Read data from CSV file and transform from string to correct data-type, including a list-of-integer column

You have to map your rows:

import csv
import io

data = u"""
True,foo,1,2.3,baz
False,bar,7,9.8,qux
"""

reader = csv.reader(io.StringIO(data, newline=""), delimiter=",")
parsed = (({'True': True}.get(row[0],False), row[1], int(row[2]), float(row[3]), row[4])
              for row in reader)
for row in parsed:
    print(row)

Results in

(True, 'foo', 1, 2.3, 'baz')
(False, 'bar', 7, 9.8, 'qux')

An alternative (although it seems a bit extreme) in lieu of using ast.literal_eval is the pyparsing module available on PyPi – and see if the http://pyparsing.wikispaces.com/file/view/parsePythonValue.py code sample is either appropriate for what you require, or can be easily adapted.

Props to Jon Clements and cortopy for teaching me about ast.literal_eval! Here’s what I ended up going with (Python 2; changes for 3 should be trivial):

from ast import literal_eval
from csv import DictReader
import csv


def csv_data(filepath, **col_conversions):
    """Yield rows from the CSV file as dicts, with column headers as the keys.

    Values in the CSV rows are converted to Python values when possible,
    and are kept as strings otherwise.

    Specific conversion functions for columns may be specified via
    `col_conversions`: if a column's header is a key in this dict, its
    value will be applied as a function to the CSV data. Specify
    `ColumnHeader=str` if all values in the column should be interpreted
    as unquoted strings, but might be valid Python literals (`True`,
    `None`, `1`, etc.).

    Example usage:

    >>> csv_data(filepath,
    ...          VariousWordsIncludingTrueAndFalse=str,
    ...          NumbersOfVaryingPrecision=float,
    ...          FloatsThatShouldBeRounded=round,
    ...          **{'Column Header With Spaces': arbitrary_function})
    """

    def parse_value(key, value):
        if key in col_conversions:
            return col_conversions[key](value)
        try:
            # Interpret the string as a Python literal
            return literal_eval(value)
        except Exception:
            # If that doesn't work, assume it's an unquoted string
            return value

    with open(filepath) as f:
        # QUOTE_NONE: don't process quote characters, to avoid the value
        # `"2"` becoming the int `2`, rather than the string `'2'`.
        for row in DictReader(f, quoting=csv.QUOTE_NONE):
            yield {k: parse_value(k, v) for k, v in row.iteritems()}

(I’m a little wary that I might have missed some corner cases involving quoting. Please comment if you see any issues!)

I know this is a fairly old question, tagged python-2.5, but here’s answer that works with Python 3.6+ which might be of interest to folks using more up-to-date versions of the language.

It leverages the built-in typing.NamedTuple class which was added in Python 3.5. What may not be evident from the documentation is that the "type" of each field can be a function.

The example usage code also uses so-called f-string literals which weren’t added until Python 3.6, but their use isn’t required to do the core data-type transformations.

#!/usr/bin/env python3.6
import ast
import csv
from typing import NamedTuple


class Record(NamedTuple):
    """ Define the fields and their types in a record. """
    IsActive: bool
    Type: str
    Price: float
    States: ast.literal_eval  # Handles string represenation of literals.

    @classmethod
    def _transform(cls: 'Record', dict_: dict) -> dict:
        """ Convert string values in given dictionary to corresponding Record
            field type.
        """
        return {name: cls.__annotations__[name](value)
                    for name, value in dict_.items()}


filename = 'test_transform.csv'

with open(filename, newline='') as file:
    for i, row in enumerate(csv.DictReader(file)):
        row = Record._transform(row)
        print(f'row {i}: {row}')

Output:

row 0: {'IsActive': True, 'Type': 'Cellphone', 'Price': 34.0, 'States': [1, 2]}
row 1: {'IsActive': False, 'Type': 'FlatTv', 'Price': 3.5, 'States': [2]}
row 2: {'IsActive': True, 'Type': 'Screen', 'Price': 100.23, 'States': [5, 1]}
row 3: {'IsActive': True, 'Type': 'Notebook', 'Price': 50.0, 'States': [1]}

Generalizing this by creating a base class with just the generic classmethod in it is not simple because of the way typing.NamedTuple is implemented.

To avoid that issue, in Python 3.7+, a dataclasses.dataclass could be used instead because they do not have the inheritance issue — so creating a generic base class that can be reused is simple:

#!/usr/bin/env python3.7
import ast
import csv
from dataclasses import dataclass, fields
from typing import Type, TypeVar

T = TypeVar('T', bound='GenericRecord')

class GenericRecord:
    """ Generic base class for transforming dataclasses. """
    @classmethod
    def _transform(cls: Type[T], dict_: dict) -> dict:
        """ Convert string values in given dictionary to corresponding type. """
        return {field.name: field.type(dict_[field.name])
                    for field in fields(cls)}


@dataclass
class CSV_Record(GenericRecord):
    """ Define the fields and their types in a record.
        Field names must match column names in CSV file header.
    """
    IsActive: bool
    Type: str
    Price: float
    States: ast.literal_eval  # Handles string represenation of literals.


filename = 'test_transform.csv'

with open(filename, newline='') as file:
    for i, row in enumerate(csv.DictReader(file)):
        row = CSV_Record._transform(row)
        print(f'row {i}: {row}')

In one sense it’s not really very important which one you use because an instance of the class in never created — using one is just a clean way of specifying and holding a definition of the field names and their type in a record data-structure.

