Learning Objectives
In this lab, you will be using the US Centers for Disease Control and Prevention's (CDC) natality data to build a model to predict baby birth weights based on a handful of features known at pregnancy. Because we're predicting a continuous value, this is a regression problem, and for that, we'll use the linear regression model built into BQML.
In [ ]:
import matplotlib.pyplot as plt
VERY IMPORTANT: In the cell below you must replace the text <YOUR PROJECT> with your GCP project id as provided during the setup of your environment. Please leave any surrounding single quotes in place.
In [ ]:
PROJECT = '<YOUR PROJECT>' #TODO Replace with your GCP PROJECT
This lab will use natality data and training on features to predict the birth weight.
The CDC's Natality data has details on US births from 1969 to 2008 and is available in BigQuery as a public data set. More details: https://bigquery.cloud.google.com/table/publicdata:samples.natality?tab=details
Start by looking at the data since 2000 with useful values, those greater than 0.
Note: "%%bigquery" is a magic which allows quick access to BigQuery from within a notebook.
In [ ]:
%%bigquery
SELECT
*
FROM
publicdata.samples.natality
WHERE
year > 2000
AND gestation_weeks > 0
AND mother_age > 0
AND plurality > 0
AND weight_pounds > 0
LIMIT 10
Looking over the data set, there are a few columns of interest that could be leveraged into features for a reasonable prediction of approximate birth weight.
Further, some feature engineering may be accomplished with the BigQuery CAST function -- in BQML, all strings are considered categorical features and all numeric types are considered continuous ones.
The hashmonth is added so that we can repeatably split the data without leakage -- the goal is to have all babies that share a birthday to be either in training set or in test set and not spread between them (otherwise, there would be information leakage when it comes to triplets, etc.)
In [ ]:
%%bigquery
SELECT
weight_pounds, -- this is the label; because it is continuous, we need to use regression
CAST(is_male AS STRING) AS is_male,
mother_age,
CAST(plurality AS STRING) AS plurality,
gestation_weeks,
FARM_FINGERPRINT(CONCAT(CAST(YEAR AS STRING), CAST(month AS STRING))) AS hashmonth
FROM
publicdata.samples.natality
WHERE
year > 2000
AND gestation_weeks > 0
AND mother_age > 0
AND plurality > 0
AND weight_pounds > 0
LIMIT 10
In [ ]:
%%bash
bq --location=US mk -d demo
With the demo dataset ready, it is possible to create a linear regression model to train the model.
This will take approximately 5 to 7 minutes to run. Feedback from BigQuery will cease in output cell and the notebook will leave the "busy" state when complete.
In [ ]:
%%bigquery
CREATE or REPLACE MODEL demo.babyweight_model_asis
OPTIONS
(model_type='linear_reg', labels=['weight_pounds'], optimize_strategy='batch_gradient_descent') AS
WITH natality_data AS (
SELECT
weight_pounds,-- this is the label; because it is continuous, we need to use regression
CAST(is_male AS STRING) AS is_male,
mother_age,
CAST(plurality AS STRING) AS plurality,
gestation_weeks,
FARM_FINGERPRINT(CONCAT(CAST(YEAR AS STRING), CAST(month AS STRING))) AS hashmonth
FROM
publicdata.samples.natality
WHERE
year > 2000
AND gestation_weeks > 0
AND mother_age > 0
AND plurality > 0
AND weight_pounds > 0
)
SELECT
weight_pounds,
is_male,
mother_age,
plurality,
gestation_weeks
FROM
natality_data
WHERE
ABS(MOD(hashmonth, 4)) < 3 -- select 75% of the data as training
For all training runs, statistics are captured in the "TRAINING_INFO" table. This table has basic performance statistics for each iteration.
The query below returns the training details.
In [ ]:
%%bigquery
SELECT * FROM ML.TRAINING_INFO(MODEL demo.babyweight_model_asis);
Some of these columns are obvious although what do the non-specific ML columns mean (specific to BQML)?
training_run - Will be zero for a newly created model. If the model is re-trained using warm_start, this will increment for each re-training.
iteration - Number of the associated training_run, starting with zero for the first iteration.
duration_ms - Indicates how long the iteration took (in ms).
Next plot the training and evaluation loss to see if the model has an overfit.
In [ ]:
%%bigquery history
SELECT * FROM ML.TRAINING_INFO(MODEL demo.babyweight_model_asis)
In [ ]:
history
In [ ]:
plt.plot('iteration', 'loss', data=history,
marker='o', color='orange', linewidth=2)
plt.plot('iteration', 'eval_loss', data=history,
marker='', color='green', linewidth=2, linestyle='dashed')
plt.xlabel('iteration')
plt.ylabel('loss')
plt.legend();
As you can see, the training loss and evaluation loss are essentially identical. There does not appear to be any overfitting.
With a trained model, it is now possible to make a prediction on the values. The only difference from the second query above is the reference to the model. The data has been limited (LIMIT 100) to reduce amount of data returned.
When the ml.predict function is leveraged, output prediction column name for the model is predicted_<label_column_name>.
In [ ]:
%%bigquery
SELECT
*
FROM
ml.PREDICT(MODEL demo.babyweight_model_asis,
(SELECT
weight_pounds,
CAST(is_male AS STRING) AS is_male,
mother_age,
CAST(plurality AS STRING) AS plurality,
gestation_weeks
FROM
publicdata.samples.natality
WHERE
year > 2000
AND gestation_weeks > 0
AND mother_age > 0
AND plurality > 0
AND weight_pounds > 0
))
LIMIT 100
Copyright 2018 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