Add Interface for Preprocessing, Training, and Prediction w/ Google ML Engine

[KatJHuang]

Google ML Engine provides an ensemble of utilities that handles tasks from training to prediction. It also talks with other Google services: Cloud Function, Cloud Storage we employ for manipulating and storing our data.

Since we already have a trained model, the most pressing need is to deploy the model and relay the result to the Android app. Automated training using existing frameworks is then a desired feature but not immediately necessary.

Acceptance Criteria GIVEN an architecture of a model WHEN I supply a training set THEN ML Engine trains the model and saves the model parameters

GIVEN an audio file and a trained model stored on Cloud Storage WHEN I signal for a prediction for the audio THEN ML engine returns me the prediction result

Update Prof Rose hinted at the peril of training and testing with different datasets in his comment to the project proposal, and suggested we obtain our own labelled datasets. Using additional datasets makes it necessary to automate the training process, as manually uploading from local training is prone to error and hard to version control. Therefore we should use Google ML Engine for training as well.

[KatJHuang]

Currently following tutorial at https://cloud.google.com/ml-engine/docs/tensorflow/getting-started-training-prediction for an introduction to ML engine.

So far have done:

• enabled ML engine service on speech-recognition project
• got credentials for using the service (credentials generated by Google as JSON, downloaded to my local machine, available upon request)
  • created a service account named: ece496emotionrecogntion
• submitted a training job from local and stored trained model to Google Cloud Storage
• deployed the model coming from the training step above and ran an online prediction
• deleted job to prevent extra cost from incurring

Observations

• the syntax for doing training is very similar to what we would do locally; except it has GCP commands wrapping the parameters for local command.
• we can run prediction as long as we have saved binary file for the trained model somewhere in Google Cloud Storage

Next Steps

• Find how to do preprocessing with ML engine (maybe put functions for that purpose in the same file?)
• Find how to input audio for a prediction job

Further readings: https://github.com/GoogleCloudPlatform/cloudml-samples/tree/master/census/