Training on Cloud AI Platform

This notebook illustrates distributed training and hyperparameter tuning on Cloud AI Platform (formerly known as Cloud ML Engine).

In [ ]:
# change these to try this notebook out
BUCKET = 'cloud-training-demos-ml'
PROJECT = 'cloud-training-demos'
REGION = 'us-central1'
In [ ]:
import os
os.environ['BUCKET'] = BUCKET
os.environ['PROJECT'] = PROJECT
os.environ['REGION'] = REGION
os.environ['TFVERSION'] = '1.13'
In [ ]:
gcloud config set project $PROJECT
gcloud config set compute/region $REGION
In [ ]:
if ! gsutil ls | grep -q gs://${BUCKET}/babyweight/preproc; then
  gsutil mb -l ${REGION} gs://${BUCKET}
  # copy canonical set of preprocessed files if you didn't do previous notebook
  gsutil -m cp -R gs://cloud-training-demos/babyweight gs://${BUCKET}
In [ ]:
gsutil ls gs://${BUCKET}/babyweight/preproc/*-00000*

Now that we have the TensorFlow code working on a subset of the data, we can package the TensorFlow code up as a Python module and train it on Cloud AI Platform.

Train on Cloud AI Platform

Training on Cloud AI Platform requires:

  1. Making the code a Python package
  2. Using gcloud to submit the training code to Cloud AI Platform

Ensure that the AI Platform API is enabled by going to this link.

Lab Task 1

The following code edits babyweight/trainer/

In [ ]:
%%writefile babyweight/trainer/
import argparse
import json
import os

from . import model

import tensorflow as tf

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
        help = 'GCS path to data. We assume that data is in gs://BUCKET/babyweight/preproc/',
        required = True
        help = 'GCS location to write checkpoints and export models',
        required = True
        help = 'Number of examples to compute gradient over.',
        type = int,
        default = 512
        help = 'this model ignores this field, but it is required by gcloud',
        default = 'junk'
        help = 'Hidden layer sizes to use for DNN feature columns -- provide space-separated layers',
        nargs = '+',
        type = int,
        default=[128, 32, 4]
        help = 'Embedding size of a cross of n key real-valued parameters',
        type = int,
        default = 3

    ## TODO 1: add the new arguments here 
        help = 'Number of examples (in thousands) to run the training job over. If this is more than actual # of examples available, it cycles through them. So specifying 1000 here when you have only 100k examples makes this 10 epochs.',
        type = int,
        default = 5000
        help = 'Specify a pattern that has to be in input files. For example 00001-of will process only one shard',
        default = 'of'
        help = 'Positive number of steps for which to evaluate model. Default to None, which means to evaluate until input_fn raises an end-of-input exception',
        type = int,       
        default = None
    ## parse all arguments
    args = parser.parse_args()
    arguments = args.__dict__

    # unused args provided by service
    arguments.pop('job_dir', None)
    arguments.pop('job-dir', None)

    ## assign the arguments to the model variables
    output_dir = arguments.pop('output_dir')
    model.BUCKET     = arguments.pop('bucket')
    model.BATCH_SIZE = arguments.pop('batch_size')
    model.TRAIN_STEPS = (arguments.pop('train_examples') * 1000) / model.BATCH_SIZE
    model.EVAL_STEPS = arguments.pop('eval_steps')    
    print ("Will train for {} steps using batch_size={}".format(model.TRAIN_STEPS, model.BATCH_SIZE))
    model.PATTERN = arguments.pop('pattern')
    model.NEMBEDS= arguments.pop('nembeds')
    model.NNSIZE = arguments.pop('nnsize')
    print ("Will use DNN size of {}".format(model.NNSIZE))

    # Append trial_id to path if we are doing hptuning
    # This code can be removed if you are not using hyperparameter tuning
    output_dir = os.path.join(
            os.environ.get('TF_CONFIG', '{}')
        ).get('task', {}).get('trial', '')

    # Run the training job

Lab Task 2

The following code edits babyweight/trainer/

In [ ]:
%%writefile babyweight/trainer/
import shutil
import numpy as np
import tensorflow as tf


BUCKET = None  # set from
PATTERN = 'of' # gets all files

# Determine CSV, label, and key columns
CSV_COLUMNS = 'weight_pounds,is_male,mother_age,plurality,gestation_weeks,key'.split(',')
LABEL_COLUMN = 'weight_pounds'
KEY_COLUMN = 'key'

# Set default values for each CSV column
DEFAULTS = [[0.0], ['null'], [0.0], ['null'], [0.0], ['nokey']]

# Define some hyperparameters
NNSIZE = [64, 16, 4]

# Create an input function reading a file using the Dataset API
# Then provide the results to the Estimator API
def read_dataset(prefix, mode, batch_size):
    def _input_fn():
        def decode_csv(value_column):
            columns = tf.decode_csv(value_column, record_defaults=DEFAULTS)
            features = dict(zip(CSV_COLUMNS, columns))
            label = features.pop(LABEL_COLUMN)
            return features, label
        # Use prefix to create file path
        file_path = 'gs://{}/babyweight/preproc/{}*{}*'.format(BUCKET, prefix, PATTERN)

