rats-processors 0.1.3.dev21

Rats Processors

rats-processors

A package to create and compose pipelines in a high level API, where processors (classes or unbound methods) are mapped into pipeline nodes, node ports are inferred from the processors signature, and edges are created by connecting node ports inputs and outputs. Pipelines defined this way are immutable objects that can be reused and composed into larger pipelines, facilitating reusability.

Example

In your python project or Jupyter notebook, you can compose a pipeline as follows:

from typing import NamedTuple

from pathlib import Path

from sklearn.base import BaseEstimator
import pandas as pd
from rats.processors import task, pipeline, Pipeline, PipelineContainer


class DataOut(NamedTuple):
    data: pd.DataFrame


class ModelOut(NamedTuple):
    model: BaseEstimator


class MyContainer(PipelineContainer):
    @task
    def load_data(self, fname: Path) -> DataOut:
        return DataOut(data=pd.read_csv(fname))

    @task
    def train_model(self, data: pd.DataFrame) -> ModelOut:
        return {"model": "trained"}

    @pipeline
    def my_pipeline(self) -> Pipeline:
        load_data = self.load_data()
        train_model = self.get(train_model)
        return self.combine(
            pipelines=[load_data, train_model],
            dependencies=(train_model.inputs.data << load_data.outputs.data),
        )

The above example helps with modularization and bringing exploratory code from notebooks to more permanent code.

The example above illustrates already several important concepts:

• rats.processors.PipelineContainer: we wire up code modularly, i.e., one container organizes and connects tasks and pipelines.
• rats.processors.ux.Pipeline: a data structure that represents a computation graph, or a direct acyclic graph (DAG) of operations.
• rats.processors.task: a decorator to define a computational task, which we refer as processor and register it into the container. The return value of this method is rats.processors.ux.Pipeline, a (single-node) pipeline.
• rats.processors.pipeline: a decorator to register a rats.processors.ux.Pipeline, which can be a combination of other pipelines, or any method that returns a pipeline.

Note that to create a pipeline, you first create tasks (processors) and then combine them into larger pipelines, e.g. MyContainer.load_data and MyContainer.train_model are processors wrapped by the task decorator, and MyContainer.my_pipeline is a pipeline wrapped by the pipeline decorator.

To run the above pipeline, you can do the following:

from rats.apps import autoid, NotebookApp


app = NotebookApp()
app.add(MyContainer())  # add a container to the notebook app
p = app.get(autoid(MyContainer.my_pipeline))  # get a pipeline by its id
app.draw(p)
app.run(p, inputs={"fname": "data.csv"})

Concepts

Concepts	Description
Pipelines	DAG organizing computation tasks
	Orchestrated in run environments
	Figure display
Tasks	Entry point for computation process
	Accepts dynamic inputs/outputs
Combined	Compose tasks & pipelines to draw more complex DAGs.
	Dependency assignment

Features

Features
Modular	Steps become independent; Plug & play		Distributed	Uses required resources (spark or GPUs)
Graph-based	Can operate on the DAG; Enables meta-pipelines		Reusable	Every pipeline is shareable allowing collaborations

Goals

• Flexibility: multiple data sources; multiple ML frameworks (pytorch, sklearn, ...), etc.
• Scalability: both data and compute.
• Usability: everyone should be able to author components and share them.
• Reproducibility: Tracking and recreating results.