Related
It's well known that comparing floats for equality is a little fiddly due to rounding and precision issues.
For example: Comparing Floating Point Numbers, 2012 Edition
What is the recommended way to deal with this in Python?
Is a standard library function for this somewhere?
Python 3.5 adds the math.isclose and cmath.isclose functions as described in PEP 485.
If you're using an earlier version of Python, the equivalent function is given in the documentation.
def isclose(a, b, rel_tol=1e-09, abs_tol=0.0):
return abs(a-b) <= max(rel_tol * max(abs(a), abs(b)), abs_tol)
rel_tol is a relative tolerance, it is multiplied by the greater of the magnitudes of the two arguments; as the values get larger, so does the allowed difference between them while still considering them equal.
abs_tol is an absolute tolerance that is applied as-is in all cases. If the difference is less than either of those tolerances, the values are considered equal.
Something as simple as the following may be good enough:
return abs(f1 - f2) <= allowed_error
I would agree that Gareth's answer is probably most appropriate as a lightweight function/solution.
But I thought it would be helpful to note that if you are using NumPy or are considering it, there is a packaged function for this.
numpy.isclose(a, b, rtol=1e-05, atol=1e-08, equal_nan=False)
A little disclaimer though: installing NumPy can be a non-trivial experience depending on your platform.
Use Python's decimal module, which provides the Decimal class.
From the comments:
It is worth noting that if you're
doing math-heavy work and you don't
absolutely need the precision from
decimal, this can really bog things
down. Floats are way, way faster to
deal with, but imprecise. Decimals are
extremely precise but slow.
The common wisdom that floating-point numbers cannot be compared for equality is inaccurate. Floating-point numbers are no different from integers: If you evaluate "a == b", you will get true if they are identical numbers and false otherwise (with the understanding that two NaNs are of course not identical numbers).
The actual problem is this: If I have done some calculations and am not sure the two numbers I have to compare are exactly correct, then what? This problem is the same for floating-point as it is for integers. If you evaluate the integer expression "7/3*3", it will not compare equal to "7*3/3".
So suppose we asked "How do I compare integers for equality?" in such a situation. There is no single answer; what you should do depends on the specific situation, notably what sort of errors you have and what you want to achieve.
Here are some possible choices.
If you want to get a "true" result if the mathematically exact numbers would be equal, then you might try to use the properties of the calculations you perform to prove that you get the same errors in the two numbers. If that is feasible, and you compare two numbers that result from expressions that would give equal numbers if computed exactly, then you will get "true" from the comparison. Another approach is that you might analyze the properties of the calculations and prove that the error never exceeds a certain amount, perhaps an absolute amount or an amount relative to one of the inputs or one of the outputs. In that case, you can ask whether the two calculated numbers differ by at most that amount, and return "true" if they are within the interval. If you cannot prove an error bound, you might guess and hope for the best. One way of guessing is to evaluate many random samples and see what sort of distribution you get in the results.
Of course, since we only set the requirement that you get "true" if the mathematically exact results are equal, we left open the possibility that you get "true" even if they are unequal. (In fact, we can satisfy the requirement by always returning "true". This makes the calculation simple but is generally undesirable, so I will discuss improving the situation below.)
If you want to get a "false" result if the mathematically exact numbers would be unequal, you need to prove that your evaluation of the numbers yields different numbers if the mathematically exact numbers would be unequal. This may be impossible for practical purposes in many common situations. So let us consider an alternative.
A useful requirement might be that we get a "false" result if the mathematically exact numbers differ by more than a certain amount. For example, perhaps we are going to calculate where a ball thrown in a computer game traveled, and we want to know whether it struck a bat. In this case, we certainly want to get "true" if the ball strikes the bat, and we want to get "false" if the ball is far from the bat, and we can accept an incorrect "true" answer if the ball in a mathematically exact simulation missed the bat but is within a millimeter of hitting the bat. In that case, we need to prove (or guess/estimate) that our calculation of the ball's position and the bat's position have a combined error of at most one millimeter (for all positions of interest). This would allow us to always return "false" if the ball and bat are more than a millimeter apart, to return "true" if they touch, and to return "true" if they are close enough to be acceptable.
