README.md

Demonstrates Google ai-platform CLI for multi-class classification as shown in Geewax chapter 18

Geewax, J. J. (2018). Google Cloud Platform in Action. Manning Publications.

This video demo's GCP ai-platform API use via their command line interfaceabove. Please unmute the video to hear sound. Alternatively below the video ere each of the 16 steps followed by a screenshot of its execution:

DEPP.434-55.DV6.mp4

https://cloud.google.com/sdk/gcloud/reference/services/enable

• step 1: Enable the API
  • gcloud services list
  • gcloud services enable ml.googleapis.com

Note: ml-engine replaced by gcloud ai-platform

https://cloud.google.com/sdk/gcloud/reference/ai-platform/models/create

• step 2: create model
  • gcloud ai-platform models create census --description 'Census example'
  • gcloud ai-platform models list
  • gcloud ai-platform versions list --model census

https://github.com/amygdala/tensorflow-workshop/tree/master/workshop_sections/wide_n_deep

• step 3: set up the environment
  • python3 -m venv .venv
  • source .venv/bin/activate
  • mkdir data
• step 4: retrieve the data
  • git clone https://github.com/amygdala/tensorflow-workshop.git
  • ls tensorflow-workshop/workshop_sections/wide_n_deep
  • cp tensorflow-workshop/workshop_sections/wide_n_deep/adult* ./data/

https://cloud.google.com/storage/docs/gsutil/commands/mb

• step 5: move data to GCP google storage
  • gsutil cp ./data/adult.test.csv data gs://dv-auto-ml-depp/data
  • gsutil cp ./data/adult.data.csv data gs://dv-auto-ml-depp/data
  • gsutil ls gs://dv-auto-ml-depp/data

https://github.com/GoogleCloudPlatform/cloudml-samples

• step 6: retrieve the canned model
  • git clone https://github.com/GoogleCloudPlatform/cloudml-samples
  • ls cloudml-samples/census/tensorflowcore/trainer
  • cd cloudml-samples/census/tensorflowcore

https://cloud.google.com/sdk/gcloud/reference/ai-platform/jobs/submit/training

https://cloud.google.com/ai-platform/prediction/docs/runtime-version-list

• step 7: submit the job
  • gcloud ai-platform jobs submit training cloud1 --stream-logs --runtime-version 1.15 --job-dir gs://dv-auto-ml-depp/census --module-name trainer.task --package-path trainer/ --region us-central1 -- --train-files gs://dv-auto-ml-depp/data/adult.data.csv --eval-files gs://dv-auto-ml-depp/data/adult.test.csv --train-steps 10000 --eval-steps 500

• step 8: read the logs
  • took 30 minutes to get log results
  • notes old version of some dependencies
  • python 2.7 end of life at google Jan 2020, but still seems to run
  • global steps / sec = 4.65 and with 10,000 steps / 4.65 / 60 = 35 minutes.

• We can also see logs on the GCP https://console.cloud.google.com/logs/query;query=resource.labels.job_id%3D%22cloud1%22%20timestamp%3E%3D%222020-10-23T07:43:09Z%22?project=msds434dv6

• step 9: review the job
  • gcloud ai-platform jobs describe cloud1

• step 10: review the output files and model stored in the bucket
  • gsutil ls gs://dv-auto-ml-depp/census
  • gsutil ls gs://dv-auto-ml-depp/census/export

https://cloud.google.com/sdk/gcloud/reference/ai-platform/versions/create

• step 11: create version 1 of the model

  • gcloud ai-platform versions create v1 --model census --staging-bucket gs://dv-auto-ml-depp --origin gs://dv-auto-ml-depp/census/export --runtime-version 1.15

• note for step 11: the documentation is a bit sketchy here.

  • version_name is the only mandatory argument and --model the only mandatory flag, but errors are thrown unless you include these flags: --staging_bucket, --origin, --runtime-version

    https://cloud.google.com/sdk/gcloud/reference/ai-platform/versions/create#--model

    https://cloud.google.com/sdk/gcloud/reference/ai-platform/versions/create#--staging-bucket

    https://cloud.google.com/sdk/gcloud/reference/ai-platform/versions/create#--origin

    https://cloud.google.com/sdk/gcloud/reference/ai-platform/versions/create#--runtime-version

https://cloud.google.com/sdk/gcloud/reference/ai-platform/predict

• step 12: locate a test.json file to make a single online predictions from version 1 of the model
  • ls ../test.*
  • gcloud ai-platform predict --model census --version v1 --json-instances ../test.json —> 78.4% confidence the correct class is “<=50k”

cat test.json

• step 13: copy the test.json file into test2.json and see model sensitivity with a change the age feature modified from 25 to 20.
  • ls ../test*
  • gcloud ai-platform predict --model census --version v1 --json-instances ../test2.json —> 84.7% confidence the correct class is “<=50k”

cat test2.json

https://cloud.google.com/sdk/gcloud/reference/ai-platform/jobs/submit/prediction

• step 14: submit prediction job for a test_batch.json dataset with 11 rows for ages 20-70
  • gcloud ai-platform jobs submit prediction prediction1 --model census --version v1 --data-format text --region us-central1 --input-paths gs://dv-auto-ml-depp/test_batch.json --output-path gs://dv-auto-ml-depp/prediction1-output
  • (somewhat disconcerting to have no message when QUEUED for > 5 minutes)

nano test_batch.json

https://console.cloud.google.com/logs/query;query=resource.labels.job_id%3D%22prediction1%22%20timestamp%3E%3D%222020-10-23T09:43:55Z%22?project=msds434dv6

• step 14 continued: Look at the GCP logs for this job

• step 14 continued: Look at the GCP logs for this job

• step15: review the prediction job output in the bucket and the description of the job
  • gsutil ls gs://dv-auto-ml-depp/prediction1-output
  • gsutil cat gs://dv-auto-ml-depp/prediction1-output/prediction.results-00000-of-00001
  • gcloud ai-platform jobs describe prediction1

• note: gcloud ai-platform jobs submit prediction …
  • again the documentation is a bit confusing: on the one hand --runtime_version “must be specified unless --master-image-uri is specified” and in the error “Runtime version should only be specified when the model URI is suppled” https://cloud.google.com/sdk/gcloud/reference/ai-platform/jobs/submit/prediction#--data-format

https://cloud.google.com/storage/docs/deleting-buckets

• step 16: don’t over pay: delete the bucket
  • gsutil rm -r gs://dv-aut-ml-depp
  • alternatives: - rm will remove bucket and any contents in one go - rb will removes bucket only if empty

https://cloud.google.com/sdk/gcloud/reference/ml-engine/versions/delete

• step 16 continued: delete version then delete model then disable api
  • gcloud ai-platform versions delete v1 --model census
  • gcloud ai-platform models delete census
  • gcloud services disable ml.googleapis.com