Memory error when using pandas read_csv

I am trying to do something fairly simple, reading a large csv file into a pandas dataframe.

data = pandas.read_csv(filepath, header = 0, sep = DELIMITER,skiprows = 2)

The code either fails with a MemoryError, or just never finishes.

Mem usage in the task manager stopped at 506 Mb and after 5 minutes of no change and no CPU activity in the process I stopped it.

I am using pandas version 0.11.0.

I am aware that there used to be a memory problem with the file parser, but according to http://wesmckinney.com/blog/?p=543 this should have been fixed.

The file I am trying to read is 366 Mb, the code above works if I cut the file down to something short (25 Mb).

It has also happened that I get a pop up telling me that it can't write to address 0x1e0baf93...

Stacktrace:

Traceback (most recent call last):
  File "F:\QA ALM\Python\new WIM data\new WIM data\new_WIM_data.py", line 25, in
 <module>
    wimdata = pandas.read_csv(filepath, header = 0, sep = DELIMITER,skiprows = 2
)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\io\parsers.py"
, line 401, in parser_f
    return _read(filepath_or_buffer, kwds)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\io\parsers.py"
, line 216, in _read
    return parser.read()
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\io\parsers.py"
, line 643, in read
    df = DataFrame(col_dict, columns=columns, index=index)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\core\frame.py"
, line 394, in __init__
    mgr = self._init_dict(data, index, columns, dtype=dtype)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\core\frame.py"
, line 525, in _init_dict
    dtype=dtype)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\core\frame.py"
, line 5338, in _arrays_to_mgr
    return create_block_manager_from_arrays(arrays, arr_names, axes)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\core\internals
.py", line 1820, in create_block_manager_from_arrays
    blocks = form_blocks(arrays, names, axes)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\core\internals
.py", line 1872, in form_blocks
    float_blocks = _multi_blockify(float_items, items)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\core\internals
.py", line 1930, in _multi_blockify
    block_items, values = _stack_arrays(list(tup_block), ref_items, dtype)
  File "C:\Program Files\Python\Anaconda\lib\site-packages\pandas\core\internals
.py", line 1962, in _stack_arrays
    stacked = np.empty(shape, dtype=dtype)
MemoryError
Press any key to continue . . .

A bit of background - I am trying to convince people that Python can do the same as R. For this I am trying to replicate an R script that does

data <- read.table(paste(INPUTDIR,config[i,]$TOEXTRACT,sep=""), HASHEADER, DELIMITER,skip=2,fill=TRUE)

R not only manages to read the above file just fine, it even reads several of these files in a for loop (and then does some stuff with the data). If Python does have a problem with files of that size I might be fighting a loosing battle...

Windows memory limitation

Memory errors happens a lot with python when using the 32bit version in Windows. This is because 32bit processes only gets 2GB of memory to play with by default.

Tricks for lowering memory usage

If you are not using 32bit python in windows but are looking to improve on your memory efficiency while reading csv files, there is a trick.

The pandas.read_csv function takes an option called dtype. This lets pandas know what types exist inside your csv data.

How this works

By default, pandas will try to guess what dtypes your csv file has. This is a very heavy operation because while it is determining the dtype, it has to keep all raw data as objects (strings) in memory.

Example

Let's say your csv looks like this:

name, age, birthday
Alice, 30, 1985-01-01
Bob, 35, 1980-01-01
Charlie, 25, 1990-01-01

This example is of course no problem to read into memory, but it's just an example.

If pandas were to read the above csv file without any dtype option, the age would be stored as strings in memory until pandas has read enough lines of the csv file to make a qualified guess.

I think the default in pandas is to read 1,000,000 rows before guessing the dtype.

Solution

By specifying dtype={'age':int} as an option to the .read_csv() will let pandas know that age should be interpreted as a number. This saves you lots of memory.

Problem with corrupt data

However, if your csv file would be corrupted, like this:

name, age, birthday
Alice, 30, 1985-01-01
Bob, 35, 1980-01-01
Charlie, 25, 1990-01-01
Dennis, 40+, None-Ur-Bz

Then specifying dtype={'age':int} will break the .read_csv() command, because it cannot cast "40+" to int. So sanitize your data carefully!

Here you can see how the memory usage of a pandas dataframe is a lot higher when floats are kept as strings:

Try it yourself

df = pd.DataFrame(pd.np.random.choice(['1.0', '0.6666667', '150000.1'],(100000, 10)))
resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
# 224544 (~224 MB)

df = pd.DataFrame(pd.np.random.choice([1.0, 0.6666667, 150000.1],(100000, 10)))
resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
# 79560 (~79 MB)

I had the same memory problem with a simple read of a tab delimited text file around 1 GB in size (over 5.5 million records) and this solved the memory problem：

df = pd.read_csv(myfile,sep='\t') # didn't work, memory error
df = pd.read_csv(myfile,sep='\t',low_memory=False) # worked fine and in less than 30 seconds

Spyder 3.2.3 Python 2.7.13 64bits

I tried chunksize while reading big CSV file

reader = pd.read_csv(filePath,chunksize=1000000,low_memory=False,header=0)

The read is now the list. We can iterate the reader and write/append to the new csv or can perform any operation

for chunk in reader:
    print(newChunk.columns)
    print("Chunk -> File process")
    with open(destination, 'a') as f:
        newChunk.to_csv(f, header=False,sep='\t',index=False)
        print("Chunk appended to the file")

I use Pandas on my Linux box and faced many memory leaks that only got resolved after upgrading Pandas to the latest version after cloning it from github.

I encountered this issue as well when I was running in a virtual machine, or somewere else where the memory is stricktly limited. It has nothing to do with pandas or numpy or csv, but will always happen if you try using more memory as you are alowed to use, not even only in python.

The only chance you have is what you already tried, try to chomp down the big thing into smaller pieces which fit into memory.

If you ever asked yourself what MapReduce is all about, you found out by yourself...MapReduce would try to distribute the chunks over many machines, you would try to process the chunke on one machine one after another.

What you found out with the concatenation of the chunk files might be an issue indeed, maybe there are some copy needed in this operation...but in the end this maybe saves you in your current situation but if your csv gets a little bit larger you might run against that wall again...

It also could be, that pandas is so smart, that it actually only loads the individual data chunks into memory if you do something with it, like concatenating to a big df?

Several things you can try:

• Don't load all the data at once, but split in in pieces
• As far as I know, hdf5 is able to do these chunks automatically and only loads the part your program currently works on
• Look if the types are ok, a string '0.111111' needs more memory than a float
• What do you need actually, if there is the adress as a string, you might not need it for numerical analysis...
• A database can help acessing and loading only the parts you actually need (e.g. only the 1% active users)