First BigQuery ML models for Taxifare Prediction

In this notebook, we will use BigQuery ML to build our first models for taxifare prediction.BigQuery ML provides a fast way to build ML models on large structured and semi-structured datasets.

Learning Objectives

Choose the correct BigQuery ML model type and specify options
Evaluate the performance of your ML model
Improve model performance through data quality cleanup
Create a Deep Neural Network (DNN) using SQL

Each learning objective will correspond to a #TODO in the student lab notebook -- try to complete that notebook first before reviewing this solution notebook.

We'll start by creating a dataset to hold all the models we create in BigQuery



In [ ]:

    
%%bash
export PROJECT=$(gcloud config list project --format "value(core.project)")
echo "Your current GCP Project Name is: "$PROJECT



In [10]:

    
import os

PROJECT = "qwiklabs-gcp-bdc77450c97b4bf6" # REPLACE WITH YOUR PROJECT NAME
REGION = "us-central1" # REPLACE WITH YOUR BUCKET REGION e.g. us-central1

# Do not change these
os.environ["PROJECT"] = PROJECT
os.environ["REGION"] = REGION
os.environ["BUCKET"] = PROJECT # DEFAULT BUCKET WILL BE PROJECT ID

if PROJECT == "your-gcp-project-here":
  print("Don't forget to update your PROJECT name! Currently:", PROJECT)

Create a BigQuery Dataset and Google Cloud Storage Bucket

A BigQuery dataset is a container for tables, views, and models built with BigQuery ML. Let's create one called serverlessml if we have not already done so in an earlier lab. We'll do the same for a GCS bucket for our project too.



In [11]:

    
%%bash

## Create a BigQuery dataset for serverlessml if it doesn't exist
datasetexists=$(bq ls -d | grep -w serverlessml)

if [ -n "$datasetexists" ]; then
    echo -e "BigQuery dataset already exists, let's not recreate it."

else
    echo "Creating BigQuery dataset titled: serverlessml"
    
    bq --location=US mk --dataset \
        --description 'Taxi Fare' \
        $PROJECT:serverlessml
   echo "\nHere are your current datasets:"
   bq ls
fi    
    
## Create GCS bucket if it doesn't exist already...
exists=$(gsutil ls -d | grep -w gs://${PROJECT}/)

if [ -n "$exists" ]; then
   echo -e "Bucket exists, let's not recreate it."
    
else
   echo "Creating a new GCS bucket."
   gsutil mb -l ${REGION} gs://${PROJECT}
   echo "\nHere are your current buckets:"
   gsutil ls
fi









    



BigQuery dataset already exists, let's not recreate it.
Bucket exists, let's not recreate it.

Model 1: Raw data

Let's build a model using just the raw data. It's not going to be very good, but sometimes it is good to actually experience this.

The model will take a minute or so to train. When it comes to ML, this is blazing fast.



In [12]:

    
%%bigquery
CREATE OR REPLACE MODEL serverlessml.model1_rawdata
OPTIONS(input_label_cols=['fare_amount'], model_type='linear_reg') AS

SELECT
  (tolls_amount + fare_amount) AS fare_amount,
  pickup_longitude AS pickuplon,
  pickup_latitude AS pickuplat,
  dropoff_longitude AS dropofflon,
  dropoff_latitude AS dropofflat,
  passenger_count*1.0 AS passengers
FROM `nyc-tlc.yellow.trips`
WHERE ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), 100000)) = 1









    Out[12]:

Once the training is done, visit the BigQuery Cloud Console and look at the model that has been trained. Then, come back to this notebook.

Note that BigQuery automatically split the data we gave it, and trained on only a part of the data and used the rest for evaluation. We can look at eval statistics on that held-out data:



In [13]:

    
%%bigquery
SELECT * FROM ML.EVALUATE(MODEL serverlessml.model1_rawdata)









    Out[13]:







  
    
      
      mean_absolute_error
      mean_squared_error
      mean_squared_log_error
      median_absolute_error
      r2_score
      explained_variance
    
  
  
    
      0
      6.086824
      93.909516
      0.32947
      4.501958
      0.00215
      0.002582

Let's report just the error we care about, the Root Mean Squared Error (RMSE)



In [14]:

    
%%bigquery
SELECT SQRT(mean_squared_error) AS rmse FROM ML.EVALUATE(MODEL serverlessml.model1_rawdata)

We told you it was not going to be good! Recall that our heuristic got 8.13, and our target is $6.

Note that the error is going to depend on the dataset that we evaluate it on. We can also evaluate the model on our own held-out benchmark/test dataset, but we shouldn't make a habit of this (we want to keep our benchmark dataset as the final evaluation, not make decisions using it all along the way. If we do that, our test dataset won't be truly independent).



