CPPItertools

Range-based for loop add-ons inspired by the Python builtins and itertools library. Like itertools and the Python3 builtins, this library uses lazy evaluation wherever possible.

Note: Everything is inside the iter namespace.

Follow @cppitertools for updates.

Build and Test Status

Status	Compilers
	gcc-7 gcc-8 gcc-9 clang-5.0 clang-6.0 clang-7 clang-8 clang-9
	MSVC 2017 MSVC 2019

Table of Contents

range
enumerate
zip
zip_longest
imap
filter
filterfalse
unique_everseen
unique_justseen
takewhile
dropwhile
cycle
repeat
count
groupby
starmap
accumulate
compress
sorted
chain
chain.from_iterable
reversed
slice
sliding_window
chunked
batched

Combinatoric fuctions

product
combinations
combinations_with_replacement
permutations
powerset

Requirements

This library is header-only and relies only on the C++ standard library. The only exception is zip_longest which uses boost::optional. #include <cppitertools/itertools.hpp> will include all of the provided tools except for zip_longest which must be included separately. You may also include individual pieces with the relevant header (#include <cppitertools/enumerate.hpp> for example).

Running tests

You may use either scons or bazel to build the tests. scons seems to work better with viewing the test output, but the same bazel command can be run from any directory.

To run tests with scons you must be within the test directory

test$ # build and run all tests
test$ scons
test$ ./test_all
test$ # build and run a specific test
test$ scons test_enumerate
test$ ./test_enumerate
test$ valgrind ./test_enumerate

bazel absolute commands can be run from any directory inside the project

$ bazel test //test:all # runs all tests
$ bazel test //test:test_enumerate # runs a specific test

Requirements of passed objects

Most itertools will work with iterables using InputIterators and not copy or move any underlying elements. The itertools that need ForwardIterators or have additional requirements are noted in this document. However, the cases should be fairly obvious: any time an element needs to appear multiple times (as in combinations or cycle) or be looked at more than once (specifically, sorted). This library takes every effort to rely on as little as possible from the underlying iterables, but if anything noteworthy is needed it is described in this document.

Guarantees of implementations

By implementations, I mean the objects returned by the API's functions. All of the implementation classes are move-constructible, not copy-constructible, not assignable. All iterators that work over another iterable are tagged as InputIterators and behave as such.

Feedback

If you find anything not working as you expect, not compiling when you believe it should, a divergence from the python itertools behavior, or any sort of error, please let me know. The preferable means would be to open an issue on GitHub. If you want to talk about an issue that you don't feel would be appropriate as a GitHub issue (or you just don't want to open one), you can email me directly with whatever code you have that describes the problem; I've been pretty responsive in the past. If I believe you are "misusing" the library, I'll try to put the blame on myself for being unclear in this document and take the steps to clarify it. So please, contact me with any concerns, I'm open to feedback.

How (not) to use this library

The library functions create and return objects that are properly templated on the iterable they are passed. These exact names of these types or precisely how they are templated is unspecified, you should rely on the functions described in this document. If you plan to use these functions in very simple, straight forward means as in the examples on this page, then you will be fine. If you feel like you need to open the header files, then I've probably under-described something, let me know.

Handling of rvalues vs lvalues

The rules are pretty simple, and the library can be largely used without knowledge of them. Let's take an example

std::vector<int> vec{2,4,6,8};
for (auto&& p : enumerate(vec)) { /* ... */ }

In this case, enumerate will return an object that has bound a reference to vec. No copies are produced here, neither of vec nor of the elements it holds.

If an rvalue was passed to enumerate, binding a reference would be unsafe. Consider:

for (auto&& p : enumerate(std::vector<int>{2,4,6,8})) { /* ... */ }

Instead, enumerate will return an object that has the temporary moved into it. That is, the returned object will contain a std::vector<int> rather than just a reference to one. This may seem like a contrived example, but it matters when enumerate is passed the result of a function call like enumerate(f()), or, more obviously, something like enumerate(zip(a, b)). The object returned from zip must be moved into the enumerate object. As a more specific result, itertools can be mixed and nested.

Pipe syntax

Wherever it makes sense, I've implemented the "pipe" operator that has become common in similar libraries. When the syntax is available, it is done by pulling out the iterable from the call and placing it before the tool. For example:

filter(pred, seq);  // regular call
seq | filter(pred);  // pipe-style
enumerate(seq);  // regular call
seq | enumerate;  // pipe-style.

