Workflows automate machine learning tasks, combining GPU instances with an expressive syntax to generate production-ready machine learning pipelines with a few lines of code.
defaults: At the top of the YAML Workflow file, you can specify default parameters to be used throughout the entire Workflow. This includes environment variables and default machine instance configuration. Instances can also be specified per-job.
inputs: The inputs block allows you to specify named inputs (for example, a versioned dataset to be referenced and consumed by your jobs). You can also collect inputs in a separate YAML and reference this file as an inputPath when creating a Workflow run. Workflow and job-level inputs can be of type: dataset (a persistent, versioned collection of data), string (for example, a generated value or ID that may be output from another job) or volume (a temporary workspace mounted onto a job’s container). Datasets must be defined in advance of being referenced in a workflow. See Create Datasets for the Workflow for more information.
jobs: Jobs are also sometimes referred to as “steps” within the Gradient Workflow. A job is an individual task that executes code (such as a training a machine learning model) and can consume inputs and produce outputs.
To run this Workflow, define datasets named test-one, test-two, and test-three as described in the Create Datasets for the Workflow documentation. Also, to make use of the secret named hello in the inputs section, define a secret.
defaults:
# clusterId defaults to the NY2 public cluster, setting this parameter this is equaivalent to using the `--clusterId` flag on the command line.
# This parameter often used for github triggered workflows running on private clusters.
clusterId: clusterId
# Default environment variables for all jobs. Can use any supported
# substitution syntax (named secrets, ephemeral secrets, etc.).
env:
# This environment variable uses a Gradient secret called "hello".
HELLO: secret:hello
# Default instance type for all jobs
resources:
instance-type: P4000
container-registries: # optional
- my-registry
# Workflow takes two inputs, neither of which have defaults. This means that
# when the Workflow is run the corresponding input for these values are
# required, for example:
#
# {"inputs": {"data": {"id": "test-one"}, "echo": {"value": "hello world"}}}
#
inputs:
data:
type: dataset
with:
ref: test-one
echo:
type: string
with:
value: "hello world"
jobs:
job-1:
# These are inputs for the "job-1" job; they are "aliases" to the
# Workflow inputs.
#
# All inputs are placed in the "/inputs/<name>" path of the run
# containers. So for this job you would have the paths "/inputs/data"
# and "/inputs/echo".
inputs:
# The "/inputs/data" directory would contain the contents for the dataset
# version. ID here refers to the name of the dataset, not its dataset ID.
data: workflow.inputs.data
# The "/inputs/echo" file would contain the string of the Workflow input
# "echo".
echo: workflow.inputs.echo
# These are outputs for the "job-1" job.
#
# All outputs are read from the "/outputs/<name>" path.
outputs:
# A directory will automatically be created for output datasets and
# any content written to that directory will be committed to a newly
# created dataset version when the jobs completes.
data2:
type: dataset
with:
id: test-two
# The container is responsible creating the file "/outputs/<name>" with the
# content being a small-ish utf-8 encoded string.
echo2:
type: string
# Set job-specific environment variables
env:
TSTVAR: test
# Set action
uses: container@v1
# Set action arguments
with:
args:
- bash
- -c
- find /inputs/data > /outputs/data2/list.txt; echo ENV $HELLO $TSTVAR > /outputs/echo2; cat /inputs/echo; echo; cat /outputs/data2/list.txt /outputs/echo2
image: bash:5
job-2:
inputs:
# These inputs use job-1 outputs instead of Workflow inputs. You must
# specify job-1 in the needs section to reference them here.
data2: job-1.outputs.data2
echo2: job-1.outputs.echo2
outputs:
data3:
type: dataset
with:
ref: test-three
# List of job IDs that must complete before this job runs
needs:
- job-1
uses: container@v1
with:
args:
- bash
- -c
- wc -l /inputs/data2/list.txt > /outputs/data3/summary.txt; cat /outputs/data3/summary.txt /inputs/echo2
image: bash:5
Below is an example of a valid workflow.yaml spec. It clones the repository from https://github.com/NVlabs/stylegan2, generates images from the repo script run_generator.py, and outputs the results to the Gradient-managed dataset demo-dataset.
jobs:
CloneRepo:
resources:
instance-type: C5
outputs:
repo:
type: volume
uses: git-checkout@v1
with:
url: https://github.com/NVlabs/stylegan2.git
StyleGan2:
resources:
instance-type: P4000
needs:
- CloneRepo
inputs:
repo: CloneRepo.outputs.repo
outputs:
generatedFaces:
type: dataset
with:
ref: demo-dataset
uses: script@v1
with:
script: |-
pip install scipy==1.3.3
pip install requests==2.22.0
pip install Pillow==6.2.1
cp -R /inputs/repo /stylegan2
cd /stylegan2
python run_generator.py generate-images \
--network=gdrive:networks/stylegan2-ffhq-config-f.pkl \
--seeds=6600-6605 \
--truncation-psi=0.5 \
--result-dir=/outputs/generatedFaces
image: tensorflow/tensorflow:1.14.0-gpu-py3