}
]
for mapper in mappers:
dataset = mapper.map(dataset)
print(len(dataset))
# >>> 5
print(dataset[0])
# >>> {
# 'input_ids': [
# 101,
# 2023,
# 2003,
# 1037,
# 6251,
# 1012,
# 102,
# 2023,
# 2003,
# 2178,
# 6251,
# 1012,
# 102
# ],
# 'attention_mask': [
# 1,
# 1,
# 1,
# 1,
# 1,
# 1,
# 1,
# 1,
# 1,
# 1,
# 1,
# 1,
# 1
# ]
# }
Building a Pipeline
Mappers can also be composed into a pipeline using the >> (or <<) operator. For example, the code above can be rewritten as follows:
pipeline = TokenizerMapper(
input_field='sentences',
tokenizer=tokenizer,
add_special_tokens=False,
truncation=True,
max_length=80
) >> MultiSequenceStriderMapper(
max_stride_count=2,
max_length=512,
tokenizer=tokenizer,
length_reference_field='input_ids'
) >> TokensSequencesPaddingMapper(
tokenizer=tokenizer,
input_field='input_ids'
) >> AttentionMaskSequencePaddingMapper(
tokenizer=tokenizer,
input_field='attention_mask'
) >> SequencesConcatenateMapper()
dataset = ...
# apply the full pipeline to the dataset
pipeline.map(dataset)
Dataset Interfaces Available
The initial version of SMASHED supports two interfaces for dataset:
interfaces.simple.Dataset: A simple dataset representation that is just a list of python dictionaries with some extra convenience methods to make it work with SMASHED. You can crate a simple dataset by passing a list of dictionaries tointerfaces.simple.Dataset.- HuggingFace
datasetslibrary. SMASHED mappers work with any datasets from HuggingFace, whether it is a regular or iterable dataset.
Developing SMASHED
To contribute to SMASHED, make sure to:
- (If you are not part of AI2) Fork the repository on GitHub.
- Clone it locally.
- Create a new branch in for the new feature.
- Install development dependencies with
pip install -r dev-requirements.txt. - Add your new mapper or feature.
- Add unit tests.
- Run tests, linting, and type checking from the root directory of the repo:
- Style:
black .(Should format for you) - Style:
flake8 .(Should return no error) - Style:
isort .(Should sort imports for you) - Static type check:
mypy .(Should return no error) - Tests:
pytest -v --color=yes tests/(Should return no error)
- Style:
- Commit, push, and create a pull request.
- Tag
soldnito review the PR.
A note about versioning
SMASHED follows Semantic Versioning. In short, this means that the version number is MAJOR.MINOR.PATCH, where:
- MAJOR version when you make incompatible API changes,
- MINOR version when you add functionality in a backwards compatible manner; adding a mapper typically falls under this category, and
- PATCH version when you make backwards compatible bug fixes.
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
