Learning Objectives
In this notebook, we'll be using Keras to create a wide and deep model to predict the weight of a baby before it is born.
We'll start by defining the CSV column names, label column, and column defaults for our data inputs. Then, we'll construct a tf.data Dataset of features and the label from the CSV files and create inputs layers for the raw features. Next, we'll set up feature columns for the model inputs and build a wide and deep neural network in Keras. We'll create a custom evaluation metric and build our wide and deep model. Finally, we'll train and evaluate our model.
Each learning objective will correspond to a #TODO in this student lab notebook -- try to complete this notebook first and then review the solution notebook.
In [1]:
import datetime
import os
import shutil
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
print(tf.__version__)
In the seventh lab of this series 4a_sample_babyweight, we sampled from BigQuery our train, eval, and test CSV files. Verify that they exist, otherwise go back to that lab and create them.
In [2]:
%%bash
ls *.csv
In [1]:
%%bash
head -5 *.csv
Now that we have verified that our CSV files exist, we need to set a few things that we will be using in our input function.
CSV_COLUMNS are going to be our header names of our columns. Make sure that they are in the same order as in the CSV filesLABEL_COLUMN is the header name of the column that is our label. We will need to know this to pop it from our features dictionary.DEFAULTS is a list with the same length as CSV_COLUMNS, i.e. there is a default for each column in our CSVs. Each element is a list itself with the default value for that CSV column.
In [6]:
# Determine CSV, label, and key columns
CSV_COLUMNS = ["weight_pounds",
"is_male",
"mother_age",
"plurality",
"gestation_weeks"]
LABEL_COLUMN = "weight_pounds"
# Set default values for each CSV column.
# Treat is_male and plurality as strings.
DEFAULTS = [[0.0], ["null"], [0.0], ["null"], [0.0]]
Next, we will write an input_fn to read the data. Since we are reading from CSV files we can save ourself from trying to recreate the wheel and can use tf.data.experimental.make_csv_dataset. This will create a CSV dataset object. However we will need to divide the columns up into features and a label. We can do this by applying the map method to our dataset and popping our label column off of our dictionary of feature tensors.
In [7]:
def features_and_labels(row_data):
"""Splits features and labels from feature dictionary.
Args:
row_data: Dictionary of CSV column names and tensor values.
Returns:
Dictionary of feature tensors and label tensor.
"""
label = row_data.pop(LABEL_COLUMN)
return row_data, label # features, label
def load_dataset(pattern, batch_size=1, mode=tf.estimator.ModeKeys.EVAL):
"""Loads dataset using the tf.data API from CSV files.
Args:
pattern: str, file pattern to glob into list of files.
batch_size: int, the number of examples per batch.
mode: tf.estimator.ModeKeys to determine if training or evaluating.
Returns:
`Dataset` object.
"""
# Make a CSV dataset
dataset = tf.data.experimental.make_csv_dataset(
file_pattern=pattern,
batch_size=batch_size,
column_names=CSV_COLUMNS,
column_defaults=DEFAULTS)
# Map dataset to features and label
dataset = dataset.map(map_func=features_and_labels) # features, label
# Shuffle and repeat for training
if mode == tf.estimator.ModeKeys.TRAIN:
dataset = dataset.shuffle(buffer_size=1000).repeat()
# Take advantage of multi-threading; 1=AUTOTUNE
dataset = dataset.prefetch(buffer_size=1)
return dataset
We'll need to get the data read in by our input function to our model function, but just how do we go about connecting the dots? We can use Keras input layers (tf.Keras.layers.Input) by defining:
In [8]:
def create_input_layers():
"""Creates dictionary of input layers for each feature.
Returns:
Dictionary of `tf.Keras.layers.Input` layers for each feature.
"""
deep_inputs = {
colname: tf.keras.layers.Input(
name=colname, shape=(), dtype="float32")
for colname in ["mother_age", "gestation_weeks"]
}
wide_inputs = {
colname: tf.keras.layers.Input(
name=colname, shape=(), dtype="string")
for colname in ["is_male", "plurality"]
}
inputs = {**wide_inputs, **deep_inputs}
return inputs
In [9]:
def categorical_fc(name, values):
"""Helper function to wrap categorical feature by indicator column.