The following tools support pipe. The remaining I left out because although some of them have multiple reasonable versions, it wasn't obvious to me how I would expect them to behave:

• accumulate
• chain.from_iterable
• chunked
• batched
• combinations
• combinations_with_replacement
• cycle
• dropwhile
• enumerate
• filter
• filterfalse
• groupby
• imap
• permutations
• powerset
• reversed
• slice
• sliding_window
• sorted
• starmap
• takewhile
• unique_everseen (*only without custom hash and equality callables)
• unique_justseen

I don't personally care for the piping style, but it seemed to be desired by the users.

range

Uses an underlying iterator to achieve the same effect of the python range function. range can be used in three different ways:

Only the stopping point is provided. Prints 0 1 2 3 4 5 6 7 8 9

for (auto i : range(10)) {
    cout << i << '\n';
}

The start and stop are both provided. Prints 10 11 12 13 14

for (auto i : range(10, 15)) {
    cout << i << '\n';
}

The start, stop, and step are all provided. Prints 20 22 24 26 28

for (auto i : range(20, 30, 2)) {
    cout << i << '\n';
}

Negative values are allowed as well. Prints 2 1 0 -1 -2

for (auto i : range(2, -3, -1)) {
    cout << i << '\n';
}

A step size of 0 results in an empty range (Python's raises an exception). The following prints nothing

for (auto i : range(0, 10, 0)) {
    cout << i << '\n';
}

In addition to normal integer range operations, doubles and other numeric types are supported through the template

Prints: 5.0 5.5 6.0 ... 9.5

for(auto i : range(5.0, 10.0, 0.5)) {
    cout << i << '\n';
}

Implementation Note: Typical ranges have their current value incremented by the step size repeatedly (value += step). Floating point range value are recomputed at each step to avoid accumulating floating point inaccuracies (value = start + (step * steps_taken). The result of the latter is a bit slower but more accurate.

range also supports the following operations:

• .size() to get the number of elements in the range (not enabled for floating point ranges).
• Accessors for .start(), .stop(), and .step().
• Indexing. Given a range r, r[n] is the nth element in the range.

enumerate

Continually "yields" containers similar to pairs. They are structs with the index in .first, and the element in .second, and also work with structured binding declarations. Usage appears as:

vector<int> vec{2, 4, 6, 8};
for (auto&& [i, e] : enumerate(vec)) {
    cout << i << ": " << e << '\n';
}

filter

Called as filter(predicate, iterable). The predicate can be any callable. filter will only yield values that are true under the predicate.

Prints values greater than 4: 5 6 7 8

vector<int> vec{1, 5, 4, 0, 6, 7, 3, 0, 2, 8, 3, 2, 1};
for (auto&& i : filter([] (int i) { return i > 4; }, vec)) {
    cout << i <<'\n';
}

If no predicate is passed, the elements themselves are tested for truth

Prints only non-zero values.

for(auto&& i : filter(vec)) {
    cout << i << '\n';
}

filterfalse

Similar to filter, but only prints values that are false under the predicate.

Prints values not greater than 4: 1 4 3 2 3 2 1

vector<int> vec{1, 5, 4, 0, 6, 7, 3, 0, 2, 8, 3, 2, 1};
for (auto&& i : filterfalse([] (int i) { return i > 4; }, vec)) {
    cout << i <<'\n';
}

If no predicate is passed, the elements themselves are tested for truth.

Prints only zero values.

for(auto&& i : filterfalse(vec)) {
    cout << i << '\n';
}

unique_everseen

Additional Requirements: Underlying values must be copy-constructible.

This is a filter adaptor that only generates values that have never been seen before.

Prints 1 2 3 4 5 6 7 8 9

vector<int> v {1,2,3,4,3,2,1,5,6,7,7,8,9,8,9,6};
for (auto&& i : unique_everseen(v)) {
    cout << i << ' ';
}

unique_everseen uses an undordered_set so it needs hashable elements. For types that don't work with std::hash or std::equal_to, unique_everseen also provides an overload taking a hash callable and an equality callable. This does not work with the pipe syntax.

vector<Widget> v { /* ... */ };
for (auto&& w : unique_everseen(v, WidgetHash{}, WidgetEq{})) {
  cout << w.name() << ' ';
}

unique_justseen

Another filter adaptor that only omits consecutive duplicates.

