I have a spark data frame which I want to divide into train, validation and test in the ratio 0.60, 0.20,0.20.
I used the following code for the same:
def data_split(x):
global data_map_var
d_map = data_map_var.value
data_row = x.asDict()
import random
rand = random.uniform(0.0,1.0)
ret_list = ()
if rand <= 0.6:
ret_list = (data_row['TRANS'] , d_map[data_row['ITEM']] , data_row['Ratings'] , 'train')
elif rand <=0.8:
ret_list = (data_row['TRANS'] , d_map[data_row['ITEM']] , data_row['Ratings'] , 'test')
else:
ret_list = (data_row['TRANS'] , d_map[data_row['ITEM']] , data_row['Ratings'] , 'validation')
return ret_list
split_sdf = ratings_sdf.map(data_split)
train_sdf = split_sdf.filter(lambda x : x[-1] == 'train').map(lambda x :(x[0],x[1],x[2]))
test_sdf = split_sdf.filter(lambda x : x[-1] == 'test').map(lambda x :(x[0],x[1],x[2]))
validation_sdf = split_sdf.filter(lambda x : x[-1] == 'validation').map(lambda x :(x[0],x[1],x[2]))
print "Total Records in Original Ratings RDD is {}".format(split_sdf.count())
print "Total Records in training data RDD is {}".format(train_sdf.count())
print "Total Records in validation data RDD is {}".format(validation_sdf.count())
print "Total Records in test data RDD is {}".format(test_sdf.count())
#help(ratings_sdf)
Total Records in Original Ratings RDD is 300001
Total Records in training data RDD is 180321
Total Records in validation data RDD is 59763
Total Records in test data RDD is 59837
My original data frame is ratings_sdf which I use to pass a mapper function which does the splitting.
If you check the total sum of train, validation and test does not sum to split (original ratings) count. And these numbers change at every run of the code.
Where is the remaining records going and why the sum is not equal?
TL;DR If you want to split DataFrame
use randomSplit
method:
ratings_sdf.randomSplit([0.6, 0.2, 0.2])
Your code is just wrong on multiple levels but there are two fundamental problems that make it broken beyond repair:
Spark transformations can be evaluated arbitrary number of times and functions you use should be referentially transparent and side effect free. Your code evaluates split_sdf
multiple times and you use stateful RNG data_split
so each time results are different.
This results in a behavior you describe where each child sees different state of the parent RDD.
You don't properly initialize RNG and in consequence random values you get are not independent.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With