ts-preprocessing 0.1.2

A preprocessing library for foundation time series models

Time Series Preprocessing Library

A comprehensive library for building streaming data pipelines for time series datasets, providing tools for downloading, transforming, and combining time series data in real-time.

Features

• Streaming Dataset Transforms: Build composable data pipelines that process time series data as it streams
• Dataset Downloaders: Download datasets from Hugging Face Hub with caching support
• Synthetic Data Generation: Generate synthetic time series data for testing and development
• Flexible Data Pipeline Builder: Chain multiple transforms together using a fluent builder pattern
• PyTorch Integration: Full compatibility with PyTorch's IterableDataset interface
• Type Safety: Comprehensive type annotations for better development experience

Installation

Using poetry:

poetry install

Using pip:

pip install .

Core Components

Dataset Transforms

The library provides a rich set of streaming dataset transforms that can be chained together:

Basic Transforms

• TransformingDataset: Apply arbitrary functions to each item in a dataset
• BatchingIterableDataset: Group items into batches with configurable batch sizes
• UnbatchingIterableDataset: Flatten batched datasets back to individual items
• SlidingWindowIterableDataset: Create sliding windows over time series data

Combining Transforms

• ConcatDataset: Concatenate multiple datasets sequentially
• CombiningDataset: Combine multiple datasets using custom operations (e.g., element-wise addition)
• ProbabilisticMixingDataset: Mix datasets with configurable probabilities and seeding

Pipeline Builder

• Builder: Fluent interface for building complex data pipelines

Dataset Downloaders

• HuggingFaceDownloader: Download datasets from Hugging Face Hub with automatic caching
• GiftEvalWrapperDataset: Wrapper for Salesforce's GiftEval pretraining datasets

Synthetic Data

• LinearTrendDataset: Generate synthetic time series with linear trends and configurable noise

Serialization

• serialize_tensor_stream: Save tensor streams to disk in sharded format
• SerializedTensorDataset: Load serialized tensor streams with lazy loading support

Usage Examples

Basic Transform Pipeline

from preprocessing.transform.dataset_builder import Builder
from preprocessing.transform.batching_dataset import BatchingIterableDataset
from preprocessing.transform.transforming_dataset import TransformingDataset

# Create a simple pipeline
pipeline = (
    Builder(your_dataset)
    .batch(batch_size=32)
    .map(lambda x: x * 2)  # Double all values
    .build()
)

# Iterate over the pipeline
for batch in pipeline:
    print(batch.shape)  # (32, sequence_length, features)

Combining Multiple Datasets

from preprocessing.transform.combining_dataset import CombiningDataset

# Combine two datasets element-wise
def add_operation(x, y):
    return x + y

combined = CombiningDataset([dataset1, dataset2], op=add_operation)

for result in combined:
    print(result)  # x + y for each pair

Concatenating Datasets

from preprocessing.transform.concat_dataset import ConcatDataset

# Concatenate datasets sequentially
concatenated = ConcatDataset([dataset1, dataset2, dataset3])

for item in concatenated:
    # Items from dataset1, then dataset2, then dataset3
    print(item)

Probabilistic Mixing

from preprocessing.transform.probabilistic_mixing_dataset import ProbabilisticMixingDataset

# Mix datasets with custom probabilities
mixed = ProbabilisticMixingDataset(
    datasets={"train": train_data, "val": val_data},
    probabilities={"train": 0.8, "val": 0.2},
    seed=42  # For reproducibility
)

for item in mixed:
    # 80% chance from train, 20% chance from val
    print(item)

Sliding Windows

from preprocessing.transform.sliding_window_dataset import SlidingWindowIterableDataset

# Create sliding windows over time series
windowed = SlidingWindowIterableDataset(
    dataset=your_dataset,
    window_size=100,
    step=50
)

for window in windowed:
    print(window.shape)  # (100, features)

Downloading Datasets

from preprocessing.downloader.huggingface import HuggingFaceDownloader
from preprocessing.config import DatasetConfig

# Configure dataset download
config = DatasetConfig(
    name="air-passengers",
    repo_id="duol/airpassengers",
    files=["AP.csv"],
    cache_dir="data/cache"
)

# Download dataset
downloader = HuggingFaceDownloader(config)
data = downloader.download()

Synthetic Data Generation

from preprocessing.synthetic.linear_trend import LinearTrendDataset

# Generate synthetic time series
synthetic_data = LinearTrendDataset(
    sequence_length=1000,
    num_sequences=100,
    trend_slope=0.01,
    noise_std=0.1,
    seed=42
)

for sequence in synthetic_data:
    print(sequence.shape)  # (1000, 1)

Serialization

from preprocessing.serialization.serialize import serialize_tensor_stream
from preprocessing.serialization.deserialize import SerializedTensorDataset

# Save dataset to disk
serialize_tensor_stream(
    dataset=your_dataset,
    output_dir="data/serialized",
    max_tensors_per_file=1000
)

# Load dataset from disk
loaded_dataset = SerializedTensorDataset(
    filepaths=["data/serialized/shard_00000.pt"],
    lazy=True  # Load on-demand
)

Project Structure

preprocessing/
├── common/              # Common types and utilities
│   └── tensor_dataset.py
├── config/              # Configuration management
│   ├── __init__.py
│   └── examples/
├── downloader/          # Dataset downloaders
│   ├── huggingface.py
│   └── gift_eval.py
├── serialization/       # Data serialization
│   ├── serialize.py
│   └── deserialize.py
├── synthetic/           # Synthetic data generation
│   └── linear_trend.py
└── transform/           # Dataset transforms
    ├── batching_dataset.py
    ├── combining_dataset.py
    ├── concat_dataset.py
    ├── dataset_builder.py
    ├── probabilistic_mixing_dataset.py
    ├── sliding_window_dataset.py
    ├── transforming_dataset.py
    └── unbatching_dataset.py

Development

Setup Development Environment

poetry install --with dev

Run Tests

poetry run pytest

Run Tests with Coverage

poetry run pytest --cov=preprocessing

Key Design Principles

1. Streaming First: All transforms work with streaming data, enabling processing of large datasets that don't fit in memory
2. Composability: Transforms can be easily chained together using the Builder pattern
3. Type Safety: Comprehensive type annotations for better development experience
4. PyTorch Integration: Full compatibility with PyTorch's data loading ecosystem
5. Reproducibility: Built-in support for random seeds and deterministic operations
6. Flexibility: Support for custom operations and transformations

License

MIT License - see LICENSE file for details