Prints 1 2 3 4 3 2 1 Example Usage:

vector<int> v {1,1,1,2,2,3,3,3,4,3,2,1,1,1};
for (auto&& i : unique_justseen(v)) {
    cout << i << ' ';
}

If elements cannot be directly compared with equality, you can pass in a key callable.

vector<Person> v { /* ... */ };
for (auto&& p : unique_justseen(v, [] (const Person& p) { return p.name; }))
  cout << p.name() << ' ' << p.age() << '\n';
}

takewhile

Yields elements from an iterable until the first element that is false under the predicate is encountered.

Prints 1 2 3 4. (5 is false under the predicate)

vector<int> ivec{1, 2, 3, 4, 5, 6, 7, 6, 5, 4, 3, 2, 1};
for (auto&& i : takewhile([] (int i) {return i < 5;}, ivec)) {
    cout << i << '\n';
}

dropwhile

Yields all elements after and including the first element that is true under the predicate.

Prints 5 6 7 1 2

vector<int> ivec{1, 2, 3, 4, 5, 6, 7, 1, 2};
for (auto&& i : dropwhile([] (int i) {return i < 5;}, ivec)) {
    cout << i << '\n';
}

cycle

Additional Requirements: Input must have a ForwardIterator

Repeatedly produces all values of an iterable. The loop will be infinite, so a break or other control flow structure is necessary to exit.

Prints 1 2 3 repeatedly until some_condition is true

vector<int> vec{1, 2, 3};
for (auto&& i : cycle(vec)) {
    cout << i << '\n';
    if (some_condition) {
        break;
    }
}

repeat

Repeatedly produces a single argument forever, or a given number of times. repeat will bind a reference when passed an lvalue and move when given an rvalue. It will then yield a reference to the same item until completion.

The below prints 1 five times.

for (auto&& e : repeat(1, 5)) {
    cout << e << '\n';
}

The below prints 2 forever

for (auto&& e : repeat(2)) {
    cout << e << '\n';
}

count

Effectively a range without a stopping point.
count() with no arguments will start counting from 0 with a positive step of 1.
count(i) will start counting from i with a positive step of 1.
count(i, st) will start counting from i with a step of st.

Technical limitations: Unlike Python which can use its long integer types when needed, count() would eventually exceed the maximum possible value for its type (or minimum with a negative step). count is actually implemented as a range with the stopping point being the std::numeric_limits<T>::max() for the integral type (long by default)

The below will print 0 1 2 ... etc

for (auto&& i : count()) {
    cout << i << '\n';
}

groupby

Additional Requirements: If the Input's iterator's operator*() returns a reference, the reference must remain valid after the iterator is incremented. Roughly equivalent to requiring the Input have a ForwardIterator.

Separate an iterable into groups sharing a common key. The following example creates a new group whenever a string of a different length is encountered.

vector<string> vec = {
    "hi", "ab", "ho",
    "abc", "def",
    "abcde", "efghi"
};

for (auto&& gb : groupby(vec, [] (const string &s) {return s.length(); })) {
    cout << "key: " << gb.first << '\n';
    cout << "content: ";
    for (auto&& s : gb.second) {
        cout << s << "  ";
    }
    cout << '\n';
}

Note: Just like Python's itertools.groupby, this doesn't do any sorting. It just iterates through, making a new group each time there is a key change. Thus, if the group is unsorted, the same key may appear multiple times.

starmap

Takes a sequence of tuple-like objects (anything that works with std::get) and unpacks each object into individual arguments for each function call. The below example takes a vector of pairs of ints, and passes them to a function expecting two ints, with the elements of the pair being the first and second arguments to the function.

vector<pair<int, int>> v = {{2, 3}, {5, 2}, {3, 4}}; // {base, exponent}
for (auto&& i : starmap([](int b, int e){return pow(b, e);}, v)) {
    // ...
}

starmap can also work over a tuple-like object of tuple-like objects even when the contained objects are different as long as the functor works with multiple types of calls. For example, a Callable struct with overloads for its operator() will work as long as all overloads have the same return type

struct Callable {
    int operator()(int i) const;
    int operator()(int i, char c) const;
    int operator()(double d, int i, char c) const;
};

This will work with a tuple of mixed types

auto t = make_tuple(
        make_tuple(5), // first form
        make_pair(3, 'c'), // second
        make_tuple(1.0, 1, '1')); // third
for (auto&& i : starmap(Callable{}, t)) {
    // ...
}

accumulate

Additional Requirements: Type return from functor (with reference removed) must be assignable.

