Awesome

Hydra-Vertex-AI-Pipeline

This repository focus on how to run code written in Hydra on the Vertex AI Pipeline.

• The first half of the README describes how to use the repository
• The second half provides a brief description of Vertex Pipeline, compatibility issues with Hydra, and how to resolve them.

The Japanese blog is here

• data prepare: download MNIST data
• train: perform training

✅ step1. Build and push Docker Images

• Decide URIs of data_prepare and train to push to GCP, and write them in the push-data-prepare-image and push-train-image of Makefile.

• Then, in the root directory of this repository, run

make push-data-prepare-image
make push-train-image

to build and pushed two Docker Images.

※ In the sample code, the Docker Image of the data prepare is built in components/data_prepare in this repository. The process of the data prepare is written by Hydra. In detail, after writing function codes in functions, You can determine the functions to be processed as parameters by writing them in config.yaml, which is a similar way to manage hyperparameters in AI training.

Also, the Docker Image of the train is from train code written in Hydra.

✅ step2. Building python environment

Run

make build-python-environment

✅ step3. Compile Vertex AI pipeline system

1. Add the image URIs used in step1 to implementation.container.image of data_prepare.yaml and train.yaml.
2. Add information about your GCP account to pipeline.yaml
3. Run

poetry run python pipeline.py

then vertex-pipelines-sample.json will be created

✅ step4. Run Vertex AI Pipeline on GCP

There are 2 ways to run a pipeline.

1. Submit JSON file to GCP console.

1. Access the console of Pipeline
2. Click CREATE RUN at the top of the console screen.

3. Click Pipeline and choose Upload file. Then upload vertex-pipelines-sample.json which was created in step4.

4. Click SUBMIT to run the pipeline.

2. Submit JSON file via python.

Run the following command

poetry run python submit_pipeline_job.py

👨‍🏭 What is the ML Pipeline?

Training process of Deep Learning usually consists of various processes such as data preprocessing, training, and evaluation. Training in which these processes are performed on a single machine or container is commonly referred to as a Monolith system (Figure 1 (a)). I guess many people's first experience with machine learning is likely to have been with this system.

On the other hand, when considering the operation of machine learning, it is important to

• Ensure reproducibility of data and models (e.g., if randomness is included in preprocessing and executed in every training process, it is difficult to find out the factors that changed the results).
• Different processes require different machine specifications (e.g., some processes require high memory while another require GPUs).
• Because the processes are independent, they can be used interchangeably.

For these reasons, using a pipeline is recommended for training, where each process (generally called a component) is processed independently (Figure 1 (b)).

For more information, see what-a-machine-learning-pipeline-is-and-why-its-important or Full Stack Deep Learning.

Also see Google's blog (Rules of Machine Learning: Best Practices for ML Engineering) is based on the assumption that Pipeline is used for training.

💻 What is Vertex AI pipeline

Let's say imagine building a pipeline system from scratch that divides each training process into components, each running on a different spec machine. You might feel that it would be very complex.

On the other hand, with Vertex AI Pipeline, you can easily build an ML Pipeline in conjunction with other GCP services, as shown in Figure 2.

For more information about Vertex AI Pipeline, please see the official documentation. If you want to know an excellent sample code, please see this sample code, which is truly excellent for starting the Vertex AI Pipeline.

Hydra and Vertex AI Pipeline

Hydra is an excellent library for hyperparameter management. Various training codes have been written, as in this example. On the other hand, problems arise when trying to use containers written in Hydra as components of the Vertex AI Pipeline.

😖 Problem

In Vertex AI Pipeline, the arguments to be passed to each component are defined in the args of the YAML file. According to the official document of Vertex AI, it is necessary to write like the below.

    command: [python3, main.py]
    args: [
      --project, {inputValue: project},
    ]

this leads following command being passed to the container

python3 main.py --project <value of project>

However, passing commands in this format to a container using Hydra will result in an error (Figure 3). This is because the code written in Hydra requires the command to be passed in the following format

python3 main.py project=<value of project>

💡 Solution

The coding style of the YAML file needs to be changed as follows (Figure 3).

    command: [python3, main.py]
    args: [
      'project={{$.inputs.parameters["project"]}}',
    ]

Commonly used arguments and the corresponding conversion methods are listed in Table 1.

Table 1: Correspondence table for converting the recommended argument passing in Vertex AI for Hydra.