In [15]:

    
%%bigquery
SELECT SQRT(mean_squared_error) AS rmse FROM ML.EVALUATE(MODEL serverlessml.model1_rawdata,
(
SELECT
  (tolls_amount + fare_amount) AS fare_amount,
  pickup_longitude AS pickuplon,
  pickup_latitude AS pickuplat,
  dropoff_longitude AS dropofflon,
  dropoff_latitude AS dropofflat,
  passenger_count*1.0 AS passengers
FROM `nyc-tlc.yellow.trips`
WHERE ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), 100000)) = 2
AND
  trip_distance > 0
  AND fare_amount >= 2.5
  AND pickup_longitude > -78
  AND pickup_longitude < -70
  AND dropoff_longitude > -78
  AND dropoff_longitude < -70
  AND pickup_latitude > 37
  AND pickup_latitude < 45
  AND dropoff_latitude > 37
  AND dropoff_latitude < 45
  AND passenger_count > 0
))

Model 2: Apply data cleanup

Recall that we did some data cleanup in the previous lab. Let's do those before training.

This is a dataset that we will need quite frequently in this notebook, so let's extract it first.



In [16]:

    
%%bigquery
CREATE OR REPLACE TABLE serverlessml.cleaned_training_data AS

SELECT
  (tolls_amount + fare_amount) AS fare_amount,
  pickup_longitude AS pickuplon,
  pickup_latitude AS pickuplat,
  dropoff_longitude AS dropofflon,
  dropoff_latitude AS dropofflat,
  passenger_count*1.0 AS passengers
FROM `nyc-tlc.yellow.trips`
WHERE ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), 100000)) = 1
AND
  trip_distance > 0
  AND fare_amount >= 2.5
  AND pickup_longitude > -78
  AND pickup_longitude < -70
  AND dropoff_longitude > -78
  AND dropoff_longitude < -70
  AND pickup_latitude > 37
  AND pickup_latitude < 45
  AND dropoff_latitude > 37
  AND dropoff_latitude < 45
  AND passenger_count > 0









    Out[16]:



In [17]:

    
%%bigquery
-- LIMIT 0 is a free query; this allows us to check that the table exists.
SELECT * FROM serverlessml.cleaned_training_data
LIMIT 0









    Out[17]:







  
    
      
      fare_amount
      pickuplon
      pickuplat
      dropofflon
      dropofflat
      passengers



In [18]:

    
%%bigquery
CREATE OR REPLACE MODEL serverlessml.model2_cleanup
OPTIONS(input_label_cols=['fare_amount'], model_type='linear_reg') AS

SELECT
*
FROM
serverlessml.cleaned_training_data









    Out[18]:



In [19]:

    
%%bigquery
SELECT SQRT(mean_squared_error) AS rmse FROM ML.EVALUATE(MODEL serverlessml.model2_cleanup)

Model 3: More sophisticated models

What if we try a more sophisticated model? Let's try Deep Neural Networks (DNNs) in BigQuery:

DNN

To create a DNN, simply specify dnn_regressor for the model_type and add your hidden layers.



In [20]:

    
%%bigquery
-- This model type is in alpha, so it may not work for you yet. This training takes on the order of 15 minutes.
CREATE OR REPLACE MODEL serverlessml.model3b_dnn
OPTIONS(input_label_cols=['fare_amount'], model_type='dnn_regressor', hidden_units=[32, 8]) AS

SELECT
*
FROM
serverlessml.cleaned_training_data









    Out[20]:



In [21]:

    
%%bigquery
SELECT SQRT(mean_squared_error) AS rmse FROM ML.EVALUATE(MODEL serverlessml.model3b_dnn)

Nice!

Evaluate DNN on benchmark dataset

Let's use the same validation dataset to evaluate -- remember that evaluation metrics depend on the dataset. You can not compare two models unless you have run them on the same withheld data.



In [22]:

    
%%bigquery
SELECT SQRT(mean_squared_error) AS rmse 
FROM ML.EVALUATE(MODEL serverlessml.model3b_dnn,
(
SELECT
  (tolls_amount + fare_amount) AS fare_amount,
  pickup_datetime,
  pickup_longitude AS pickuplon,
  pickup_latitude AS pickuplat,
  dropoff_longitude AS dropofflon,
  dropoff_latitude AS dropofflat,
  passenger_count*1.0 AS passengers,
  'unused' AS key
FROM `nyc-tlc.yellow.trips`
WHERE ABS(MOD(FARM_FINGERPRINT(CAST(pickup_datetime AS STRING)), 10000)) = 2
AND
  trip_distance > 0
  AND fare_amount >= 2.5
  AND pickup_longitude > -78
  AND pickup_longitude < -70
  AND dropoff_longitude > -78
  AND dropoff_longitude < -70
  AND pickup_latitude > 37
  AND pickup_latitude < 45
  AND dropoff_latitude > 37
  AND dropoff_latitude < 45
  AND passenger_count > 0
))

Wow! Later in this sequence of notebooks, we will get to below $4, but this is quite good, for very little work.

In this notebook, we showed you how to use BigQuery ML to quickly build ML models. We will come back to BigQuery ML when we want to experiment with different types of feature engineering. The speed of BigQuery ML is very attractive for development.

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.