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 [ ]:
%%bash
gcloud config set project $PROJECT
gcloud config set compute/region $REGION

In [ ]:
%%bash
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}
fi

In [ ]:
%%bash
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/task.py. You should use add hyperparameters needed by your model through the command-line using the parser module. Look at how batch_size is passed to the model in the code below. Do this for the following hyperparameters (defaults in parentheses): train_examples (5000), eval_steps (None), pattern (of).


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

from . import model

import tensorflow as tf
from tensorflow.contrib.learn.python.learn import learn_runner

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

    ## TODOs after this line
    ################################################################################
    
    ## TODO 1: add the new arguments here 

    ## 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(
        output_dir,
        json.loads(
            os.environ.get('TF_CONFIG', '{}')
        ).get('task', {}).get('trial', '')
    )

    # Run the training job
    model.train_and_evaluate(output_dir)

Lab Task 2

Address all the TODOs in the following code in babyweight/trainer/model.py with the cell below. This code is similar to the model training code we wrote in Lab 3.

After addressing all TODOs, run the cell to write the code to the model.py file.


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

tf.logging.set_verbosity(tf.logging.INFO)

BUCKET = None  # set from task.py
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
TRAIN_STEPS = 10000
EVAL_STEPS = None
BATCH_SIZE = 512
NEMBEDS = 3
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 = (tf.data.TextLineDataset(file_list)  # 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)
        else:
            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 = \
        [\
            tf.feature_column.categorical_column_with_vocabulary_list('is_male', 
                        ['True', 'False', 'Unknown']),
            tf.feature_column.numeric_column('mother_age'),
            tf.feature_column.categorical_column_with_vocabulary_list('plurality',
                        ['Single(1)', 'Twins(2)', 'Triplets(3)',
                         'Quadruplets(4)', 'Quintuplets(5)','Multiple(2+)']),
            tf.feature_column.numeric_column('gestation_weeks')
        ]

    # Discretize
    age_buckets = tf.feature_column.bucketized_column(mother_age, 
                        boundaries=np.arange(15,45,1).tolist())
    gestation_buckets = tf.feature_column.bucketized_column(gestation_weeks, 
                        boundaries=np.arange(17,47,1).tolist())
      
    # Sparse columns are wide, have a linear relationship with the output
    wide = [is_male,
            plurality,
            age_buckets,
            gestation_buckets]
    
    # 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,
            gestation_weeks,
            embed]
    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)}

## TODOs after this line
################################################################################

# 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 = None,
                                        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 = None,
        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, None),
        max_steps = None)
    
    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 = None,
        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 the small notebook VM. Change as appropriate for your model).

Even with smaller data, this might take 3-5 minutes in which you won't see any output ...


In [ ]:
%%bash
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 [ ]:
%%bash
sudo find "/usr/lib/google-cloud-sdk/lib/googlecloudsdk/command_lib/ml_engine" -name '*.pyc' -delete

In [ ]:
%%bash
MODEL_LOCATION=$(ls -d $(pwd)/babyweight_trained/export/exporter/* | tail -1)
echo $MODEL_LOCATION
gcloud ai-platform local predict --model-dir=$MODEL_LOCATION --json-instances=inputs.json

Lab Task 5

Once the code works in standalone mode, you can run it on Cloud AI Platform. Change the parameters to the model (-train_examples for example may not be part of your model) appropriately.

Because this is on the entire dataset, it will take a while. The training run took about 2 hours for me. You can monitor the job from the GCP console in the Cloud AI Platform section.


In [ ]:
%%bash
OUTDIR=gs://${BUCKET}/babyweight/trained_model
JOBNAME=babyweight_$(date -u +%y%m%d_%H%M%S)
echo $OUTDIR $REGION $JOBNAME
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=200000

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.

Optional: 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
trainingInput:
  scaleTier: STANDARD_1
  hyperparameters:
    hyperparameterMetricTag: rmse
    goal: MINIMIZE
    maxTrials: 20
    maxParallelTrials: 5
    enableTrialEarlyStopping: True
    params:
    - 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 [ ]:
%%bash
OUTDIR=gs://${BUCKET}/babyweight/hyperparam
JOBNAME=babyweight_$(date -u +%y%m%d_%H%M%S)
echo $OUTDIR $REGION $JOBNAME
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 \
  --train_examples=20000

Repeat training

This time with tuned parameters (note last line)


In [ ]:
%%bash
OUTDIR=gs://${BUCKET}/babyweight/trained_model_tuned
JOBNAME=babyweight_$(date -u +%y%m%d_%H%M%S)
echo $OUTDIR $REGION $JOBNAME
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 http://www.apache.org/licenses/LICENSE-2.0 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