Orchestrate an AI Experiment

In this part of the docs, we will walk through how to execute an AI experiment on a Kubernetes cluster using SigOpt. SigOpt should now be connected to a Kubernetes cluster of your choice.

Set Up

If you haven't connected to a cluster yet, you can launch a cluster on AWS, connect to an existing Kubernetes cluster, or connect to an existing, shared K8s cluster.

Then, test whether or not you are connected to a cluster with SigOpt by running:

$ sigopt cluster test

SigOpt will output:

Successfully connected to kubernetes cluster: tiny-cluster

If you're using a custom Kubernetes cluster, you will need to install plugins to get the controller image working:

$ sigopt cluster install-plugins

SigOpt works when all of the files for your model are located in the same folder. So, please create an example directory (mkdir), and then change directories (cd) into that directory:

$ mkdir example && cd example

Then auto-generate templates for a Dockerfile and an SigOpt Configuration YAML file

$ sigopt init

Next, you will create some files and put them in this example directory.

Dockerfile: Define your model environment

For the tutorial, we'll be using a very simple Dockerfile. For instructions on how to specify more requirements see our guide on Dockerfiles. Please copy and paste the following snippet into the autogenerated file named Dockerfile.

FROM python:3.9

RUN pip install --no-cache-dir sigopt

RUN pip install --no-cache-dir scipy==1.7.1
RUN pip install --no-cache-dir scikit-learn==0.24.2
RUN pip install --no-cache-dir numpy==1.21.2

COPY . /sigopt
WORKDIR /sigopt

Define a Model

This code defines a simple SGDClassifier model that measures accuracy classifying labels for the Iris flower dataset. Copy and paste the snippet below to a file titled model.py. Note the snippet below uses SigOpt’s Runs to track model attributes.

# model.py
# SGDClassifier example

# You'll use the SigOpt Training Runs API to communicate with SigOpt
# while your model is running on the cluster.
import sigopt


# These packages will need to be installed in order to run your model.
# To do this, define a requirements.txt file, and provide instructions
from sklearn import datasets
from sklearn.linear_model import SGDClassifier
from sklearn.model_selection import cross_val_score
import numpy

# https://en.wikipedia.org/wiki/Iris_flower_data_set
def load_data():
  iris = datasets.load_iris()
  return (iris.data, iris.target)


def evaluate_model(X, y):
  sigopt.params.setdefaults(
    loss="log",
    penalty="elasticnet",
    log_alpha=-4,
    l1_ratio=0.15,
    max_iter=1000,
    tol=0.001,
  )
  classifier = SGDClassifier(
    loss=sigopt.params.loss,
    penalty=sigopt.params.penalty,
    alpha=10 ** sigopt.params.log_alpha,
    l1_ratio=sigopt.params.l1_ratio,
    max_iter=sigopt.params.max_iter,
    tol=sigopt.params.tol,
  )
  cv_accuracies = cross_val_score(classifier, X, y, cv=5)
  return (numpy.mean(cv_accuracies), numpy.std(cv_accuracies))


# Each execution of model.py should represent one evaluation of your model.
# When this file is run, it loads data, evaluates the model using assignments
if __name__ == "__main__":
  (X, y) = load_data()
  (mean, std) = evaluate_model(X=X, y=y)
  print("Accuracy: {} +/- {}".format(mean, std))
  sigopt.log_metric("accuracy", mean, std)

Notes on implementing your model

When your model runs on a node in the cluster it can use all of the CPUs on that node with multithreading. This is good for performance if your model is the only process running on the node, but in many cases it will need to share those CPUs with other processes (ex. other model runs). For this reason it is a good idea to limit the number of threads that your model library can create in conjunction with the amount of cpu specified in your resources_per_model. This varies by implementation, but some common libraries are listed below:

Numpy

Threads spawned by Numpy can be configured with environment variables, which can be set in your Dockerfile:

ENV MKL_NUM_THREADS=N
ENV NUMEXPR_NUM_THREADS=N
ENV OMP_NUM_THREADS=N

Tensorflow/Keras

Can be configured in the Tensorflow module, see: https://www.tensorflow.org/api_docs/python/tf/config/threading

PyTorch

Can be configured in the PyTorch module, see: https://pytorch.org/docs/stable/generated/torch.set_num_threads.html

Create an orchestration configuration file

Here's a sample SigOpt configuration file that specifies an AI Experiment for the model.py specified above on one CPU.

Please copy and paste the following to a file named run.yml.

# Choose a descriptive name for your model
name: SGD Classifier

# Here, we run the model
run: python model.py
resources:
  requests:
    cpu: 0.5
    memory: 512Mi
  limits:
    cpu: 2
    memory: 512Mi

# We don't need any GPUs for this example, so we'll leave this commented out
# gpus: 1

# SigOpt creates a container for your model. Since we're using an AWS
# cluster, it's easy to securely store the model in the Amazon Elastic Container Registry.
# Choose a descriptive and unique name for each new experiment configuration file.
image: sgd-classifier

Please copy and paste the following to a file named experiment.yml.

# experiment.yml
name: SGD Classifier HPO

metrics:
  - name: accuracy
parameters:
  - name: l1_ratio
    type: double
    bounds:
      min: 0
      max: 1.0
  - name: log_alpha
    type: double
    bounds:
      min: -5
      max: 2

parallel_bandwidth: 2
budget: 60

Execute

So far, SigOpt is connected to your cluster, the Dockerfile defines your model requirements, and you've updated the SigOpt configuration file. SigOpt can now execute an AI Experiment on your cluster.

$ sigopt cluster optimize -r run.yml -e experiment.yml

Monitor progress through CLI

You can monitor the status of SigOpt AI Experiments from the command line using the run name or the Experiment ID.

$ sigopt cluster status experiment/99999

experiment/999999:
  Experiment Name: hoco
  5.0 / 64.0 Observation budget
  5 Observation(s) failed
  Run Name            	Pod phase      	Status         	Link
  run-25ne1woa        	Succeeded      	failed         	https://app.sigopt.com/run/49950
  run-2lkc1ppa        	Succeeded      	failed         	https://app.sigopt.com/run/49975
  run-zggujklx        	Succeeded      	failed         	https://app.sigopt.com/run/49980
  run-zhc9c5q0        	Succeeded      	failed         	https://app.sigopt.com/run/49967
  run-zydkuibj        	Succeeded      	failed         	https://app.sigopt.com/run/49966
  Follow logs: sigopt cluster kubectl logs -ltype=run,experiment=374876 --max-log-requests=4 -f
  View more at: https://app.sigopt.com/experiment/999999

The status will include a command that you can run in your terminal to follow the logs as they are generated by your code.

Monitor progress in the web app

You can monitor experiment progress on https://app.sigopt.com/experiment/[id].

The History tab, https://app.sigopt.com/experiment/[id]/history, shows a complete table of training runs created in the experiment. The State column displays the current state of each training run.

Stop

You can stop your AI Experiment at any point while it's running. This command stops and deletes an AI Experiment on the cluster. All in-progress Training Runs will be terminated.

$ sigopt cluster stop <experiment-id>