Running Validations¶

A validation consists in evaluating same entity (e.g., a model or a feature) on two different input datasets. We can validate model output for registered models in the Notebook. This is done by calling create_data function of corridor library

In [1]:

# import spark library
import findspark; findspark.init(); import pyspark
from pyspark.sql.functions import *
spark = pyspark.sql.SparkSession.builder.getOrCreate()

# read in the Application Level Dataset
application_table = spark.read.parquet('s3a://corridor.dev/master/sampleAppData.parquet')

# create 'application_table_1' as applications before 2015
# application_table_1 = application_table.limit(1000)
application_table_1= application_table.filter(year(application_table.corridor_application_date)<2015).limit(1000)


# create 'application_table_2' as applications in or after 2015
application_table_2= application_table.filter(year(application_table.corridor_application_date)>=2015).limit(1000)

# take a look at the first 5 rows of the application_table dataframe by runningthe line below:
# application_table_1.limit(5).toPandas()

# CHecking if two datasets application_table_1 and application_table_2 have identical schema
print(f'Rows in application_table_1: {application_table_1.count()}') 
print(f'Rows in application_table_2: {application_table_2.count()}')

# import spark library import findspark; findspark.init(); import pyspark from pyspark.sql.functions import * spark = pyspark.sql.SparkSession.builder.getOrCreate() # read in the Application Level Dataset application_table = spark.read.parquet('s3a://corridor.dev/master/sampleAppData.parquet') # create 'application_table_1' as applications before 2015 # application_table_1 = application_table.limit(1000) application_table_1= application_table.filter(year(application_table.corridor_application_date)<2015).limit(1000) # create 'application_table_2' as applications in or after 2015 application_table_2= application_table.filter(year(application_table.corridor_application_date)>=2015).limit(1000) # take a look at the first 5 rows of the application_table dataframe by runningthe line below: # application_table_1.limit(5).toPandas() # CHecking if two datasets application_table_1 and application_table_2 have identical schema print(f'Rows in application_table_1: {application_table_1.count()}') print(f'Rows in application_table_2: {application_table_2.count()}')

Rows in application_table_1: 22
Rows in application_table_2: 61

In [2]:

# Import Corridor Package Objects
from corridor import create_data

# Run simulation for first dataset - application_table_1: 
#    PD Model Strict: pd_model_ver1 is output feature alias

df_1 = create_data('pd_model_ver1',
                 data={'application': application_table_1})
df_1.describe().show()

# Import Corridor Package Objects from corridor import create_data # Run simulation for first dataset - application_table_1: # PD Model Strict: pd_model_ver1 is output feature alias df_1 = create_data('pd_model_ver1', data={'application': application_table_1}) df_1.describe().show()

+-------+-------------------+
|summary|      pd_model_ver1|
+-------+-------------------+
|  count|                 22|
|   mean|0.16363636363636364|
| stddev|0.06931951244198711|
|    min|                0.1|
|    max|               0.25|
+-------+-------------------+

In [3]:

# Run simulation for second dataset - application_table_2: 
#    PD Model Strict: pd_model_ver1 is output feature alias

df_2 = create_data('pd_model_ver1',
                 data={'application': application_table_2})
df_2.describe().show()

# Run simulation for second dataset - application_table_2: # PD Model Strict: pd_model_ver1 is output feature alias df_2 = create_data('pd_model_ver1', data={'application': application_table_2}) df_2.describe().show()

+-------+--------------------+
|summary|       pd_model_ver1|
+-------+--------------------+
|  count|                  61|
|   mean| 0.16803278688524592|
| stddev|0.061259548157304895|
|    min|                0.05|
|    max|                0.25|
+-------+--------------------+