Pandas - Slice large dataframe into chunks

Question

I have a large dataframe (>3MM rows) that I'm trying to pass through a function (the one below is largely simplified), and I keep getting a Memory Error message.

I think I'm passing too large of a dataframe into the function, so I'm trying to:

1) Slice the dataframe into smaller chunks (preferably sliced by AcctName)

2) Pass the dataframe into the function

3) Concatenate the dataframes back into one large dataframe

def trans_times_2(df):
    df['Double_Transaction'] = df['Transaction'] * 2

large_df 
AcctName   Timestamp    Transaction
ABC        12/1         12.12
ABC        12/2         20.89
ABC        12/3         51.93    
DEF        12/2         13.12
DEF        12/8          9.93
DEF        12/9         92.09
GHI        12/1         14.33
GHI        12/6         21.99
GHI        12/12        98.81

I know that my function works properly, since it will work on a smaller dataframe (e.g. 40,000 rows). I tried the following, but I was unsuccessful with concatenating the small dataframes back into one large dataframe.

def split_df(df):
    new_df = []
    AcctNames = df.AcctName.unique()
    DataFrameDict = {elem: pd.DataFrame for elem in AcctNames}
    key_list = [k for k in DataFrameDict.keys()]
    new_df = []
    for key in DataFrameDict.keys():
        DataFrameDict[key] = df[:][df.AcctNames == key]
        trans_times_2(DataFrameDict[key])
    rejoined_df = pd.concat(new_df)

How I envision the dataframes being split:

df1
AcctName   Timestamp    Transaction  Double_Transaction
ABC        12/1         12.12        24.24
ABC        12/2         20.89        41.78
ABC        12/3         51.93        103.86

df2
AcctName   Timestamp    Transaction  Double_Transaction
DEF        12/2         13.12        26.24
DEF        12/8          9.93        19.86
DEF        12/9         92.09        184.18

df3
AcctName   Timestamp    Transaction  Double_Transaction
GHI        12/1         14.33        28.66
GHI        12/6         21.99        43.98
GHI        12/12        98.81        197.62

Solution 1

You can use list comprehension to split your dataframe into smaller dataframes contained in a list.

n = 200000  #chunk row size
list_df = [df[i:i+n] for i in range(0,df.shape[0],n)]

Or use numpy array_split:

list_df = np.array_split(df, n)

You can access the chunks with:

list_df[0]
list_df[1]
etc...

Then you can assemble it back into a one dataframe using pd.concat.

By AcctName

list_df = []

for n,g in df.groupby('AcctName'):
    list_df.append(g)

Solution 2

I'd suggest using a dependency more_itertools. It handles all edge cases like uneven partition of the dataframe and returns an iterator that will make things a tiny bit more efficient.

(updated using code from @Acumenus)

from more_itertools import sliced
CHUNK_SIZE = 5

index_slices = sliced(range(len(df)), CHUNK_SIZE)

for index_slice in index_slices:
  chunk = df.iloc[index_slice] # your dataframe chunk ready for use

Solution 3

I love @ScottBoston answer, although, I still haven't memorized the incantation. Here's a more verbose function that does the same thing:

def chunkify(df: pd.DataFrame, chunk_size: int):
    start = 0
    length = df.shape[0]

    # If DF is smaller than the chunk, return the DF
    if length <= chunk_size:
        yield df[:]
        return

    # Yield individual chunks
    while start + chunk_size <= length:
        yield df[start:chunk_size + start]
        start = start + chunk_size

    # Yield the remainder chunk, if needed
    if start < length:
        yield df[start:]

To rebuild the data frame, accumulate each chunk in a list, then pd.concat(chunks, axis=1)

Comments

[Walt Reed]

Thanks Scott! Is there a way to split into smaller dataframes based on AcctName instead of chunk size?

[Scott Boston]

@WaltReed Try that second part using groupby.

[Walt Reed]

Okay great, that worked! I'm calling this inside a function, but when I try to view the new dataframe after running the function, I get the error NameError: name 'new_df' is not defined. What am I missing here?

[Scott Boston]

If you created that list onside the function then it is a local variable. You may need to put the keyword global in front of list_df = []

[Asclepius]

Can use len(df) instead of df.shape[0].

[santon]

Can't you just do something a bit more direct: for chunk in sliced(df, CHUNK_SIZE)?

[roganjosh]

Your numpy method is not equivalent to the list comprehension. I think you intended it to be np.array_split(df, math.ceil(len(df) / N))? If the intention was to show two different approaches to chunking a df, I think the numpy method warrants some initial explanation. Even using math.ceil() wouldn't guarantee the same behaviour as shown in the second example in the docs

[Martin Ueding]

@santon: Yes, no difference. The generator can only be iterated once anyway, so it doesn't make much sense to store it as a variable.