Create Keras DNN model

Learning Objectives

  1. Use the tf.data API to create our ML datasets
  2. Use the Keras functional API to create our DNN architecture
  3. Train and evaluate the model and visualize the result
  4. Deploy a version of the model on Cloud AI Platform

Note: This notebook requires TensorFlow 2.0 as we are creating a model using Keras.

TODO: Complete the lab notebook #TODO sections. You can refer to the ../solutions/3_keras_dnn.ipynb notebook for reference.


In [ ]:
# change these to try this notebook out
BUCKET = 'cloud-training-demos-ml'
PROJECT = 'cloud-training-demos'
REGION = 'us-east1' #'us-central1'

In [ ]:
import os
os.environ['BUCKET'] = BUCKET
os.environ['PROJECT'] = PROJECT
os.environ['REGION'] = REGION

In [ ]:
%%bash
if ! gsutil ls | grep -q gs://${BUCKET}/; then
  gsutil mb -l ${REGION} gs://${BUCKET}
fi

In [ ]:
%%bash
ls *.csv

Create Keras model

First, write an input_fn to read the data.


In [ ]:
import shutil
import numpy as np
import tensorflow as tf
print(tf.__version__)

In [ ]:
# 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. Treat is_male and plurality as strings.
DEFAULTS = [[0.0], ['null'], [0.0], ['null'], [0.0], ['nokey']]

In [ ]:
# load the training data
# use the tf.data API to create and load a dataset. Be sure to map features and labels.
# https://www.tensorflow.org/api_docs/python/tf/data/experimental
def load_dataset(pattern, batch_size=1, mode=tf.estimator.ModeKeys.EVAL):
  dataset = (
              # TODO create the dataset
              )
 
  if mode == tf.estimator.ModeKeys.TRAIN:
        # TODO add shuffling to the dataset if it's in training mode:
        dataset = dataset
  dataset = dataset.prefetch(1) # take advantage of multi-threading; 1=AUTOTUNE
  return dataset

Next, define the feature columns. mother_age and gestation_weeks should be numeric. The others (is_male, plurality) should be categorical.


In [ ]:
## Build a simple Keras DNN using its Functional API
def rmse(y_true, y_pred):
    return tf.sqrt(tf.reduce_mean(tf.square(y_pred - y_true))) 

# Helper function to handle categorical columns
def categorical_fc(name, values):
  return tf.feature_column.indicator_column(
    tf.feature_column.categorical_column_with_vocabulary_list(name, values))

def build_dnn_model():
    # input layer
    inputs = {
        colname : tf.keras.layers.Input(name=colname, shape=(), dtype='float32')
           for colname in [] # TODO complete array of numeric input columns
    }
    inputs.update({
        colname : tf.keras.layers.Input(name=colname, shape=(), dtype='string')
            for colname in [] # TODO complete array of string input columns   
    })

    # feature columns from inputs
    feature_columns = {
        colname : tf.feature_column.numeric_column(colname)
            for colname in [] # TODO complete array of numeric feature columns
    }
    if False:
        # Until TF-serving supports 2.0, so as to get servable model
        feature_columns['is_male'] = categorical_fc('is_male', ['True', 'False', 'Unknown'])
        feature_columns['plurality'] = categorical_fc('plurality',
                          ['Single(1)', 'Twins(2)', 'Triplets(3)',
                           'Quadruplets(4)', 'Quintuplets(5)','Multiple(2+)'])

    # TODO: Use the feature columns and inputs above to create a DNN of [64, 32]

    

    model = tf.keras.models.Model(inputs, output)
    model.compile(optimizer='adam', loss='mse', metrics=[rmse, 'mse'])    
    return model

print("Here is our DNN architecture so far:\n")

# note how to use strategy to do distributed training
strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
    model = build_dnn_model()
print(model.summary())

We can visualize the DNN using the Keras plot_model utility.


In [ ]:
tf.keras.utils.plot_model(model, 'dnn_model.png', show_shapes=False, rankdir='LR')

Train and evaluate


In [ ]:
TRAIN_BATCH_SIZE = 32
NUM_TRAIN_EXAMPLES = 10000 * 5 # training dataset repeats, so it will wrap around
NUM_EVALS = 5  # how many times to evaluate
NUM_EVAL_EXAMPLES = 10000 # enough to get a reasonable sample, but not so much that it slows down

trainds = load_dataset('train*', TRAIN_BATCH_SIZE, tf.estimator.ModeKeys.TRAIN)
evalds = load_dataset('eval*', 1000, tf.estimator.ModeKeys.EVAL).take(NUM_EVAL_EXAMPLES//1000)

steps_per_epoch = NUM_TRAIN_EXAMPLES // (TRAIN_BATCH_SIZE * NUM_EVALS)

history = model.fit(trainds, 
                    validation_data=evalds,
                    epochs=NUM_EVALS, 
                    steps_per_epoch=steps_per_epoch)

Visualize loss curve


In [ ]:
# plot
import matplotlib.pyplot as plt
nrows = 1
ncols = 2
fig = plt.figure(figsize=(10, 5))

for idx, key in enumerate(['loss', 'rmse']):
    ax = fig.add_subplot(nrows, ncols, idx+1)
    plt.plot(history.history[key])
    plt.plot(history.history['val_{}'.format(key)])
    plt.title('model {}'.format(key))
    plt.ylabel(key)
    plt.xlabel('epoch')
    plt.legend(['train', 'validation'], loc='upper left');

