Learning Objectives
In the last notebook, we learned about the Keras Sequential API. The Keras Functional API provides an alternate way of building models which is more flexible. With the Functional API, we can build models with more complex topologies, multiple input or output layers, shared layers or non-sequential data flows (e.g. residual layers).
In this notebook we'll use what we learned about feature columns to build a Wide & Deep model. Recall, that the idea behind Wide & Deep models is to join the two methods of learning through memorization and generalization by making a wide linear model and a deep learning model to accommodate both. You can have a look at the original research paper here: Wide & Deep Learning for Recommender Systems.
The Wide part of the model is associated with the memory element. In this case, we train a linear model with a wide set of crossed features and learn the correlation of this related data with the assigned label. The Deep part of the model is associated with the generalization element where we use embedding vectors for features. The best embeddings are then learned through the training process. While both of these methods can work well alone, Wide & Deep models excel by combining these techniques together.
In [ ]:
!sudo chown -R jupyter:jupyter /home/jupyter/training-data-analyst
In [ ]:
# Ensure the right version of Tensorflow is installed.
!pip freeze | grep tensorflow==2.1 || pip install tensorflow==2.1
Start by importing the necessary libraries for this lab.
In [ ]:
import datetime
import os
import shutil
import numpy as np
import pandas as pd
import tensorflow as tf
from matplotlib import pyplot as plt
from tensorflow import keras
from tensorflow import feature_column as fc
from tensorflow.keras import Model
from tensorflow.keras.layers import (
Input, Dense, DenseFeatures, concatenate)
from tensorflow.keras.callbacks import TensorBoard
print(tf.__version__)
In [ ]:
%matplotlib inline
In [ ]:
!ls -l ../data/*.csv
We wrote these functions for reading data from the csv files above in the previous notebook. For this lab we will also include some additional engineered features in our model. In particular, we will compute the difference in latitude and longitude, as well as the Euclidean distance between the pick-up and drop-off locations. We can accomplish this by adding these new features to the features dictionary with the function add_engineered_features below.
Note that we include a call to this function when collecting our features dict and labels in the features_and_labels function below as well.
In [ ]:
CSV_COLUMNS = [
'fare_amount',
'pickup_datetime',
'pickup_longitude',
'pickup_latitude',
'dropoff_longitude',
'dropoff_latitude',
'passenger_count',
'key'
]
LABEL_COLUMN = 'fare_amount'
DEFAULTS = [[0.0], ['na'], [0.0], [0.0], [0.0], [0.0], [0.0], ['na']]
UNWANTED_COLS = ['pickup_datetime', 'key']
def features_and_labels(row_data):
label = row_data.pop(LABEL_COLUMN)
features = row_data
for unwanted_col in UNWANTED_COLS:
features.pop(unwanted_col)
return features, label
def create_dataset(pattern, batch_size=1, mode='eval'):
dataset = tf.data.experimental.make_csv_dataset(
pattern, batch_size, CSV_COLUMNS, DEFAULTS)
dataset = dataset.map(features_and_labels)
if mode == 'train':
dataset = dataset.shuffle(buffer_size=1000).repeat()
# take advantage of multi-threading; 1=AUTOTUNE
dataset = dataset.prefetch(1)
return dataset
For the Wide columns, we will create feature columns of crossed features. To do this, we'll create a collection of Tensorflow feature columns to pass to the tf.feature_column.crossed_column constructor. The Deep columns will consist of numeric columns and the embedding columns we want to create.
In [ ]:
# TODO 1
# 1. Bucketize latitudes and longitudes
NBUCKETS = 16
latbuckets = np.linspace(start=38.0, stop=42.0, num=NBUCKETS).tolist()
lonbuckets = np.linspace(start=-76.0, stop=-72.0, num=NBUCKETS).tolist()
fc_bucketized_plat = fc.bucketized_column(
source_column=fc.numeric_column("pickup_longitude"), boundaries=lonbuckets)
fc_bucketized_plon = fc.bucketized_column(
source_column=fc.numeric_column("pickup_latitude"), boundaries=latbuckets)
fc_bucketized_dlat = fc.bucketized_column(
source_column=fc.numeric_column("dropoff_longitude"), boundaries=lonbuckets)
fc_bucketized_dlon = fc.bucketized_column(
source_column=fc.numeric_column("dropoff_latitude"), boundaries=latbuckets)
# 2. Cross features for locations
fc_crossed_dloc = fc.crossed_column(
[fc_bucketized_dlat, fc_bucketized_dlon],
hash_bucket_size=NBUCKETS * NBUCKETS)
fc_crossed_ploc = fc.crossed_column(
[fc_bucketized_plat, fc_bucketized_plon],
hash_bucket_size=NBUCKETS * NBUCKETS)
fc_crossed_pd_pair = fc.crossed_column(
[fc_crossed_dloc, fc_crossed_ploc],
hash_bucket_size=NBUCKETS**4)
# 3. Create embedding columns for the crossed columns
fc_pd_pair = fc.embedding_column(categorical_column=fc_crossed_pd_pair, dimension=3)
fc_dloc = fc.embedding_column(categorical_column=fc_crossed_dloc, dimension=3)
fc_ploc = fc.embedding_column(categorical_column=fc_crossed_ploc, dimension=3)
Next we gather the list of wide and deep feature columns we'll pass to our Wide & Deep model in Tensorflow. Recall, wide columns are sparse, have linear relationship with the output while continuous columns are deep, have a complex relationship with the output. We will use our previously bucketized columns to collect crossed feature columns and sparse feature columns for our wide columns, and embedding feature columns and numeric features columns for the deep columns.
