🍋 pipelime

If life gives you lemons, use pipelime.

Welcome to pipelime, a swiss army knife for data processing!

pipelime is a full-fledge framework for data science: read your datasets, manipulate them, write back to disk or upload to a remote data lake. Then build up your dataflow with Piper and manage the configuration with Choixe. Finally, embed your custom commands into the pipelime workspace, to act both as dataflow nodes and advanced command line interface.

Maybe too much for you? No worries, pipelime is modular and you can just take out what you need:

• data processing scripts: use the powerful SamplesSequence and create your own data processing pipelines, with a simple and intuitive API. Parallelization works out-of-the-box and, moreover, you can easily serialize your pipelines to yaml/json. Integrations with popular frameworks, e.g., pytorch, are also provided.
• easy dataflow: Piper can manage and execute directed acyclic graphs (DAGs), giving back feedback on the progress through sockets or custom callbacks.
• configuration management: Choixe is a simple and intuitive mini scripting language designed to ease the creation of configuration files with the help of variables, symbol importing, for loops, switch statements, parameter sweeps and more.
• command line interface: pipelime can remove all the boilerplate code needed to create a beautiful CLI for you scripts and packages. You focus on what matters and we provide input parsing, advanced interfaces for complex arguments, automatic help generation, configuration management. Also, any PipelimeCommand can be used as a node in a dataflow for free!
• pydantic tools: most of the classes in pipelime derive from pydantic.BaseModel, so we have built some useful tools to, e.g., inspect their structure, auto-generate human-friendly documentation and more (including a TUI to help you writing input data to deserialize any pydantic model).

---

Installation

Install pipelime using pip:

pip install pipelime-python

To be able to draw the dataflow graphs, you need the draw variant:

pip install pipelime-python[draw]

Warning

The draw variant needs Graphviz https://www.graphviz.org/ installed on your system On Linux Ubuntu/Debian, you can install it with:

sudo apt-get install graphviz graphviz-dev

Alternatively you can use conda

conda install --channel conda-forge pygraphviz

Please see the full options at https://github.com/pygraphviz/pygraphviz/blob/main/INSTALL.txt

Basic Usage

Underfolder Format

The Underfolder format is the preferred pipelime dataset formats, i.e., a flexible way to model and store a generic dataset through filesystem.

An Underfolder dataset is a collection of samples. A sample is a collection of items. An item is a unitary block of data, i.e., a multi-channel image, a python object, a dictionary and more. Any valid underfolder dataset must contain a subfolder named data with samples and items. Also, global shared items can be stored in the root folder.

Items are named using the following naming convention:

Where:

• $ID is the sample index, must be a unique integer for each sample.
• ITEM is the item name.
• EXT is the item extension.

We currently support many common file formats and others can be added by users:

• .png, .jpeg/.jpg/.jfif/.jpe, .bmp for images
• .tiff/.tif for multi-page images and multi-dimensional numpy arrays
• .yaml/.yml, .json and .toml/.tml for metadata
• .txt for numpy 2D matrix notation
• .npy for general numpy arrays
• .pkl/.pickle for picklable python objects
• .bin for generic binary data

Root files follow the same convention but they lack the sample identifier part, i.e., $ITEM.$EXT

Reading an Underfolder Dataset

pipelime provides an intuitive interface to read, manipulate and write Underfolder Datasets. No complex signatures, weird object iterators, or boilerplate code, you just need a SamplesSequence:

    from pipelime.sequences import SamplesSequence

    # Read an underfolder dataset with a single line of code
    dataset = SamplesSequence.from_underfolder('tests/sample_data/datasets/underfolder_minimnist')

    # A dataset behaves like a Sequence
    print(len(dataset))             # the number of samples
    sample = dataset[4]             # get the fifth sample

    # A sample is a mapping
    print(len(sample))              # the number of items
    print(set(sample.keys()))       # the items' keys

    # An item is an object wrapping the actual data
    image_item = sample["image"]    # get the "image" item from the sample
    print(type(image_item))         # <class 'pipelime.items.image_item.PngImageItem'>
    image = image_item()            # actually loads the data from disk (may have been on the cloud as well)
    print(type(image))              # <class 'numpy.ndarray'>

Writing an Underfolder Dataset

You can write a dataset by calling the associated operation:

    # Attach a "write" operation to the dataset
    dataset = dataset.to_underfolder('/tmp/my_output_dataset')

    # Now run over all the samples
    dataset.run()

    # You can easily spawn multiple processes if needed
    dataset.run(num_workers=4)