Skip to content

Pipeline Profiler is a tool for visualizing machine learning pipelines generated by AutoML tools.

License

Notifications You must be signed in to change notification settings

VIDA-NYU/PipelineVis

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PipelineProfiler

AutoML Pipeline exploration tool compatible with Jupyter Notebooks. Supports Auto-Sklearn, Alpha-AutoML and D3M pipeline format.

arxiv badge

System screen

(Shift click to select multiple pipelines)

Paper: https://arxiv.org/abs/2005.00160

Video: https://youtu.be/2WSYoaxLLJ8

Blog: Medium post

Demo

Live demo (Google Colab):

In Jupyter Notebook:

import PipelineProfiler
data = PipelineProfiler.get_heartstatlog_data()
PipelineProfiler.plot_pipeline_matrix(data)

You can also find multiple examples of PipelineProfiler in the Alpha-AutoML repository, an extensible AutoML system for multiple ML tasks.

Install

Option 1: install via pip:

pip install pipelineprofiler

Option 2: Run the docker image:

docker build -t pipelineprofiler .
docker run -p 9999:8888 pipelineprofiler

Then copy the access token and log in to jupyter in the browser url:

localhost:9999

Data preprocessing

PipelineProfiler reads data from the D3M Metalearning database. You can download this data from: https://metalearning.datadrivendiscovery.org/dumps/2020/03/04/metalearningdb_dump_20200304.tar.gz

You need to merge two files in order to explore the pipelines: pipelines.json and pipeline_runs.json. To do so, run

python -m PipelineProfiler.pipeline_merge [-n NUMBER_PIPELINES] pipeline_runs_file pipelines_file output_file

Pipeline exploration

import PipelineProfiler
import json

In a jupyter notebook, load the output_file

with open("output_file.json", "r") as f:
    pipelines = json.load(f)

and then plot it using:

PipelineProfiler.plot_pipeline_matrix(pipelines[:10])

Data postprocessing

You might want to group pipelines by problem type, and select the top k pipelines from each team. To do so, use the code:

def get_top_k_pipelines_team(pipelines, k):
    team_pipelines = defaultdict(list)
    for pipeline in pipelines:
        source = pipeline['pipeline_source']['name']
        team_pipelines[source].append(pipeline)
    for team in team_pipelines.keys():
        team_pipelines[team] = sorted(team_pipelines[team], key=lambda x: x['scores'][0]['normalized'], reverse=True)
        team_pipelines[team] = team_pipelines[team][:k]
    new_pipelines = []
    for team in team_pipelines.keys():
        new_pipelines.extend(team_pipelines[team])
    return new_pipelines

def sort_pipeline_scores(pipelines):
    return sorted(pipelines, key=lambda x: x['scores'][0]['value'], reverse=True)    

pipelines_problem = {}
for pipeline in pipelines:  
    problem_id = pipeline['problem']['id']
    if problem_id not in pipelines_problem:
        pipelines_problem[problem_id] = []
    pipelines_problem[problem_id].append(pipeline)
for problem in pipelines_problem.keys():
    pipelines_problem[problem] = sort_pipeline_scores(get_top_k_pipelines_team(pipelines_problem[problem], k=100))

About

Pipeline Profiler is a tool for visualizing machine learning pipelines generated by AutoML tools.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •