Return generator instead of list from df.to_dict()

Question

I am working on a large Pandas DataFrame which needs to be converted into dictionaries before being processed by another API.

The required dictionaries can be generated by calling the .to_dict(orient='records') method. As stated in the docs, the returned value depends on the orient option:

Returns: dict, list or collections.abc.Mapping

Return a collections.abc.Mapping object representing the DataFrame. The resulting transformation depends on the orient parameter.

For my case, passing orient='records', a list of dictionaries is returned. When dealing with lists, the complete memory required to store the list items, is reserved/allocated. As my dataframe can get rather large, this might lead to memory issues especially as the code might be executed on lower spec target systems.

I could certainly circumvent this issue by processing the dataframe chunk-wise and generate the list of dictionaries for each chunk which is then passed to the API. Furthermore, calling iter(df.to_dict(orient='records')) would return the desired generator, but would not reduce the required memory footprint as the list is created intermediately.

Is there a way to directly return a generator expression from df.to_dict(orient='records') instead of a list in order to reduce the memory footprint?

Answer 1

There is not a way to get a generator directly from to_dict(orient='records'). However, it is possible to modify the to_dict source code to be a generator instead of returning a list comprehension:

from pandas.core.common import standardize_mapping
from pandas.core.dtypes.cast import maybe_box_native


def dataframe_records_gen(df_):
    columns = df_.columns.tolist()
    into_c = standardize_mapping(dict)

    for row in df_.itertuples(index=False, name=None):
        yield into_c(
            (k, maybe_box_native(v)) for k, v in dict(zip(columns, row)).items()
        )

Sample Code:

import pandas as pd

df = pd.DataFrame({
    'A': [1, 2],
    'B': [3, 4]
})

# Using Generator
for row in dataframe_records_gen(df):
    print(row)

# For Comparison with to_dict function
print("to_dict", df.to_dict(orient='records'))

Output:

{'A': 1, 'B': 3}
{'A': 2, 'B': 4}
to_dict [{'A': 1, 'B': 3}, {'A': 2, 'B': 4}]

---

For more natural syntax, it's also possible to register a custom accessor:

import pandas as pd
from pandas.core.common import standardize_mapping
from pandas.core.dtypes.cast import maybe_box_native


@pd.api.extensions.register_dataframe_accessor("gen")
class GenAccessor:
    def __init__(self, pandas_obj):
        self._obj = pandas_obj

    def records(self):
        columns = self._obj.columns.tolist()
        into_c = standardize_mapping(dict)

        for row in self._obj.itertuples(index=False, name=None):
            yield into_c(
                (k, maybe_box_native(v))
                for k, v in dict(zip(columns, row)).items()
            )

Which makes this generator accessible via the gen accessor in this case:

df = pd.DataFrame({
        'A': [1, 2],
        'B': [3, 4]
    })

# Using Generator through registered custom accessor
for row in df.gen.records():
    print(row)

# For Comparison with to_dict function
print("to_dict", df.to_dict(orient='records'))

Output:

{'A': 1, 'B': 3}
{'A': 2, 'B': 4}
to_dict [{'A': 1, 'B': 3}, {'A': 2, 'B': 4}]