TL;DR: Scroll to the end for the provided code with suggested improvements
Suggestions
1. Standalone cartesian_product helper function outside scope of combinations
A preferable solution would be to simply use itertools.product as most Python-savvy readers will be familiar with it - and it is well-documented for those that aren’t. However, since you explicitly mention it...
If you still don't want to use itertools.product:
While it does improve clarity to construct a hierarchy of scopes that expose things only when necessary, in this case, I believe that making cartesian_product a nested function inside combinations serves only to confuse the purpose of combinations. It is better to define cartesian_product above, making the code below cleaner and easier to understand. Now someone reading your code will first see cartesian_product’s definition and understand its relatively simple purpose. Then, inside ParametricMapping.combinations, readers are already familiar with cartesian_product, and their train of thought isn’t derailed by trying to understand a nested function.
2. flatten helper function
This helper function should be used thusly:
for term in cartesian_product(*pools):
results.append(flatten(term))
This may seem silly as flattening is a simple operation, but in this example there are a few fairly tricky list/dict comprehensions nearby. Therefore, it may help to replace that piece with a simple flatten call to clean up some confusion and to emphasize that this is a straight-forward operation - A fact that could be lost on some readers of the code in its current state. My point here is that stacking lots of loops and comprehensions on top of each other (especially without documentation/comments) can quickly get messy and confusing.
3. Consolidate some code
If you took both of the above suggestions, a nice opportunity to consolidate some code has presented itself. After the above changes, the section that was originally
results = []
for term in cartesian_product(*pools):
results.append([pp for p in term for pp in p])
should now be
results = []
for term in itertools.product(*pools):
results.append(flatten(term))
This block can be expressed clearly and concisely with the following list comprehension:
results = [flatten(term) for term in itertools.product(*pools)]
Note that because of the separation of functionality into helper functions, the purpose of this list comprehension is abundantly clear. It creates results, which is a list containing the flattened outputs of product.
4. Check for empty mappings at top of combinations
Instead of the if/else at the end of combinations that checks for len(tmp) == 0, make the first two lines of combinations the following:
if self.mappings == [{}]:
return [{}]
The else clause at the bottom of combinations is no longer necessary, and you can simply return tmp. This is cleaner because it handles the case where mappings is empty immediately, which means those cases bypass all the code that was being executed before when len(tmp) was evaluated at the end of combinations. This also allows readers to make the assumption that mappings is non-empty when they get into the actual work being done by combinations, which is just one less thing to worry about. Alternatively, a more concise option would be to replace
if len(tmp) == 0:
return [{}]
else:
return tmp
with
return tmp or [{}]
This works because if mappings is empty, the value of tmp will be an empty list, which will evaluate to False, returning the value following the or. This more concise version may come at the cost of decreased readability.
5. Define labels and pools using helper methods or non-init fields
Remove the first two code lines from combinations, and move them to either helper methods, or additional fields that are processed post-initialization. I recommend doing this to further clarify the work actually being done by combinations and to clean up the code overall. These both have the added benefit of exposing labels and pools as additional attributes of the dataclass that can be accessed outside of combinations.
See the ParametricMapping1 class defined in the section below for an implementation using helper methods, and see the alternative ParametricMapping2 class defined below it for an implementation using non-init fields.
TL;DR
In the end, if all suggestions are followed, the code should include the below declaration of flatten, along with one of the two following blocks (ParametricMapping1, or ParametricMapping2):
def flatten(l):
return [item for sublist in l for item in sublist]
With either ...
@dataclass
class ParametricMapping1:
mappings: List[Mapping[Tuple[str], Sequence[float]]] = field(default_factory=lambda: [{}])
def _labels(self) -> List[Tuple[str]]:
return flatten(self.mappings)
def _pools(self) -> List[List[Sequence[float]]]:
return [list(map(tuple, zip(*arg.values()))) for arg in self.mappings]
@property
def combinations(self) -> List[Mapping[Tuple[str], float]]:
if self.mappings == [{}]:
return [{}]
pool_values = [flatten(term) for term in itertools.product(*self._pools())]
return [dict(zip(self._labels(), v)) for v in pool_values]
Or ...
@dataclass
class ParametricMapping2:
mappings: List[Mapping[Tuple[str], Sequence[float]]] = field(default_factory=lambda: [{}])
labels: List[Tuple[str]] = field(init=False, repr=False)
pools: List[List[Sequence[float]]] = field(init=False, repr=False)
def __post_init__(self):
self.labels = flatten(self.mappings)
self.pools = [list(map(tuple, zip(*arg.values()))) for arg in self.mappings]
@property
def combinations(self) -> List[Mapping[Tuple[str], float]]:
pool_values = [flatten(term) for term in itertools.product(*self.pools)]
return [dict(zip(self.labels, v)) for v in pool_values] or [{}]
Edit (2019-01-09 - 1530):
- The definitions of
_labels and self.labels in the above two code blocks, respectively, have been simplified per @MathiasEttinger's excellent suggestion. See revision history for their original definitions.
