Wide & Deep with TensorFlow

This directory contains the code for running a Wide and Deep model. It also runs in Cloud ML Engine. This code has been tested on Python 2.7 but should also run on Python 3.5

Follow along the codelab here: http://bit.ly/widendeep-slides

About the dataset and model

Wide and deep jointly trains wide linear models and deep neural networks -- to combine the benefits of memorization and generalization for recommender systems. See the research paper for more details. The code is based on the TensorFlow wide and deep tutorial.

We will use the Kaggle Criteo Dataset to predict the probability that an ad is clicked.

The dataset is downloaded as part of the script (and in the cloud directly uses the copy stored online).

If you wish to download a copy, the data are located here:

gs://dataset-uploader/criteo-kaggle/small_version -- 2.5MB, 10K rows
gs://dataset-uploader/criteo-kaggle/medium_version -- 273MB, 1M rows
gs://dataset-uploader/criteo-kaggle/large_version -- 2.7GB, 10M rows

Each folder contains 2 files: train.csv and eval.csv

The command line tool gsutil is part of the Google Cloud SDK, and comes with gcloud. No gsutil but still want to download? Replace "gs://" with 'https://storage.googleapis.com/', for example: https://storage.googleapis.com/dataset-uploader/criteo-kaggle/small_version/train.csv

Training and evaluation

This repo presents 3 methods of running the model: locally, on a jupyter notebook, and on Google Cloud ML Engine.

The commands below assume you are in this directory (wide_n_deep).

You should move to it with cd workshop_sections/wide_n_deep

Local

python trainer/task.py

Jupyter Notebook

Run the notebook, and step through the cells.

jupyter notebook

Google Cloud Machine Learning Engine

The workflow to run this on Cloud Machine Learning Engine is to do a local run first, then move to the cloud.

Test it locally:

$ gcloud ml-engine local train --package-path=trainer --module-name=trainer.task
TensorFlow version 1.0.0
model directory = models/model_WIDE_AND_DEEP_1491431579
estimator built
fit done
evaluate done
Accuracy: 0.84125
Model exported to models/model_WIDE_AND_DEEP_1491431579/exports

Run it in the cloud:

Ensure you have the project you want to work in selected. You can check using gcloud config list

If you need to set it to a new value, do so using gcloud config set <YOUR_PROJECT_ID_HERE>

You should also make sure that the Cloud ML Engine API is turned on for your project. More info about getting setup is here: https://cloud.google.com/ml-engine/docs/quickstarts/command-line

Next, set the following environment variables and submit a training job.

gcloud config set compute/region us-central1
gcloud config set compute/zone us-central1-c

export PROJECT_ID=`gcloud config list project --format "value(core.project)"`
export BUCKET=gs://${PROJECT_ID}-ml
export JOB_NAME=widendeep_${USER}_$(date +%Y%m%d_%H%M%S)
export TRAIN_PATH=${BUCKET}/${JOB_NAME}

gcloud ml-engine jobs submit training ${JOB_NAME} --package-path=trainer --module-name=trainer.task  --region=us-central1 --job-dir=${TRAIN_PATH}

When you are ready to run it on a more power cluster, you can customize a config.yaml file. Included in this repo are 2 examples, one for the STANDARD_1 sizing, and one custom setup which includes 3 GPU-enabled machines.

gcloud ml-engine jobs submit training ${JOB_NAME} --package-path=trainer --module-name=trainer.task --job-dir=${TRAIN_PATH} --config config_standard.yaml

gcloud ml-engine jobs submit training ${JOB_NAME} --package-path=trainer --module-name=trainer.task --job-dir=${TRAIN_PATH} --config config_gpu.yaml

You can check the status of your training job with the command:

gcloud ml-engine jobs describe $JOB_NAME

You can also see it's progress in your cloud console and view the logs.

To run another job (in your dev workflow), simply set a new JOB_NAME and TRAIN_PATH and then re-run the jobs.summit call. Job names must be unique.

export JOB_NAME=widendeep_${USER}_$(date +%Y%m%d_%H%M%S)
export TRAIN_PATH=${BUCKET}/${JOB_NAME}
gcloud ml-engine jobs submit training ${JOB_NAME} --package-path=trainer --module-name=trainer.task  --region=us-central1 --job-dir=${TRAIN_PATH}

Your trained model

Whether you ran your training locally or in the cloud, you should now have a set of files exported. If you ran this locally, it will be located in someplace similar to models/model_WIDE_AND_DEEP_1234567890/exports/1234567890. If you ran it in the cloud, it will be located in the GCS bucket that you passed.

The trained model files that were exported are ready to be used for prediction.

Prediction

You can run prediction jobs in Cloud ML Engine as well, using the Prediction Service.

Before we begin, if you trained a model locally, you should upload the contents that were exported (something like saved_model.pb and a folder called variables) to a Google Cloud Storage location and make a note of its address (gs://<BUCKET_ID>/path/to/model)

Now we are ready to create a model

export MODEL_NAME='my_model'
gcloud ml-engine models create $MODEL_NAME

Next, create a 'version' of that model

export VERSION_NAME='my_version'
export DEPLOYMENT_SOURCE='gs://LOCATION_OF_MODEL_FILES'
gcloud ml-engine versions create $VERSION_NAME --model $MODEL_NAME --origin $DEPLOYMENT_SOURCE

Finally, make a prediction with your newly deployed version!

gcloud ml-engine predict --model $MODEL_NAME --version $VERSION_NAME --json-instances test_instance.json

sa757 / widendeep