Differs from std::accumulate (which in my humble opinion should be named std::reduce or std::foldl). It is similar to a functional reduce where one can see all of the intermediate results. By default, it keeps a running sum. Prints: 1 3 6 10 15

for (auto&& i : accumulate(range(1, 6))) {
    cout << i << '\n';
}

A second, optional argument may provide an alternative binary function to compute results. The following example multiplies the numbers, rather than adding them. Prints: 1 2 6 24 120

for (auto&& i : accumulate(range(1, 6), std::multiplies<int>{})) {
    cout << i << '\n';
}

Note: The intermediate result type must support default construction and assignment.

zip

Takes an arbitrary number of ranges of different types and efficiently iterates over them in parallel (so an iterator to each container is incremented simultaneously). When you dereference an iterator to "zipped" range you get a tuple of the elements the iterators were holding.

Example usage:

array<int,4> iseq{{1,2,3,4}};
vector<float> fseq{1.2,1.4,12.3,4.5,9.9};
vector<string> sseq{"i","like","apples","a lot","dude"};
array<double,5> dseq{{1.2,1.2,1.2,1.2,1.2}};

for (auto&& [i, f, s, d] : zip(iseq, fseq, sseq, dseq)) {
    cout << i << ' ' << f << ' ' << s << ' ' << d << '\n';
    f = 2.2f; // modifies the underlying 'fseq' sequence
}

zip_longest

Terminates on the longest sequence instead of the shortest. Repeatedly yields a tuple of boost::optional<T>s where T is the type yielded by the sequences' respective iterators. Because of its boost dependency, zip_longest is not in itertools.hpp and must be included separately. The following loop prints either "Just " or "Nothing" for each element in each tuple yielded.

vector<int> v1 = {0, 1, 2, 3};
vector<int> v2 = {10, 11};
for (auto&& [x, y] : zip_longest(v1, v2)) {
    cout << '{';
    if (x) {
        cout << "Just " << *x;
    } else {
        cout << "Nothing";
    }
    cout << ", ";
    if (y) {
        cout << "Just " << *y;
    } else {
        cout << "Nothing";
    }
    cout << "}\n";
}

The output is:

{Just 0, Just 10}
{Just 1, Just 11}
{Just 2, Nothing}
{Just 3, Nothing}

imap

Takes a function and one or more iterables. The number of iterables must match the number of arguments to the function. Applies the function to each element (or elements) in the iterable(s). Terminates on the shortest sequence.

Prints the squares of the numbers in vec: 1 4 9 16 25

vector<int> vec{1, 2, 3, 4, 5};
for (auto&& i : imap([] (int x) {return x * x;}, vec)) {
    cout << i << '\n';
}

With more than one sequence, the below adds corresponding elements from each vector together, printing 11 23 35 47 59 71

vector<int> vec1{1, 3, 5, 7, 9, 11};
vector<int> vec2{10, 20, 30, 40, 50, 60};
for (auto&& i : imap([] (int x, int y) { return x + y; }, vec1, vec2)) {
    cout << i << '\n';
}

Note: The name imap is chosen to prevent confusion/collision with std::map, and because it is more related to itertools.imap than the python builtin map.

compress

Yields only the values corresponding to true in the selectors iterable. Terminates on the shortest sequence.

Prints 2 6

vector<int> ivec{1, 2, 3, 4, 5, 6};
vector<bool> bvec{false, true, false, false, false, true};
for (auto&& i : compress(ivec, bvec) {
    cout << i << '\n';
}

sorted

Additional Requirements: Input must have a ForwardIterator

Allows iteration over a sequence in sorted order. sorted does not produce a new sequence, copy elements, or modify the original sequence. It only provides a way to iterate over existing elements. sorted also takes an optional second comparator argument. If not provided, defaults to std::less.
Iterables passed to sorted are required to have an iterator with an operator*() const member.

The below outputs 0 1 2 3 4.

unordered_set<int> nums{4, 0, 2, 1, 3};
for (auto&& i : sorted(nums)) {
    cout << i << '\n';
}

chain

Additional Requirements: The underlying iterators of all containers' operator* must have the exact same type

This can chain any set of ranges together as long as their iterators dereference to the same type.

vector<int> empty{};
vector<int> vec1{1,2,3,4,5,6};
array<int,4> arr1{{7,8,9,10}};

for (auto&& i : chain(empty,vec1,arr1)) {
    cout << i << '\n';
}

chain.from_iterable

Similar to chain, but rather than taking a variadic number of iterables, it takes an iterable of iterables and chains the contained iterables together. A simple example is shown below using a vector of vectors to represent a 2d ragged array, and prints it in row-major order.