Args:
name: str, name of feature.
values: list, list of strings of categorical values.
Returns:
Categorical and indicator column of categorical feature.
"""
cat_column = tf.feature_column.categorical_column_with_vocabulary_list(
key=name, vocabulary_list=values)
ind_column = tf.feature_column.indicator_column(
categorical_column=cat_column)
return cat_column, ind_column
def create_feature_columns(nembeds):
"""Creates wide and deep dictionaries of feature columns from inputs.
Args:
nembeds: int, number of dimensions to embed categorical column down to.
Returns:
Wide and deep dictionaries of feature columns.
"""
deep_fc = {
colname: tf.feature_column.numeric_column(key=colname)
for colname in ["mother_age", "gestation_weeks"]
}
wide_fc = {}
is_male, wide_fc["is_male"] = categorical_fc(
"is_male", ["True", "False", "Unknown"])
plurality, wide_fc["plurality"] = categorical_fc(
"plurality", ["Single(1)", "Twins(2)", "Triplets(3)",
"Quadruplets(4)", "Quintuplets(5)", "Multiple(2+)"])
# Bucketize the float fields. This makes them wide
age_buckets = tf.feature_column.bucketized_column(
source_column=deep_fc["mother_age"],
boundaries=np.arange(15, 45, 1).tolist())
wide_fc["age_buckets"] = tf.feature_column.indicator_column(
categorical_column=age_buckets)
gestation_buckets = tf.feature_column.bucketized_column(
source_column=deep_fc["gestation_weeks"],
boundaries=np.arange(17, 47, 1).tolist())
wide_fc["gestation_buckets"] = tf.feature_column.indicator_column(
categorical_column=gestation_buckets)
# Cross all the wide columns, have to do the crossing before we one-hot
crossed = tf.feature_column.crossed_column(
keys=[age_buckets, gestation_buckets],
hash_bucket_size=1000)
deep_fc["crossed_embeds"] = tf.feature_column.embedding_column(
categorical_column=crossed, dimension=nembeds)
return wide_fc, deep_fc
So we've figured out how to get our inputs ready for machine learning but now we need to connect them to our desired output. Our model architecture is what links the two together. We need to create a wide and deep model now. The wide side will just be a linear regression or dense layer. For the deep side, let's create some hidden dense layers. All of this will end with a single dense output layer. This is regression so make sure the output layer activation is correct and that the shape is right.
In [10]:
def get_model_outputs(wide_inputs, deep_inputs, dnn_hidden_units):
"""Creates model architecture and returns outputs.
Args:
wide_inputs: Dense tensor used as inputs to wide side of model.
deep_inputs: Dense tensor used as inputs to deep side of model.
dnn_hidden_units: List of integers where length is number of hidden
layers and ith element is the number of neurons at ith layer.
Returns:
Dense tensor output from the model.
"""
# Hidden layers for the deep side
layers = [int(x) for x in dnn_hidden_units]
deep = deep_inputs
for layerno, numnodes in enumerate(layers):
deep = tf.keras.layers.Dense(
units=numnodes,
activation="relu",
name="dnn_{}".format(layerno+1))(deep)
deep_out = deep
# Linear model for the wide side
wide_out = tf.keras.layers.Dense(
units=10, activation="relu", name="linear")(wide_inputs)
# Concatenate the two sides
both = tf.keras.layers.concatenate(
inputs=[deep_out, wide_out], name="both")
# Final output is a linear activation because this is regression
output = tf.keras.layers.Dense(
units=1, activation="linear", name="weight")(both)
return output
We want to make sure that we have some useful way to measure model performance for us. Since this is regression, we would like to know the RMSE of the model on our evaluation dataset, however, this does not exist as a standard evaluation metric, so we'll have to create our own by using the true and predicted labels.
In [11]:
def rmse(y_true, y_pred):
"""Calculates RMSE evaluation metric.
Args:
y_true: tensor, true labels.
y_pred: tensor, predicted labels.
