octopipes 0.2.3

Pipeline library for AI workflows.

Octopipes

Octopipes is a pipeline library for AI workflows. Not only it allows for easy definition of multi-step pipelines, but also handles testing and collection of information of multiple workflows.

pip install octopipes

Introduction

When using multiple-step (AI) pipelines, octopipes helps you define workflows and work with workflows in an easy way. It allows for instance, adding post-workflow hooks that can clean up GPU memory (or anything else).

Workflows are defined through the Workflow class, each workflow can contain a chain of processes (some common ones are already defined). Every step of the pipeline is saved and can be used later. For easier managements of different outputs, we can provide an output handler for each specific step (Some basic ones are already defined as well such BboxesHandler etc.).

To run multiple workflows on the same input, you can use the AggregateFlows class for that. The library also provides a way to read datasets and run benchmarks

To keep octopipes ML library agnostic, it does not require pytorch or tensorflow to be installed as it can work with both just fine.

Get started

Workflows

To add steps to a workflow, the class provides the add method which takes as input the function (process) and optionally an OutputHandler. The whole workflow will act as the | (pipe) operator in Unix terminals by successively feeding the output of a process as the input of the next process. To run a workflow, you just need to iterate over it with the appropriate input(s).

p1 -> p2 -> p3

In math terms, the workflow runs the function composition of all the steps.

p3 \circ p2 \circ p1 (x)

This shows how to define and run a simple workflow:

from octopipes.workflow import Workflow

# Define a workflow with two steps
wf = Workflow('wf_name').add(lambda x: x ** 2).add(lambda x: x - 4)
print(wf.nsteps)
# output: 1

# We can now run the workflow on a specific starting input value (in this case 4)
# `wf` will first run the input on the first function x ** 2, then run the second x - 4 with the result of the previous step.
# So in the first step of the iteration the result will be 16 (4 ** 2) then 12 (16 - 4)
wf_iter = wf(4)
for step, result in wf_iter:
    pass

# return a frozen instance of the workflow run.
# This is especially important, when working with memory intensive GPU workflows
frozen_res = wf_iter.freeze()

# Get the duration recap of each step of the workflow
wf_iter.recap()

# Define a workflow with some metadata attached
# this metadata can then be used to differentiate
# workflows with the same name but different params
wf = Workflow('wf_name', metadata={'thresh': 0.4}).add(lambda x: x ** 2)

Output handlers

When adding a new step that outputs a certain results that you want to be processed in a particular way, you can pass a class that implements the OutputHandler interface.

The interface has 3 methods:

• output_on_image: used when outputting the results on an image (Used in computer vision mostly)
• len_output: give the output size of the result
• to_json: returns a serialized json object of the result

The library already provides some basic ones such as:

• BboxesHandler
• SegmentationMaskHandler
• CmapBboxesHandler
• CirclesHandler

Handlers are added this way. If None were supplied, DefaultHandler is used. (In most cases a handler needs to be passed)

wf = Workflow('wf_name').add(some_func, some_handler)

Workflow steps with additional inputs

The add method can take an optional requires flag in cases a step needs additional inputs to work properly. This flag is comma separated string that declares the requires inputs needed for a specific process. These requirements are then injected as separate arguments to the step (unpacked). Currently, the requires flag supposed two types of injections:

• Injections from previous steps (or initial input) -- These are declared by their order number (0, 1, etc)
• Dependency injections that are independent of the flow of the pipeline. -- These are prefix with the letter d (d0,d2, etc.)

It is also possible to mix the requires with both types such as 0,d1,2 or d0,1. And the order in the requires flags is retained when injecting the arguments to the function.

Previous input injection

In cases where a workflow step requires inputs from previous steps other than the one strictly before it, we can inject whose inputs using the requires flag. The flag is a series of comma separated step numbers like (0, 0,1,2) where 0 is the first input of the workflow and 1 the output of the first step.

To understand this flag, let's suppose that we have such a requirement:

(input) -> p1 -> p2 -> p3

Now let's say p3 not only requires the output of p2 but also the first input (The initial starting value when running the workflow). In this case when defining your workflow you can specify a requires flag requires='0' that will inject input as the second argument to your process.

.add(lambda p2, input: some_calculation(input, p2), requires='0')

Note that the order of the steps in the requires flag is retained when injecting the arguments, and the previous step of the workflow is ignored in case it is present in requires as it is automatically injected by default.

Dependency injections

In some cases you might need a workflow to use some data or dependency further down in the workflow without having to propagate that variable through all your steps. This is where you can use the requires flag of the add method to tell octopipes to inject dependencies as arguments to a specific step. The way to specify them is to prefix the dependency number with the letter d. So d0,d2 would inject the dependency number 0 and 2 to your step.

After declaring where the dependencies should be injected, you simply has to specify them when running the workflow:

wf = Workflow('name').add(lambda x y: x + y, requires='d0')
wf_iter = wf(input=get_input(), dependencies=[dependency0, dependency1])

AggregateFlows

AggregateFlows allows running multiple workflows on the same input. This is usually used when either benchmarking multiple pipelines at the same time or wanting to select the "best" output out of different workflows.

from octopipes.workflow import Workflow
from octopipes.aggreage_flows import AggregateFlows

wf1 = Workflow('wf_name').add(some_func, some_handler)
wf2 = Workflow('wf_name').add(some_func, some_handler)

flows = AggregateFlows(input, workflows=[wf1, wf2])
flows.run_workflows()

# get results for wf1
flows.results[0]

Benchmark

As the name suggest, Benchmark allows testing your workflows on a dataset and then being able to calculate easily your metrics. Depending on the batch size of the dataset loader, the tests will be run simultaneously (as many processes as the batch size). Take note however that as of now, a single AggregateFlow is run synchronously.

If you're using pytorch or tensorflow, some memory freeing hooks might be needed. For that, you can pass an instance of DefaultAggregateFlowsFactory with the specific hooks needed. Otherwise, you can always define implement your own AggregateFlowsFactory.

Here's a simple example:

from octopipes.workflow import Workflow
from octopipes.benchmark import Benchmark

dataloader = ...

wf1 = Workflow('wf1_name').add(some_func, some_handler)
wf2 = Workflow('wf2_name').add(some_func, some_handler)

bench = Benchmark(dataloader=dataloader, workflows=[wf1, wf2])
bench.run_tests()

Requirements & Installation

The module is tested against versions >=3.10. However, this requirement is due to using type hinting so the module can be altered to work on lower version of the interpreter.

Installing the package is pretty standard:

git clone https://github.com/octomiro/octopipes

pip install -r requirements.txt
# or requirements.in

pip install -r dev-requirements.in

# Running tests
pytest
# or
tox

Contributions

PRs are more than welcome! If the change is big enough to require some discussion, it's better to open an issue for it. To keep the history of the repo clean, all PRs are rebased instead of merged so make sure everything is correct be submitting anything.

Why Octopipes?

This library was born from a need to define, benchmark, and debug in an easy way workflows that use foundational models. In the course of our work, we did not find something that met that need in terms of flexibility or features. So we created octopipes internally, and decided to open source it.

Octopipes is developed and maintained by octomiro, an AI company that makes ERP systems intelligent.