Using Pachyderm to Create a Batch Inferencing Pipeline

In this guide, we’ll help you create a simple batch inferencing project in Pachyderm, train your model using a Determined cluster, and then use the model in an inferencing pipeline.

Note

Inference describes the process of making predictions with a trained model. Specifically, when you input new data into a model to obtain predictions, you are performing inference.

Objectives

After completing the steps in this tutorial, you will have a a fully-built batch inferencing pipeline with results and you will understand how to leverage Pachyderm when working with your Determined cluster.

By following these instructions, you will:

• Install Determined and Pachyderm locally.

• Create a batch inferencing project and train a model.

• Bring a downloaded checkpoint from Determined into Pachyderm.

• Create pipelines for training the data, inferencing, and collecting the results.

Prerequisites

The following prerequisites are required:

• Docker with Kubernetes enabled

• A running Determined cluster

• Pachyderm installed

Note

• To set up Determined locally, visit the quick installation instructions: Quick Installation

• To set up Pachyderm locally, visit First-Time Setup or Pachyderm Local Deployment Guide

Get the Tutorial Files

• Before starting this tutorial, ensure you have copied the files from the Github repository to your local directory.

Note

The tutorial files include the experiment configuration file, the files needed to create each pipeline, as well as the data needed to train the model.

Verify Your Installation of Determined and Pachyderm

Verify you have installed the latest version of both products.

DeterminedPachyderm

Run the following command to verify your installation of Determined:

det version

You should see a configuration summary with details like the version number, master address, and more.

Create a Project in Pachyderm for Batch Inferencing

To begin, you’ll need a project to group together all your training and inference repositories (repos) and pipelines. To do this, you’ll create a new project in Pachyderm.

Run the following command to create a new project:

pachctl create project batch-inference-1

Next, update your Pachyderm configuration to set the context to the project you’ve just created:

pachctl config update context --project batch-inference-1

By setting up a dedicated project, you ensure that all related repos and pipelines are organized cohesively. This encapsulation makes it easier to manage batch inferencing workflows in the future.

You are now ready to create a project repo.

Create Repos in Pachyderm for Training Data

To manage our training data effectively, we’ll first need to create repos for storing the data. We’ll use a typical 80:20 train/test split.

To create the train/test repos, run the following commands:

pachctl create repo test
pachctl create repo train
pachctl list repo

The expected output should be something like this:

PROJECT	NAME	CREATED	SIZE (MASTER)	DESCRIPTION
batch-inference-1	train	3 seconds ago	≤ 0B
batch-inference-1	test	6 seconds ago	≤ 0B

To see the repos in the Pachyderm UI, visit localhost in your browser and then select batch-inference-1 to view the project.

Create a Pipeline for Training Data

Before we add data to our train and test repos, we’ll need to create a pipeline. This pipeline will merge the data from the train and test repos, then compress them into a tar file. This provides easy data access and also serves as a convenient checkpoint for data cleanup or transformations.

Run the following commands to create the pipeline:

pachctl create pipeline -f ./pachyderm/pipelines/compress/compress.json
pachctl list pipeline

You should see the new pipeline in your browser.

Add Data to the Train and Test Repos

We are now ready to add files to our repos. Once we add the files, Pachyderm will send them through the compress pipeline and produce a single tar file.

To add data, run the following commands:

pachctl put file -r train@master:/ -f ./data/train_1000
pachctl put file -r test@master:/ -f ./data/test_1000

To verify the repos:

pachctl list repo

The repo list displays and includes the size of each repo.

PROJECT	NAME	CREATED	SIZE (MASTER)	DESCRIPTION
batch-inference-1	compress	5 seconds ago	≤ 21.13MiB	Output repo for pipeline batch-inference-1/compress.
batch-inference-1	train	2 minutes ago	≤ 17.36MiB
batch-inference-1	test	2 minutes ago	≤ 4.207MiB

Train Your Models in Determined

We’ll use a Determined cluster to train our model based on the compressed data that is stored and versioned in Pachyderm. To run an experiment in Determined, you need, at minimum, a script and an experiment configuration (YAML) file.

To run our experiment in Determined using our data in our Pachyderm pipeline, we’ll just need to provide the following information to Determined: the Pachyderm host, port, project, repo, and branch. With this information, Determined will know where to download the data we want to use for our experiment. We provide this information to Determined via the experiment configuration file.