A TypedDict was added to the typing module in Python 3.8 that can also be used to provide the typing information, but must be used in a slightly different manner since it doesn’t actually define a new type like NamedTuple and dataclasses do — so it requires having a standalone transforming function:

#!/usr/bin/env python3.8
import ast
import csv
from dataclasses import dataclass, fields
from typing import TypedDict


def transform(dict_, typed_dict) -> dict:
    """ Convert values in given dictionary to corresponding types in TypedDict . """
    fields = typed_dict.__annotations__
    return {name: fields[name](value) for name, value in dict_.items()}


class CSV_Record_Types(TypedDict):
    """ Define the fields and their types in a record.
        Field names must match column names in CSV file header.
    """
    IsActive: bool
    Type: str
    Price: float
    States: ast.literal_eval


filename = 'test_transform.csv'

with open(filename, newline='') as file:
    for i, row in enumerate(csv.DictReader(file), 1):
        row = transform(row, CSV_Record_Types)
        print(f'row {i}: {row}')

I love @martineau‘s answer. It’s very clean.

One thing I needed was to convert only a couple of values and leave all the other fields as strings, like having strings as default and just updating the type for specific keys.

To do that, just replace this line:

row = CSV_Record._transform(row)

by this one:

row.update(CSV_Record._transform(row))

The ‘update‘ function updates the variable row directly, merging the raw data from the csv extract with the values converted to the correct type by the ‘_transform‘ method.

Note there is no ‘row = ‘ in the updated version.

Hope this will help in case anyone has a similar requirement.

(PS: I’m quite new to posting on stackoverflow, so please let me know if the above is not clear)

I too really liked @martineau’s approach and was especially intrigued by his comment that the essence of his code was a clean mapping between fields and types. That suggested to me that a dictionary would work also. Hence the variation on his theme shown below. It’s worked nicely for me.

Clearly the value field in the dictionary is really just a callable and thus could be used to provide a hook for data massaging as well as typecasting if one so chose.

import ast
import csv

fix_type = {'IsActive': bool, 'Type': str, 'Price': float, 'States': ast.literal_eval}

filename = 'test_transform.csv'

with open(filename, newline='') as file:
    for i, row in enumerate(csv.DictReader(file)):
        row = {k: fix_type[k](v) for k, v in row.items()}
        print(f'row {i}: {row}')

Output

row 0: {'IsActive': True, 'Type': 'Cellphone', 'Price': 34.0, 'States': [1, 2]}
row 1: {'IsActive': False, 'Type': 'FlatTv', 'Price': 3.5, 'States': [2]}
row 2: {'IsActive': True, 'Type': 'Screen', 'Price': 100.23, 'States': [5, 1]}
row 3: {'IsActive': True, 'Type': 'Notebook', 'Price': 50.0, 'States': [1]}

this is my take on the question in case you have to deal with multiple csv formats, some additional custom data wrangling to perform on some columns, and to output as list of lists or tuple of tuples.

The types are presented as a string, as the columns types are stored in a database outside of the Python code. It also enable to add some custom types if need be.

I haven’t tested this against super large files, though, as in my production code I use pandas and this code is here for some tests setup. But I guess this consumes more memory than some other answers, as all the data from the csv is loaded at once.

dict_headers_type = {
    "IsActive": "bool",
    "Type": "str",
    "Price": "float",
    "State": "list",
}

dict_converters = {
    "bool": x: bool(x),
    "float": x: float(x),
    "list": x: ast.literal_eval(x),
}

dict_header_converter = {
    header: dict_converters[my_type]
    for header, my_type in dict_headers_type.items()
    if my_type in dict_converters.keys()
}

That in place, we can perform the conversion:

with open(csv_path) as f:
    data = [line for line in csv.reader(f)]

# list of the converters to apply
ls_f = [
    dict_header_converter[header]
    if header in dict_header_converter.keys() else None
    for header in data[0]
]


ls_records = [f(datapoint) if f else datapoint
     for f, datapoint in zip(ls_f, row)]
    for row in data[1:]]

# to add headers, if needed:
ls_records.insert(0, data[0])

outputs:

[
  ['IsActive','Type','Price','State']
  [True, 'Cellphone', 34.0, [1, 2]],
  [False, 'FlatTv', 3.5, [2]],
  [True, 'Screen', 100.23, [5, 1]],
  [True, 'Notebook', 50.0, [1]],
]

Here’s a modified version of @user647772’s answer that makes use of the ast.literal_eval() function so it can handle a list-of-integer column (as well as any other valid Python literal expression) in a field in a row of a CSV formatted file.

It works in both Python 2.17 and 3.x.

from ast import literal_eval
import csv
import io

data = u"""
True,foo,1,2.3,baz,"[1, 2]"
False,bar,7,9.8,qux,"[5, 1]"
"""

def evaluate(expression):
    try:
        return literal_eval(expression)
    except ValueError:
        return str(expression)

reader = csv.reader(io.StringIO(data, newline=""), delimiter=",")
parsed = (tuple(evaluate(field) for field in row) for row in reader)
for row in parsed:
    print(row)

Results:

(True, 'foo', 1, 2.3, 'baz', [1, 2])
(False, 'bar', 7, 9.8, 'qux', [5, 1])

If you’re using JSON Schema, you can use singer.transform():

import json
import singer
import csv

with open("my.schema.json") as f:
  schema = json.load(f)

fieldnames = list(schema["properties"].keys())

with open("my.csv", newline="") as f:
  reader = csv.DictReader(f, fieldnames=fieldnames)
  for row in reader:
    print(singer.transform(row, schema))