        # Create list of files that match pattern
        file_list = tf.gfile.Glob(file_path)

        # Create dataset from file list
        dataset = (  # Read text file
                    .map(decode_csv))  # Transform each elem by applying decode_csv fn
        if mode == tf.estimator.ModeKeys.TRAIN:
            num_epochs = None # indefinitely
            dataset = dataset.shuffle(buffer_size = 10 * batch_size)
            num_epochs = 1 # end-of-input after this
        dataset = dataset.repeat(num_epochs).batch(batch_size)
        return dataset.make_one_shot_iterator().get_next()
    return _input_fn

# Define feature columns
def get_wide_deep():
    # Define column types
    is_male,mother_age,plurality,gestation_weeks = \
                        ['True', 'False', 'Unknown']),
                        ['Single(1)', 'Twins(2)', 'Triplets(3)',
                         'Quadruplets(4)', 'Quintuplets(5)','Multiple(2+)']),

    # Discretize
    age_buckets = tf.feature_column.bucketized_column(mother_age, 
    gestation_buckets = tf.feature_column.bucketized_column(gestation_weeks, 
    # Sparse columns are wide, have a linear relationship with the output
    wide = [is_male,
    # Feature cross all the wide columns and embed into a lower dimension
    crossed = tf.feature_column.crossed_column(wide, hash_bucket_size=20000)
    embed = tf.feature_column.embedding_column(crossed, NEMBEDS)
    # Continuous columns are deep, have a complex relationship with the output
    deep = [mother_age,
    return wide, deep

# Create serving input function to be able to serve predictions later using provided inputs
def serving_input_fn():
    feature_placeholders = {
        'is_male': tf.placeholder(tf.string, [None]),
        'mother_age': tf.placeholder(tf.float32, [None]),
        'plurality': tf.placeholder(tf.string, [None]),
        'gestation_weeks': tf.placeholder(tf.float32, [None]),
        KEY_COLUMN: tf.placeholder_with_default(tf.constant(['nokey']), [None])
    features = {
        key: tf.expand_dims(tensor, -1)
        for key, tensor in feature_placeholders.items()
    return tf.estimator.export.ServingInputReceiver(features, feature_placeholders)

# create metric for hyperparameter tuning
def my_rmse(labels, predictions):
    pred_values = predictions['predictions']
    return {'rmse': tf.metrics.root_mean_squared_error(labels, pred_values)}

# Create estimator to train and evaluate
def train_and_evaluate(output_dir):
    tf.summary.FileWriterCache.clear() # ensure filewriter cache is clear for TensorBoard events file
    wide, deep = get_wide_deep()
    EVAL_INTERVAL = 300 # seconds

    ## TODO 2a: set the save_checkpoints_secs to the EVAL_INTERVAL
    run_config = tf.estimator.RunConfig(save_checkpoints_secs = EVAL_INTERVAL,
                                        keep_checkpoint_max = 3)
    ## TODO 2b: change the dnn_hidden_units to NNSIZE
    estimator = tf.estimator.DNNLinearCombinedRegressor(
        model_dir = output_dir,
        linear_feature_columns = wide,
        dnn_feature_columns = deep,
        dnn_hidden_units = NNSIZE,
        config = run_config)
    # illustrates how to add an extra metric
    estimator = tf.contrib.estimator.add_metrics(estimator, my_rmse)
    # for batch prediction, you need a key associated with each instance
    estimator = tf.contrib.estimator.forward_features(estimator, KEY_COLUMN)
    ## TODO 2c: Set the third argument of read_dataset to BATCH_SIZE 
    ## TODO 2d: and set max_steps to TRAIN_STEPS
    train_spec = tf.estimator.TrainSpec(
        input_fn = read_dataset('train', tf.estimator.ModeKeys.TRAIN, BATCH_SIZE),
        max_steps = TRAIN_STEPS)
    exporter = tf.estimator.LatestExporter('exporter', serving_input_fn, exports_to_keep=None)

    ## TODO 2e: Lastly, set steps equal to EVAL_STEPS
    eval_spec = tf.estimator.EvalSpec(
        input_fn = read_dataset('eval', tf.estimator.ModeKeys.EVAL, 2**15),  # no need to batch in eval
        steps = EVAL_STEPS,
        start_delay_secs = 60, # start evaluating after N seconds
        throttle_secs = EVAL_INTERVAL,  # evaluate every N seconds
        exporters = exporter)
    tf.estimator.train_and_evaluate(estimator, train_spec, eval_spec)

Lab Task 3

After moving the code to a package, make sure it works standalone. (Note the --pattern and --train_examples lines so that I am not trying to boil the ocean on my laptop). Even then, this takes about 3 minutes in which you won't see any output ...

