pandas 0.18: out of memory error when reading CSV file with categoricals

Question

I am trying to read 3GB file (2.5 million rows, mostly categorical (string) data) into Pandas dataframe with read_csv function and get error: out of memory

• I am on PC with Pandas 0.18 version and 16GB of RAM, so 3GB data should easily fit on 16GB. (Update: This is not a duplicate question)
• I know that I can provide dtype to improve reading the CSV, but there are too many columns in my data set and I want to load it first, then decide on data type.

The Traceback is:

Traceback (most recent call last):
  File "/home/a/Dropbox/Programming/Python/C and d/main.com.py", line 9, in <module>
    preprocessing()
  File "/home/a/Dropbox/Programming/Python/C and d/main.com.py", line 5, in preprocessing
    df = pd.read_csv(filepath_or_buffer = file_path, sep ='\t', low_memory = False)
  File "/usr/lib/python3/dist-packages/pandas/io/parsers.py", line 498, in parser_f
    return _read(filepath_or_buffer, kwds)
  File "/usr/lib/python3/dist-packages/pandas/io/parsers.py", line 285, in _read
    return parser.read()
  File "/usr/lib/python3/dist-packages/pandas/io/parsers.py", line 747, in read
    ret = self._engine.read(nrows)
  File "/usr/lib/python3/dist-packages/pandas/io/parsers.py", line 1197, in read
    data = self._reader.read(nrows)
  File "pandas/parser.pyx", line 769, in pandas.parser.TextReader.read (pandas/parser.c:8011)
  File "pandas/parser.pyx", line 857, in pandas.parser.TextReader._read_rows (pandas/parser.c:9140)
  File "pandas/parser.pyx", line 1833, in pandas.parser.raise_parser_error (pandas/parser.c:22649)
pandas.parser.CParserError: Error tokenizing data. C error: out of memory

My code:

import pandas as pd
def preprocessing():
    file_path = r'/home/a/Downloads/main_query.txt'
    df = pd.read_csv(filepath_or_buffer = file_path, sep ='\t', low_memory = False)

The above code produced error message, which I posted above.

I then tried to remove low_memory = False, and everything worked, it only gave warning:

sys:1: DtypeWarning: Columns (17,20,23,24,33,44,58,118,134,
135,137,142,145,146,147) have mixed types.
Specify dtype option on import or set low_memory=False.

Answer 1

UPDATE: in Pandas 0.19.0 it should be possible to specify categorical dtype when using read_csv() method:

pd.read_csv(filename, dtype={'col1': 'category'})

so you may try to use pandas 0.19.0 RC1

OLD answer:

you can read your CSV in chunks and concatenate it to the resulting DF on each step:

chunksize = 10**5
df = pd.DataFrame()

for chunk in (pd.read_csv(filename,
                          dtype={'col1':np.int8, 'col2':np.int32, ...}
                          chunksize=chunksize)
             ):
    df = pd.concat([df, chunk], ignore_index=True)

NOTE: parameter dtype is unsupported with engine=’python’

Answer 2

The question is a duplicate. When you do read_csv but don't specify dtypes, if you read in floats, ints, dates and categoricals as (unique) strings, you can easily use up Gigabytes. So take a little time to specify dtypes.

1. categoricals read in and stored as string (as opposed to categorical) take tons of memory.

• (pandas will underreport memory usage for dataframes with strings, unless you use df.info(memory_usage='deep') or df.memory_usage(deep=True))

1. As of pandas 0.19, you now don't need to specify each Categorical variable's levels. Just do pd.read_csv(..., dtype={'foo': 'category', 'bar': 'category', ...})

• 0.19 whatsnew: read_csv supports parsing Categorical directly

1. That should solve everything. In the extremely unlikely event you still run out of memory, then also debug like this:

• only read in a subset of columns, say usecols = ['foo', 'bar', 'baz']
• only read in a subset of rows (say nrows=1e5 or see also skiprows=...)
• and iteratively figure out each categorical's levels and how much memory it uses. You don't need to read in all rows or columns to figure out one categorical column's levels.