So, how you decide what to return when comparing floating-point numbers depends very much on your specific situation.
As to how you go about proving error bounds for calculations, that can be a complicated subject. Any floating-point implementation using the IEEE 754 standard in round-to-nearest mode returns the floating-point number nearest to the exact result for any basic operation (notably multiplication, division, addition, subtraction, square root). (In case of tie, round so the low bit is even.) (Be particularly careful about square root and division; your language implementation might use methods that do not conform to IEEE 754 for those.) Because of this requirement, we know the error in a single result is at most 1/2 of the value of the least significant bit. (If it were more, the rounding would have gone to a different number that is within 1/2 the value.)
Going on from there gets substantially more complicated; the next step is performing an operation where one of the inputs already has some error. For simple expressions, these errors can be followed through the calculations to reach a bound on the final error. In practice, this is only done in a few situations, such as working on a high-quality mathematics library. And, of course, you need precise control over exactly which operations are performed. High-level languages often give the compiler a lot of slack, so you might not know in which order operations are performed.
There is much more that could be (and is) written about this topic, but I have to stop there. In summary, the answer is: There is no library routine for this comparison because there is no single solution that fits most needs that is worth putting into a library routine. (If comparing with a relative or absolute error interval suffices for you, you can do it simply without a library routine.)
math.isclose() has been added to Python 3.5 for that (source code). Here is a port of it to Python 2. It's difference from one-liner of Mark Ransom is that it can handle "inf" and "-inf" properly.
def isclose(a, b, rel_tol=1e-09, abs_tol=0.0):
'''
Python 2 implementation of Python 3.5 math.isclose()
https://github.com/python/cpython/blob/v3.5.10/Modules/mathmodule.c#L1993
'''
# sanity check on the inputs
if rel_tol < 0 or abs_tol < 0:
raise ValueError("tolerances must be non-negative")
# short circuit exact equality -- needed to catch two infinities of
# the same sign. And perhaps speeds things up a bit sometimes.
if a == b:
return True
# This catches the case of two infinities of opposite sign, or
# one infinity and one finite number. Two infinities of opposite
# sign would otherwise have an infinite relative tolerance.
# Two infinities of the same sign are caught by the equality check
# above.
if math.isinf(a) or math.isinf(b):
return False
# now do the regular computation
# this is essentially the "weak" test from the Boost library
diff = math.fabs(b - a)
result = (((diff <= math.fabs(rel_tol * b)) or
(diff <= math.fabs(rel_tol * a))) or
(diff <= abs_tol))
return result
I'm not aware of anything in the Python standard library (or elsewhere) that implements Dawson's AlmostEqual2sComplement function. If that's the sort of behaviour you want, you'll have to implement it yourself. (In which case, rather than using Dawson's clever bitwise hacks you'd probably do better to use more conventional tests of the form if abs(a-b) <= eps1*(abs(a)+abs(b)) + eps2 or similar. To get Dawson-like behaviour you might say something like if abs(a-b) <= eps*max(EPS,abs(a),abs(b)) for some small fixed EPS; this isn't exactly the same as Dawson, but it's similar in spirit.
If you want to use it in testing/TDD context, I'd say this is a standard way:
from nose.tools import assert_almost_equals
assert_almost_equals(x, y, places=7) # The default is 7
In terms of absolute error, you can just check
if abs(a - b) <= error:
print("Almost equal")
Some information of why float act weird in Python:
Python 3 Tutorial 03 - if-else, logical operators and top beginner mistakes
You can also use math.isclose for relative errors.