Returns:
Tensor with value of RMSE between true and predicted labels.
"""
return tf.sqrt(tf.reduce_mean((y_pred - y_true) ** 2))
Excellent! We've assembled all of the pieces, now we just need to tie them all together into a Keras Model. This is NOT a simple feedforward model with no branching, side inputs, etc. so we can't use Keras' Sequential Model API. We're instead going to use Keras' Functional Model API. Here we will build the model using tf.keras.models.Model giving our inputs and outputs and then compile our model with an optimizer, a loss function, and evaluation metrics.
In [13]:
def build_wide_deep_model(dnn_hidden_units=[64, 32], nembeds=3):
"""Builds wide and deep model using Keras Functional API.
Returns:
`tf.keras.models.Model` object.
"""
# Create input layers
inputs = create_input_layers()
# Create feature columns for both wide and deep
wide_fc, deep_fc = create_feature_columns(nembeds)
# The constructor for DenseFeatures takes a list of numeric columns
# The Functional API in Keras requires: LayerConstructor()(inputs)
wide_inputs = tf.keras.layers.DenseFeatures(
feature_columns=wide_fc.values(), name="wide_inputs")(inputs)
deep_inputs = tf.keras.layers.DenseFeatures(
feature_columns=deep_fc.values(), name="deep_inputs")(inputs)
# Get output of model given inputs
output = get_model_outputs(wide_inputs, deep_inputs, dnn_hidden_units)
# Build model and compile it all together
model = tf.keras.models.Model(inputs=inputs, outputs=output)
model.compile(optimizer="adam", loss="mse", metrics=[rmse, "mse"])
return model
print("Here is our wide and deep architecture so far:\n")
model = build_wide_deep_model()
print(model.summary())
We can visualize the wide and deep network using the Keras plot_model utility.
In [14]:
tf.keras.utils.plot_model(
model=model, to_file="wd_model.png", show_shapes=False, rankdir="LR")
Out[14]:
We've built our Keras model using our inputs from our CSV files and the architecture we designed. Let's now run our model by training our model parameters and periodically running an evaluation to track how well we are doing on outside data as training goes on. We'll need to load both our train and eval datasets and send those to our model through the fit method. Make sure you have the right pattern, batch size, and mode when loading the data. Also, don't forget to add the callback to TensorBoard.
In [15]:
TRAIN_BATCH_SIZE = 32
NUM_TRAIN_EXAMPLES = 10000 * 5 # training dataset repeats, it'll wrap around
NUM_EVALS = 5 # how many times to evaluate
# Enough to get a reasonable sample, but not so much that it slows down
NUM_EVAL_EXAMPLES = 10000
trainds = load_dataset(
pattern="train*",
batch_size=TRAIN_BATCH_SIZE,
mode=tf.estimator.ModeKeys.TRAIN)
evalds = load_dataset(
pattern="eval*",
batch_size=1000,
mode=tf.estimator.ModeKeys.EVAL).take(count=NUM_EVAL_EXAMPLES // 1000)
steps_per_epoch = NUM_TRAIN_EXAMPLES // (TRAIN_BATCH_SIZE * NUM_EVALS)
logdir = os.path.join(
"logs", datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=logdir, histogram_freq=1)
history = model.fit(
trainds,
validation_data=evalds,
epochs=NUM_EVALS,
steps_per_epoch=steps_per_epoch,
callbacks=[tensorboard_callback])
In [16]:
# Plot
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");
In [17]:
OUTPUT_DIR = "babyweight_trained_wd"
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(
obj=model, export_dir=EXPORT_PATH) # with default serving function
print("Exported trained model to {}".format(EXPORT_PATH))
In [ ]:
!ls $EXPORT_PATH
In this lab, we started by defining the CSV column names, label column, and column defaults for our data inputs. Then, we constructed a tf.data Dataset of features and the label from the CSV files and created inputs layers for the raw features. Next, we set up feature columns for the model inputs and built a wide and deep neural network in Keras. We created a custom evaluation metric and built our wide and deep model. Finally, we trained and evaluated our model.
Copyright 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