(local) Build & Setup Instructions

General Setup

Install the GCP SDK

https://cloud.google.com/sdk/docs/quickstarts

Clone the repo

• Make sure you have your github SSH key setup
• git clone git@github.com:mmwebster/voxsrc-2020.git

Setup the environment

• Install miniconda: https://docs.conda.io/projects/conda/en/latest/user-guide/install/linux.html
• create the env from existing yaml file: conda env create -f voxsrc-2020/env/conda-vox-env-ubuntu.yaml
• activate the environment: conda activate voxsrc-2020
• set it as the default env: echo "conda activate voxsrc-2020" >> ~/.bashrc

Setup for local runs

Install a dataset

• Move into data dir cd voxsrc-2020/data/
• Run the install script python utils.py --install-local-dataset --src-bucket voxsrc-2020-voxceleb-v4 --src-dataset no_cuda --dst-data-path ./datasets --dst-list-path ./lists --dst-tmp-path ./tmp
• Setup symlinks
  • ln -s ./datasets/vox1_no_cuda ../components/train/tmp/data/vox1_no_cuda
  • ln -s ./datasets/vox1_no_cuda.txt ../components/train/tmp/data/vox1_no_cuda.txt
  • ln -s ./datasets/vox2_no_cuda ../components/train/tmp/data/vox2_no_cuda
  • ln -s ./datasets/vox2_no_cuda.txt ../components/train/tmp/data/vox2_no_cuda.txt
• NOTE: To setup another dataset, replace "no_cuda" with another dataset's name. For example, "full" to install the complete, original dataset

Run the train component (locally, standalone)

• Move into component dir cd ../components/train/
• Execute the component's local run script with default config ./run_local.sh

Setup for remote runs on the Kubeflow cluster

Install Docker

• Perform the initial install: https://docs.docker.com/engine/install/ubuntu/
  • NOTE: If you're using WSL2, you may have to run sudo service docker start after the installation. It doesn't appear to start by default.
• Make sure docker is working by running sudo docker run hello-world
• Enable docker to run with non-root user: https://docs.docker.com/engine/install/linux-postinstall/#manage-docker-as-a-non-root-user
• Make sure docker is working with a non-root user by running docker run hello-world

Authenticate Docker to our container registry on GCP

• https://cloud.google.com/container-registry/docs/advanced-authentication#helpers
  • NOTE: If you don't have a service account for your local machine, ask to have one set up
  • NOTE: You can use gcr.io as the credential helper, with gcloud auth configure-docker gcr.io

Authenticate wandb for your runs on the cluster

• Find your API key: https://app.wandb.ai/authorize
• Set an API key environment variable: echo "export WANDB_API_KEY='[YOUR_KEY_HERE]'" >> ~/.bashrc
• NOTE: Python code in the pipeline file grabs the API key and sets it in the container for the train component

Build an image, compile a pipeline, and upload to Kubeflow web-app

• NOTE: These steps are now identical to running on Kubeflow via the dev-01 machine
• Build the component image (and the image for any other component who's code you modified) with: build_image.sh
• Compile the pipeline files at the top level of the project with python [pipeline-name].py
• Open the Kubeflow web-app and upload the compiled [pipeline-name].tar.gz as a new pipeline (or pipeline version) and start a run
  • For the near future, the web app will be hosted at: https://kf-train-6.endpoints.voxsrc-2020-dev-1.cloud.goog/#