In [ ]:
echo "bucket=${BUCKET}"
rm -rf babyweight_trained
export PYTHONPATH=${PYTHONPATH}:${PWD}/babyweight
python -m trainer.task \
  --bucket=${BUCKET} \
  --output_dir=babyweight_trained \
  --job-dir=./tmp \
  --pattern="00000-of-" --train_examples=1 --eval_steps=1

Lab Task 4

The JSON below represents an input into your prediction model. Write the input.json file below with the next cell, then run the prediction locally to assess whether it produces predictions correctly.

In [ ]:
%%writefile inputs.json
{"key": "b1", "is_male": "True", "mother_age": 26.0, "plurality": "Single(1)", "gestation_weeks": 39}
{"key": "g1", "is_male": "False", "mother_age": 26.0, "plurality": "Single(1)", "gestation_weeks": 39}
In [ ]:
sudo find "/usr/lib/google-cloud-sdk/lib/googlecloudsdk/command_lib/ml_engine" -name '*.pyc' -delete
In [ ]:
MODEL_LOCATION=$(ls -d $(pwd)/babyweight_trained/export/exporter/* | tail -1)
gcloud ai-platform local predict --model-dir=$MODEL_LOCATION --json-instances=inputs.json

Lab Task 5

Once the code works in standalone mode, you can run it on Cloud AI Platform. Because this is on the entire dataset, it will take a while. The training run took about two hours for me. You can monitor the job from the GCP console in the Cloud AI Platform section.

In [ ]:
JOBNAME=babyweight_$(date -u +%y%m%d_%H%M%S)
gsutil -m rm -rf $OUTDIR
gcloud ai-platform jobs submit training $JOBNAME \
  --region=$REGION \
  --module-name=trainer.task \
  --package-path=$(pwd)/babyweight/trainer \
  --job-dir=$OUTDIR \
  --staging-bucket=gs://$BUCKET \
  --scale-tier=STANDARD_1 \
  --runtime-version=$TFVERSION \
  -- \
  --bucket=${BUCKET} \
  --output_dir=${OUTDIR} \

When I ran it, I used train_examples=2000000. When training finished, I filtered in the Stackdriver log on the word "dict" and saw that the last line was:

Saving dict for global step 5714290: average_loss = 1.06473, global_step = 5714290, loss = 34882.4, rmse = 1.03186
The final RMSE was 1.03 pounds.

Hyperparameter tuning

All of these are command-line parameters to my program. To do hyperparameter tuning, create hyperparam.xml and pass it as --configFile. This step will take up to 2 hours -- you can increase maxParallelTrials or reduce maxTrials to get it done faster. Since maxParallelTrials is the number of initial seeds to start searching from, you don't want it to be too large; otherwise, all you have is a random search.

In [ ]:
%%writefile hyperparam.yaml
  scaleTier: STANDARD_1
    hyperparameterMetricTag: rmse
    goal: MINIMIZE
    maxTrials: 20
    maxParallelTrials: 5
    enableTrialEarlyStopping: True
    - parameterName: batch_size
      type: INTEGER
      minValue: 8
      maxValue: 512
      scaleType: UNIT_LOG_SCALE
    - parameterName: nembeds
      type: INTEGER
      minValue: 3
      maxValue: 30
      scaleType: UNIT_LINEAR_SCALE
    - parameterName: nnsize
      type: INTEGER
      minValue: 64
      maxValue: 512
      scaleType: UNIT_LOG_SCALE
In [ ]:
JOBNAME=babyweight_$(date -u +%y%m%d_%H%M%S)
gsutil -m rm -rf $OUTDIR
gcloud ai-platform jobs submit training $JOBNAME \
  --region=$REGION \
  --module-name=trainer.task \
  --package-path=$(pwd)/babyweight/trainer \
  --job-dir=$OUTDIR \
  --staging-bucket=gs://$BUCKET \
  --scale-tier=STANDARD_1 \
  --config=hyperparam.yaml \
  --runtime-version=$TFVERSION \
  -- \
  --bucket=${BUCKET} \
  --output_dir=${OUTDIR} \
  --eval_steps=10 \

Repeat training

This time with tuned parameters (note last line)

In [ ]:
JOBNAME=babyweight_$(date -u +%y%m%d_%H%M%S)
gsutil -m rm -rf $OUTDIR
gcloud ai-platform jobs submit training $JOBNAME \
  --region=$REGION \
  --module-name=trainer.task \
  --package-path=$(pwd)/babyweight/trainer \
  --job-dir=$OUTDIR \
  --staging-bucket=gs://$BUCKET \
  --scale-tier=STANDARD_1 \
  --runtime-version=$TFVERSION \
  -- \
  --bucket=${BUCKET} \
  --output_dir=${OUTDIR} \
  --train_examples=20000 --batch_size=35 --nembeds=16 --nnsize=281

Copyright 2017 Google Inc. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License