SeqTools 1.1.0

A library for transparent transformation of indexable containers (lists, etc.)

SeqTools

SeqTools facilitates the manipulation of datasets and the evaluation of a transformation pipeline. Some of the provided functionalities include: mapping element-wise operations, reordering, reindexing, concatenation, joining, slicing, minibatching, etc.

To improve ease of use, SeqTools manipulates list-like objects, otherwise known as a sequences (objects with a length supporting integer or slice based indexing).

Manipulating a dataset as a whole can be slow and resource/memory intensive. To circumvent this issue, SeqTools implements on-demand evaluation under the hood: operations and transformations on a dataset are only applied to individual items when they are actually accessed. This is particularly convenient for prototyping.

When comes the transition from prototyping to execution, the list-like container interface facilitates serial evaluation. Besides, SeqTools also provides simple helpers to dispatch work between multiple workers (threads or processes).

SeqTools originally targets data science, more precisely the data preprocessing stages. Being aware of the experimental nature of this usage, on-demand execution is made as transparent as possible to users by providing fault-tolerant functions and insightful error reporting. Moreover, internal code is kept concise and clear with comments to facilitate error tracing through a failing transformation pipeline.

Example

>>> def f1(x):
... return x + 1
...
>>> def f2(x): # slow and memory heavy transformation
... time.sleep(.01)
... return [x for _ in range(500)]
...
>>> def f3(x):
... return sum(x) / len(x)
...
>>> data = list(range(1000))

Without delayed evaluation, defining the pipeline and reading values looks like so:

>>> tmp1 = [f1(x) for x in data]
>>> tmp2 = [f2(x) for x in tmp1] # takes 10 seconds and a lot of memory
>>> res = [f3(x) for x in tmp2]
>>> print(res[2])
3.0
>>> print(max(tmp2[2])) # requires to store 499 500 useless values along
3

With seqtools:

>>> tmp1 = seqtools.smap(f1, data)
>>> tmp2 = seqtools.smap(f2, tmp1)
>>> res = seqtools.smap(f3, tmp2) # no computations so far
>>> print(res[2]) # takes 0.01 seconds
3.0
>>> print(max(tmp2[2])) # easy access to intermediate results
3

Batteries included!

The library comes with a set of functions to manipulate sequences:

concatenation	batching	reindexing
prefetching	interleaving

and others (suggestions are also welcome).

Installation

pip install seqtools

Documentation

The documentation is hosted at https://seqtools-doc.readthedocs.io.

Contributing and Support

Use the issue tracker to request features, propose improvements or report issues. For questions regarding usage, please send an email.

Related libraries

Joblib, proposes low-level functions with many optimization settings to optimize pipelined transformations. This library notably provides advanced caching mechanisms which are not the primary concern of SeqTool. SeqTool uses a simpler container-oriented interface with multiple utility functions in order to assist fast prototyping. On-demand evaluation is its default behaviour and applies at all layers of a transformation pipeline. In particular, parallel evaluation can be inserted in the middle of the transformation pipeline and won’t block the execution to wait for the computation of all elements from the dataset.

SeqTools is conceived to connect nicely to the data loading pipeline of Machine Learning libraries such as PyTorch’s torch.utils.data and torchvision.transforms or Tensorflow’s tf.data. The interface of these libraries focuses on iterators to access transformed elements, contrary to SeqTools which also provides arbitrary reads via indexing.