In [ ]:
# TODO 2
wide_columns = [
# One-hot encoded feature crosses
fc.indicator_column(fc_crossed_dloc),
fc.indicator_column(fc_crossed_ploc),
fc.indicator_column(fc_crossed_pd_pair)
]
deep_columns = [
# Embedding_column to "group" together ...
fc.embedding_column(fc_crossed_pd_pair, dimension=10),
# Numeric columns
fc.numeric_column("pickup_latitude"),
fc.numeric_column("pickup_longitude"),
fc.numeric_column("dropoff_longitude"),
fc.numeric_column("dropoff_latitude")
]
To build a wide-and-deep network, we connect the sparse (i.e. wide) features directly to the output node, but pass the dense (i.e. deep) features through a set of fully connected layers. Here’s that model architecture looks using the Functional API.
First, we'll create our input columns using tf.keras.layers.Input.
In [ ]:
INPUT_COLS = [
'pickup_longitude',
'pickup_latitude',
'dropoff_longitude',
'dropoff_latitude',
'passenger_count'
]
inputs = {colname : Input(name=colname, shape=(), dtype='float32')
for colname in INPUT_COLS
}
Then, we'll define our custom RMSE evaluation metric and build our wide and deep model.
In [ ]:
def rmse(y_true, y_pred):
return tf.sqrt(tf.reduce_mean(tf.square(y_pred - y_true)))
# TODO 3
def build_model(dnn_hidden_units):
# Create the deep part of model
deep = DenseFeatures(deep_columns, name='deep_inputs')(inputs)
for num_nodes in dnn_hidden_units:
deep = Dense(num_nodes, activation='relu')(deep)
# Create the wide part of model
wide = DenseFeatures(wide_columns, name='wide_inputs')(inputs)
# Combine deep and wide parts of the model
combined = concatenate(inputs=[deep, wide], name='combined')
# Map the combined outputs into a single prediction value
output = Dense(units=1, activation=None, name='prediction')(combined)
# Finalize the model
model = Model(inputs=list(inputs.values()), outputs=output)
# Compile the keras model
model.compile(optimizer="adam", loss="mse", metrics=[rmse, "mse"])
return model
Next, we can call the build_model to create the model. Here we'll have two hidden layers, each with 10 neurons, for the deep part of our model. We can also use plot_model to see a diagram of the model we've created.
In [ ]:
HIDDEN_UNITS = [10,10]
model = build_model(dnn_hidden_units=HIDDEN_UNITS)
tf.keras.utils.plot_model(model, show_shapes=False, rankdir='LR')
Next, we'll set up our training variables, create our datasets for training and validation, and train our model.
(We refer you the the blog post ML Design Pattern #3: Virtual Epochs for further details on why express the training in terms of NUM_TRAIN_EXAMPLES and NUM_EVALS and why, in this training code, the number of epochs is really equal to the number of evaluations we perform.)
In [ ]:
BATCH_SIZE = 1000
NUM_TRAIN_EXAMPLES = 10000 * 5 # training dataset will repeat, wrap around
NUM_EVALS = 50 # how many times to evaluate
NUM_EVAL_EXAMPLES = 10000 # enough to get a reasonable sample
trainds = create_dataset(
pattern='../data/taxi-train*',
batch_size=BATCH_SIZE,
mode='train')
evalds = create_dataset(
pattern='../data/taxi-valid*',
batch_size=BATCH_SIZE,
mode='eval').take(NUM_EVAL_EXAMPLES//1000)
In [ ]:
%%time
steps_per_epoch = NUM_TRAIN_EXAMPLES // (BATCH_SIZE * NUM_EVALS)
OUTDIR = "./taxi_trained"
shutil.rmtree(path=OUTDIR, ignore_errors=True) # start fresh each time
history = model.fit(x=trainds,
steps_per_epoch=steps_per_epoch,
epochs=NUM_EVALS,
validation_data=evalds,
callbacks=[TensorBoard(OUTDIR)])
Just as before, we can examine the history to see how the RMSE changes through training on the train set and validation set.
In [ ]:
RMSE_COLS = ['rmse', 'val_rmse']
pd.DataFrame(history.history)[RMSE_COLS].plot()
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