open selected rows with pandas using "chunksize" and/or "iterator"

Question

I have a large csv file and I open it with pd.read_csv as it follows:

df = pd.read_csv(path//fileName.csv, sep = ' ', header = None)

As the file is really large I would like to be able to open it in rows

from 0 to 511
from 512 to 1023
from 1024 to 1535
...
from 512*n to 512*(n+1) - 1

Where n = 1, 2, 3 ...

If I add chunksize = 512 into the arguments of read_csv

df = pd.read_csv(path//fileName.csv, sep = ' ', header = None, chunksize = 512)

and I type

df.get_chunk(5)

Than I am able to open rows from 0 to 5 or I may be able to divide the file in parts of 512 rows using a for loop

data = []
for chunks in df:
    data = data + [chunk]

But this is quite useless as still the file has to be completelly opened and takes time. How can I read only rows from 512*n to 512*(n+1).

Looking around I often saw that "chunksize" is used together with "iterator" as it follows

 df = pd.read_csv(path//fileName.csv, sep = ' ', header = None, iterator = True, chunksize = 512)

But after many attempts I still don't understand which benefits provide me this boolean variable. Could you explain me it, please?

Answer 1

How can I read only rows from 512*n to 512*(n+1)?

df = pd.read_csv(fn, header=None, skiprows=512*n, nrows=512)

You can do it this way (and it's pretty useful):

for chunk in pd.read_csv(f, sep = ' ', header = None, chunksize = 512):
    # process your chunk here

Demo:

In [61]: fn = 'd:/temp/a.csv'

In [62]: pd.DataFrame(np.random.randn(30, 3), columns=list('abc')).to_csv(fn, index=False)

In [63]: for chunk in pd.read_csv(fn, chunksize=10):
   ....:     print(chunk)
   ....:
          a         b         c
0  2.229657 -1.040086  1.295774
1  0.358098 -1.080557 -0.396338
2  0.731741 -0.690453  0.126648
3 -0.009388 -1.549381  0.913128
4 -0.256654 -0.073549 -0.171606
5  0.849934  0.305337  2.360101
6 -1.472184  0.641512 -1.301492
7 -2.302152  0.417787  0.485958
8  0.492314  0.603309  0.890524
9 -0.730400  0.835873  1.313114
          a         b         c
0  1.393865 -1.115267  1.194747
1  3.038719 -0.343875 -1.410834
2 -1.510598  0.664154 -0.996762
3 -0.528211  1.269363  0.506728
4  0.043785 -0.786499 -1.073502
5  1.096647 -1.127002  0.918172
6 -0.792251 -0.652996 -1.000921
7  1.582166 -0.819374  0.247077
8 -1.022418 -0.577469  0.097406
9 -0.274233 -0.244890 -0.352108
          a         b         c
0 -0.317418  0.774854 -0.203939
1  0.205443  0.820302 -2.637387
2  0.332696 -0.655431 -0.089120
3 -0.884916  0.274854  1.074991
4  0.412295 -1.561943 -0.850376
5 -1.933529 -1.346236 -1.789500
6  1.652446 -0.800644 -0.126594
7  0.520916 -0.825257 -0.475727
8 -2.261692  2.827894 -0.439698
9 -0.424714  1.862145  1.103926

In which case "iterator" can be useful?

when using chunksize - all chunks will have the same length. Using iterator parameter you can define how much data (get_chunk(nrows)) you want to read in each iteration:

In [66]: reader = pd.read_csv(fn, iterator=True)

let's read first 3 rows

In [67]: reader.get_chunk(3)
Out[67]:
          a         b         c
0  2.229657 -1.040086  1.295774
1  0.358098 -1.080557 -0.396338
2  0.731741 -0.690453  0.126648

now we'll read next 5 rows:

In [68]: reader.get_chunk(5)
Out[68]:
          a         b         c
0 -0.009388 -1.549381  0.913128
1 -0.256654 -0.073549 -0.171606
2  0.849934  0.305337  2.360101
3 -1.472184  0.641512 -1.301492
4 -2.302152  0.417787  0.485958

next 7 rows:

In [69]: reader.get_chunk(7)
Out[69]:
          a         b         c
0  0.492314  0.603309  0.890524
1 -0.730400  0.835873  1.313114
2  1.393865 -1.115267  1.194747
3  3.038719 -0.343875 -1.410834
4 -1.510598  0.664154 -0.996762
5 -0.528211  1.269363  0.506728
6  0.043785 -0.786499 -1.073502