View the configuration for the experiment by running the following command:

cat ./determined/train.yaml

The configuration should resemble the following:

description: catdog_single_train
data:
  pachyderm:
    host: PACHD_HOST
    port: PACHD_PORT
    project: batch-inference-1
    repo: compress
    branch: master
hyperparameters:
  learning_rate: 0.005
  global_batch_size: 16
  weight_decay: 1.0e-4
  nesterov: true
searcher:
  name: single
  metric: accuracy
  max_length:
    batches: 100
  smaller_is_better: false
entrypoint: model_def:CatDogModel
scheduling_unit: 10
min_validation_period:
  batches: 10

Create the experiment by first modifying and then running the following command:

Note

To modify the command and avoid a failed experiment, you’ll need to retrieve the host and port of your Pachyderm cluster.

det experiment create ./determined/train.yaml ./determined --config data.pachyderm.host=Yourhostname --config data.pachyderm.port=80

Upon successful creation, you should see the following output:

Preparing files to send to master… 19.0KB and 11 files Created experiment 10

Visit the Determined dashboard to view the progress of your experiment. One way to do this is to enter the following URL: http://localhost:8080/ in your browser. This is the cluster address for your local training environment.

Accept the default username of determined, and click Sign In. After signing in, you’ll need to create a strong password.

Wait until Determined displays Best Checkpoint before continuing on to the next step. Then, obtain the ID of the completed trial, you’ll need this to download the checkpoint.

Download Checkpoints from Determined

After training your model using Determined, you’ll likely want to access and retain the best-performing checkpoints. By following the steps below, you can download the desired checkpoint and subsequently store it within a Pachyderm repo for future reference.

To run the following command, ensure you replace the trial ID with the ID of your recently completed trial.

det trial download 10 --best -o ./data/checkpoints/catdog1000

Upon execution, you should expect to see the following output:

Local checkpoint path: data/checkpoints/catdog1000

Create Repos for Storing Our Models

Now, let’s create a new repo to store our models:

pachctl create repo models

Verify the repo’s creation by listing all available repos:

pachctl list repo

Lastly, to add the checkpoint to your newly created repo, run the following command:

pachctl put file -r models@master:/catdog1000 -f ./data/checkpoints/catdog1000

Create a Repo and Pipeline for Inferencing

Now that we have our trained model stored in the models repo, let’s establish a new repository and pipeline dedicated to inferencing. This step allows for the model’s utilization in predicting batches of files. Additionally, to enhance the processing speed and manage higher loads, we can introduce a parallelism specification in our pipeline spec.

Start by creating the predict repo:

pachctl create repo predict

To verify the repo creation, list all the available repos:

pachctl list repo

The table output should be as follows:

PROJECT	NAME	CREATED	SIZE (MASTER)	DESCRIPTION
batch-inference-1	predict	2 seconds ago	≤ 0B
batch-inference-1	models	36 seconds ago	≤ 179.8MiB
batch-inference-1	compress	38 minutes ago	≤ 21.13MiB	Output repo for pipeline batch-inference-1/compress.
batch-inference-1	train	41 minutes ago	≤ 17.36MiB
batch-inference-1	test	41 minutes ago	≤ 4.207MiB

Next, create the pipeline for prediction:

pachctl create pipeline -f ./pachyderm/pipelines/predict/predict.json

Add Some Files for Pachyderm/Determined to Inference

After setting up the pipeline, we can now push some files for the prediction. This is flexible; you can add any number of files to the predict repo at any time. Keep in mind that our pipelines will not only generate an image as output but also store the prediction result as a row in a CSV.

To add files for prediction, run:

pachctl put file -r predict@master -f ./data/predict/batch_10

Add a Results Pipeline

Next, we’ll set up a results pipeline. Its role is to gather all the predictions and then process them to generate various visualizations like charts. Additionally, it can store these predictions in a structured database format.

Start by creating the results pipeline:

pachctl create pipeline -f ./pachyderm/pipelines/results/results.json

Add More Files for Prediction and Results Pipelines

To watch all of the prediction and results pipelines run, add more files by running the following commands:

pachctl put file -r predict@master -f ./data/predict/batch_5_2

Next Steps

Congratulations! You’ve successfully streamlined the task of building a batch inferencing pipeline using Pachyderm and Determined together! Now you can try applying the same steps to your own use case.