This is useful for the case where you want to make sure two numbers are the same 'up to precision', and there isn't any need to specify the tolerance:
Find minimum precision of the two numbers
Round both of them to minimum precision and compare
def isclose(a, b):
astr = str(a)
aprec = len(astr.split('.')[1]) if '.' in astr else 0
bstr = str(b)
bprec = len(bstr.split('.')[1]) if '.' in bstr else 0
prec = min(aprec, bprec)
return round(a, prec) == round(b, prec)
As written, it only works for numbers without the 'e' in their string representation (meaning 0.9999999999995e-4 < number <= 0.9999999999995e11)
Example:
>>> isclose(10.0, 10.049)
True
>>> isclose(10.0, 10.05)
False
For some of the cases where you can affect the source number representation, you can represent them as fractions instead of floats, using integer numerator and denominator. That way you can have exact comparisons.
See Fraction from fractions module for details.
I liked Sesquipedal's suggestion, but with modification (a special use case when both values are 0 returns False). In my case, I was on Python 2.7 and just used a simple function:
if f1 ==0 and f2 == 0:
return True
else:
return abs(f1-f2) < tol*max(abs(f1),abs(f2))
If you want to do it in a testing or TDD context using the pytest package, here's how:
import pytest
PRECISION = 1e-3
def assert_almost_equal():
obtained_value = 99.99
expected_value = 100.00
assert obtained_value == pytest.approx(expected_value, PRECISION)
I found the following comparison helpful:
str(f1) == str(f2)
To compare up to a given decimal without atol/rtol:
def almost_equal(a, b, decimal=6):
return '{0:.{1}f}'.format(a, decimal) == '{0:.{1}f}'.format(b, decimal)
print(almost_equal(0.0, 0.0001, decimal=5)) # False
print(almost_equal(0.0, 0.0001, decimal=4)) # True
This maybe is a bit ugly hack, but it works pretty well when you don't need more than the default float precision (about 11 decimals).
The round_to function uses the format method from the built-in str class to round up the float to a string that represents the float with the number of decimals needed, and then applies the eval built-in function to the rounded float string to get back to the float numeric type.
The is_close function just applies a simple conditional to the rounded up float.
def round_to(float_num, prec):
return eval("'{:." + str(int(prec)) + "f}'.format(" + str(float_num) + ")")
def is_close(float_a, float_b, prec):
if round_to(float_a, prec) == round_to(float_b, prec):
return True
return False
>>>a = 10.0
10.0
>>>b = 10.0001
10.0001
>>>print is_close(a, b, prec=3)
True
>>>print is_close(a, b, prec=4)
False
Update:
As suggested by #stepehjfox, a cleaner way to build a rount_to function avoiding "eval" is using nested formatting:
def round_to(float_num, prec):
return '{:.{precision}f}'.format(float_num, precision=prec)
Following the same idea, the code can be even simpler using the great new f-strings (Python 3.6+):
def round_to(float_num, prec):
return f'{float_num:.{prec}f}'
So, we could even wrap it up all in one simple and clean 'is_close' function:
def is_close(a, b, prec):
return f'{a:.{prec}f}' == f'{b:.{prec}f}'
If you want to compare floats, the options above are great, but in my case, I ended up using Enum's, since I only had few valid floats my use case was accepting.
from enum import Enum
class HolidayMultipliers(Enum):
EMPLOYED_LESS_THAN_YEAR = 2.0
EMPLOYED_MORE_THAN_YEAR = 2.5
Then running:
testable_value = 2.0
HolidayMultipliers(testable_value)
If the float is valid, it's fine, but otherwise it will just throw an ValueError.
Use == is a simple good way, if you don't care about tolerance precisely.
# Python 3.8.5
>>> 1.0000000000001 == 1
False
>>> 1.00000000000001 == 1
True
But watch out for 0:
>>> 0 == 0.00000000000000000000000000000000000000000001
False
The 0 is always the zero.
Use math.isclose if you want to control the tolerance.
The default a == b is equivalent to math.isclose(a, b, rel_tol=1e-16, abs_tol=0).