vector<vector<int>> matrix = {
    {1, 2, 3},
    {4, 5},
    {6, 8, 9, 10, 11, 12}
};

for (auto&& i : chain.from_iterable(matrix)) {
    cout << i << '\n';
}

reversed

Additional Requirements: Input must be compatible with std::rbegin() and std::rend()

Iterates over elements of a sequence in reverse order.

for (auto&& i : reversed(a)) {
    cout << i << '\n';
}

slice

Returns selected elements from a range, parameters are start, stop and step. the range returned is [start,stop) where you only take every step element

This outputs 0 3 6 9 12

vector<int> a{0,1,2,3,4,5,6,7,8,9,10,11,12,13};
for (auto&& i : slice(a,0,15,3)) {
    cout << i << '\n';
}

sliding_window

Additional Requirements: Input must have a ForwardIterator

Takes a section from a range and increments the whole section. If the window size is larger than the length of the input, the sliding_window will yield nothing (begin == end).

Example: [1, 2, 3, 4, 5, 6, 7, 8, 9]

take a section of size 4, output is:

1 2 3 4
2 3 4 5
3 4 5 6
4 5 6 7
5 6 7 8
6 7 8 9

Example Usage:

vector<int> v = {1,2,3,4,5,6,7,8,9};
for (auto&& sec : sliding_window(v,4)) {
    for (auto&& i : sec) {
        cout << i << ' ';
        i.get() = 90;
    }
    cout << '\n';
}

chunked

chunked will yield subsequent chunks of an iterable in blocks of a specified size. The final chunk may be shorter than the rest if the chunk size given does not evenly divide the length of the iterable.

Example usage:

vector<int> v {1,2,3,4,5,6,7,8,9};
for (auto&& sec : chunked(v,4)) {
    for (auto&& i : sec) {
        cout << i << ' ';
    }
    cout << '\n';
}

The above prints:

1 2 3 4
5 6 7 8
9

batched

batched will yield a given number N of batches containing subsequent elements from an iterable, assuming the iterable contains at least N elements. The size of each batch is immaterial, but the implementation guarantees that no two batches will differ in size by more than 1.

Example usage:

vector<int> v {1,2,3,4,5,6,7,8,9};
for (auto&& sec : batched(v,4)) {
    for (auto&& i : sec) {
        cout << i << ' ';
    }
    cout << '\n';
}

The above prints:

1 2 3
4 5
6 7
8 9

product

Additional Requirements: Input must have a ForwardIterator

Generates the cartesian product of the given ranges put together.

Example usage:

vector<int> v1{1,2,3};
vector<int> v2{7,8};
vector<string> v3{"the","cat"};
vector<string> v4{"hi","what's","up","dude"};
for (auto&& [a, b, c, d] : product(v1,v2,v3,v4)) {
    cout << a << ", " << b << ", " << c << ", " << d << '\n';
}

Product also accepts a "repeat" as a template argument. Currently this is the only way to do repeats. If you are reading this and need product(seq, 3) instead of product<3>(seq) please open an issue.

Example usage:

std::string s = "abc";
// equivalent of product(s, s, s);
for (auto&& t : product<3>(s)) {
   // ...
}

combinations

Additional Requirements: Input must have a ForwardIterator

Generates n length unique sequences of the input range.

Example usage:

vector<int> v = {1,2,3,4,5};
for (auto&& i : combinations(v,3)) {
    for (auto&& j : i ) cout << j << " ";
    cout << '\n';
}

combinations_with_replacement

Additional Requirements: Input must have a ForwardIterator

Like combinations, but with replacement of each element. The below is printed by the loop that follows:

{A, A}
{A, B}
{A, C}
{B, B}
{B, C}
{C, C}

for (auto&& v : combinations_with_replacement(s, 2)) {
    cout << '{' << v[0] << ", " << v[1] << "}\n";
}

permutations

Additional Requirements: Input must have a ForwardIterator. Iterator must have an operator*() const.

Generates all the permutations of a range using std::next_permutation.

Example usage:

vector<int> v = {1,2,3,4,5};
for (auto&& vec : permutations(v)) {
    for (auto&& i : vec) {
        cout << i << ' ';
    }
    cout << '\n';
}

powerset

Additional Requirements: Input must have a ForwardIterator

Generates every possible subset of a set, runs in O(2^n).

Example usage:

vector<int> vec {1,2,3,4,5,6,7,8,9};
for (auto&& v : powerset(vec)) {
    for (auto&& i : v) {
        cout << i << " ";
    }
    cout << '\n';
}