Save the model

Let's wrap the model so that we can supply keyed predictions, and get the key back in our output


In [ ]:
# Serving function that passes through keys
@tf.function(input_signature=[{
      'is_male': tf.TensorSpec([None,], dtype=tf.string, name='is_male'),
      'mother_age': tf.TensorSpec([None,], dtype=tf.float32, name='mother_age'),
      'plurality': tf.TensorSpec([None,], dtype=tf.string, name='plurality'),
      'gestation_weeks': tf.TensorSpec([None,], dtype=tf.float32, name='gestation_weeks'),
      'key': tf.TensorSpec([None,], dtype=tf.string, name='key')
}])
def my_serve(inputs):
    feats = inputs.copy()
    key = feats.pop('key')
    output = model(feats)
    return {'key': key, 'babyweight': output}

In [ ]:
import shutil, os, datetime
OUTPUT_DIR = './export/babyweight'
shutil.rmtree(OUTPUT_DIR, ignore_errors=True)
EXPORT_PATH = os.path.join(OUTPUT_DIR, datetime.datetime.now().strftime('%Y%m%d%H%M%S'))
tf.saved_model.save(model, EXPORT_PATH, signatures={'serving_default': my_serve})
print("Exported trained model to {}".format(EXPORT_PATH))
os.environ['EXPORT_PATH'] = EXPORT_PATH

In [ ]:
!find $EXPORT_PATH

Monitor and experiment with training

To begin TensorBoard from within AI Platform Notebooks, click the + symbol in the top left corner and select the Tensorboard icon to create a new TensorBoard.

In TensorBoard, look at the learned embeddings. Are they getting clustered? How about the weights for the hidden layers? What if you run this longer? What happens if you change the batchsize?

Deploy trained model to Cloud AI Platform


In [ ]:
!saved_model_cli show --tag_set serve --signature_def serving_default --dir {EXPORT_PATH}

In [ ]:
%%bash
MODEL_NAME="babyweight"
VERSION_NAME="dnn"
MODEL_LOCATION=$EXPORT_PATH
echo "Deleting and deploying $MODEL_NAME $MODEL_VERSION from $MODEL_LOCATION ... this will take a few minutes"

if [[ $(gcloud ai-platform models list --format='value(name)' | grep $MODEL_NAME) ]]; then
    echo "The model named $MODEL_NAME already exists."
else
    # create model
    echo "Creating $MODEL_NAME model now."
    gcloud ai-platform models create --regions=$REGION $MODEL_NAME
fi

if [[ $(gcloud ai-platform versions list --model $MODEL_NAME --format='value(name)' | grep $VERSION_NAME) ]]; then
    echo "Deleting already the existing model $MODEL_NAME:$VERSION_NAME ... "
    gcloud ai-platform versions delete --model=$MODEL_NAME $VERSION_NAME
    echo "Please run this cell again if you don't see a Creating message ... "
    sleep 2
fi

# TODO create model on Cloud AI Platform. Use python-version 3.5 and runtime-version 1.14
# https://cloud.google.com/sdk/gcloud/reference/ai-platform/versions/create
echo "Creating $MODEL_NAME:$VERSION_NAME"
gcloud ai-platform versions create # TODO complete the statement

Monitor the model creation at GCP Console > AI Platform and once the model version dnn is created, proceed to the next cell.


In [ ]:
%%writefile input.json
{"key": "b1", "is_male": "True", "mother_age": 26.0, "plurality": "Single(1)", "gestation_weeks": 39}
{"key": "b2", "is_male": "True", "mother_age": 33.0, "plurality": "Single(1)", "gestation_weeks": 41}
{"key": "g1", "is_male": "False", "mother_age": 26.0, "plurality": "Single(1)", "gestation_weeks": 39}
{"key": "g2", "is_male": "False", "mother_age": 33.0, "plurality": "Single(1)", "gestation_weeks": 41}

In [ ]:
!gcloud ai-platform predict --model babyweight --json-instances input.json --version dnn

main.py

This is the code that exists in serving/application/main.py, i.e. the code in the web application that accesses the ML API.


In [ ]:
from oauth2client.client import GoogleCredentials
from googleapiclient import discovery

credentials = GoogleCredentials.get_application_default()
api = discovery.build('ml', 'v1', credentials=credentials)
project = PROJECT
model_name = 'babyweight'
version_name = 'dnn'

input_data = {
  'instances': [
    {
      'key': 'b1',
      'is_male': 'True',
      'mother_age': 26.0,
      'plurality': 'Single(1)',
      'gestation_weeks': 39
    },
    {
      'key': 'g1',
      'is_male': 'False',
      'mother_age': 29.0,
      'plurality': 'Single(1)',
      'gestation_weeks': 38
    },
    {
      'key': 'b2',
      'is_male': 'True',
      'mother_age': 26.0,
      'plurality': 'Triplets(3)',
      'gestation_weeks': 39
    },
    {
      'key': 'u1',
      'is_male': 'Unknown',
      'mother_age': 29.0,
      'plurality': 'Multiple(2+)',
      'gestation_weeks': 38
    },
  ]
}

parent = 'projects/%s/models/%s/versions/%s' % (project, model_name, version_name)
prediction = api.projects().predict(body=input_data, name=parent).execute()
print(prediction)
print(prediction['predictions'][0]['babyweight'][0])

Copyright 2017-2019 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