DigitalGlobe / tensorflow-docker

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Example Project using the Tensorflow Docker image

Includes scripts to build, run and push docker container

JSON files

taskDefinition.json

This is the task definition for the task in GBDx. It will be registered as test-tensorflow. It tries to import tensorflow in Python and captures the stdout/stderr to the dump directory as shown in testTensorflow.sh

To add a task definition, use the Register Task Definition API

To view a task defintion, use the Get Task Definition API

To delete a task, use the Delete a Task API

tensorflowWorkflow.json

This is the sample workflow. It mounts the s3 path as input, runs the test-tensorflow task, and writes its output to s3 using the StageToS3 task

Shell scripts

buildTestTensorflow.sh

Constructs the tdgp/testtensorflow image

pushTestTensorflow.sh

Pushes the tdgp/testtensorflow image to DockerHub

runTestTensorflow.sh

Runs the tdgp/testtensorflow image locally. Assumes that /opt/mnt has the folder structure that GBDx expects. In particular, the directory /opt/mnt/work/input/data/ must exist. If one is trying to fully mimic the GBDx folder structure, the following files should be present:

  • Folder /opt/mnt/work/input/<input_dir_name> for all directory inputs
  • File /opt/mnt/work/gbdx_runtime.json if using task impersonation. This should be in the format: { "user_token": "<GBDX_TOKEN>" }
  • File /opt/mnt/work/input/ports.json for string inputs. This should be in the format: { "<param_1_name>": "<param_1_value>", "<param_2_name>": "<param_2_value>" }

NOTE: The /opt/mnt/work/output folder will not be created by the system. It must be manually created

testTensorflow.sh

Imports tensorflow within Python and captures the output

About


Languages

Language:Shell 100.0%