How to convert a pyspark dataframe column to numpy array

Question

I am trying to convert a pyspark dataframe column having approximately 90 million rows into a numpy array.

I need the array as an input for scipy.optimize.minimize function.

I have tried both converting to Pandas and using collect(), but these methods are very time consuming.

I am new to PySpark, If there is a faster and better approach to do this, Please help.

Thanks

This is how my dataframe looks like.

+----------+
|Adolescent|
+----------+
|       0.0|
|       0.0|
|       0.0|
|       0.0|
|       0.0|
|       0.0|
|       0.0|
|       0.0|
|       0.0|
|       0.0|
+----------+

Have you tried `df['Adolescent'].to_numpy()` or `df['Adolescent'].array`? — Nils Werner, Sep 30 '19 at 07:40
It looks like to_numpy() works for pandas data frame only and not pyspark. I tried df["Adolescent"].array which gives the output: "Column". I don't know how to use this as array. — Vaibhav Rathi, Sep 30 '19 at 08:59

pissall · Accepted Answer · 2019-09-30T09:58:14.733

26

#1

You will have to call a .collect() in any way. To create a numpy array from the pyspark dataframe, you can use:

adoles = np.array(df.select("Adolescent").collect()) #.reshape(-1) for 1-D array

#2

You can convert it to a pandas dataframe using toPandas(), and you can then convert it to numpy array using .values.

pdf = df.toPandas()
adoles = df["Adolescent"].values

Or simply:

adoles = df.select("Adolescent").toPandas().values #.reshape(-1) for 1-D array

#3

For distributed arrays, you can try Dask Arrays

I haven't tested this, but assuming it would work the same as numpy (might have inconsistencies):

import dask.array as da
adoles = da.array(df.select("Adolescent").collect()) #.reshape(-1) for 1-D array

edited Sep 30 '19 at 09:58

answered Sep 30 '19 at 08:01

pissall

7,109
2
25
45

I have tried using toPandas() but it is taking a lot of time. – Vaibhav Rathi Sep 30 '19 at 08:26
You should have a look at [Dask Arrays](https://docs.dask.org/en/latest/array.html) – pissall Sep 30 '19 at 09:52
Thanks for your help. I am trying with dask arrays now. – Vaibhav Rathi Oct 01 '19 at 06:18

score 1 · Answer 2 · answered Oct 25 '21 at 03:48

Another way is to convert the selected column to RDD, then flatten by extracting the value of each Row (can abuse .keys()), then convert to numpy array:

x = df.select("colname").rdd.map(lambda r: r[0]).collect()  # python list
np.array(x)  # numpy array

How to convert a pyspark dataframe column to numpy array

2 Answers2

#1

#2

#3

Linked