If you still want to use == with a self-defined tolerance:
>>> class MyFloat(float):
def __eq__(self, another):
return math.isclose(self, another, rel_tol=0, abs_tol=0.001)
>>> a == MyFloat(0)
>>> a
0.0
>>> a == 0.001
True
So far, I didn't find anywhere to config it globally for float. Besides, mock is also not working for float.__eq__.
It's well known that comparing floats for equality is a little fiddly due to rounding and precision issues.
For example: Comparing Floating Point Numbers, 2012 Edition
What is the recommended way to deal with this in Python?
Is a standard library function for this somewhere?
Python 3.5 adds the math.isclose and cmath.isclose functions as described in PEP 485.
If you're using an earlier version of Python, the equivalent function is given in the documentation.
def isclose(a, b, rel_tol=1e-09, abs_tol=0.0):
return abs(a-b) <= max(rel_tol * max(abs(a), abs(b)), abs_tol)
rel_tol is a relative tolerance, it is multiplied by the greater of the magnitudes of the two arguments; as the values get larger, so does the allowed difference between them while still considering them equal.
abs_tol is an absolute tolerance that is applied as-is in all cases. If the difference is less than either of those tolerances, the values are considered equal.
Something as simple as the following may be good enough:
return abs(f1 - f2) <= allowed_error
I would agree that Gareth's answer is probably most appropriate as a lightweight function/solution.
But I thought it would be helpful to note that if you are using NumPy or are considering it, there is a packaged function for this.
numpy.isclose(a, b, rtol=1e-05, atol=1e-08, equal_nan=False)
A little disclaimer though: installing NumPy can be a non-trivial experience depending on your platform.
Use Python's decimal module, which provides the Decimal class.
From the comments:
It is worth noting that if you're
doing math-heavy work and you don't
absolutely need the precision from
decimal, this can really bog things
down. Floats are way, way faster to
deal with, but imprecise. Decimals are
extremely precise but slow.
The common wisdom that floating-point numbers cannot be compared for equality is inaccurate. Floating-point numbers are no different from integers: If you evaluate "a == b", you will get true if they are identical numbers and false otherwise (with the understanding that two NaNs are of course not identical numbers).
The actual problem is this: If I have done some calculations and am not sure the two numbers I have to compare are exactly correct, then what? This problem is the same for floating-point as it is for integers. If you evaluate the integer expression "7/3*3", it will not compare equal to "7*3/3".
So suppose we asked "How do I compare integers for equality?" in such a situation. There is no single answer; what you should do depends on the specific situation, notably what sort of errors you have and what you want to achieve.
Here are some possible choices.
If you want to get a "true" result if the mathematically exact numbers would be equal, then you might try to use the properties of the calculations you perform to prove that you get the same errors in the two numbers. If that is feasible, and you compare two numbers that result from expressions that would give equal numbers if computed exactly, then you will get "true" from the comparison. Another approach is that you might analyze the properties of the calculations and prove that the error never exceeds a certain amount, perhaps an absolute amount or an amount relative to one of the inputs or one of the outputs. In that case, you can ask whether the two calculated numbers differ by at most that amount, and return "true" if they are within the interval. If you cannot prove an error bound, you might guess and hope for the best. One way of guessing is to evaluate many random samples and see what sort of distribution you get in the results.
Of course, since we only set the requirement that you get "true" if the mathematically exact results are equal, we left open the possibility that you get "true" even if they are unequal. (In fact, we can satisfy the requirement by always returning "true". This makes the calculation simple but is generally undesirable, so I will discuss improving the situation below.)
If you want to get a "false" result if the mathematically exact numbers would be unequal, you need to prove that your evaluation of the numbers yields different numbers if the mathematically exact numbers would be unequal. This may be impossible for practical purposes in many common situations. So let us consider an alternative.
A useful requirement might be that we get a "false" result if the mathematically exact numbers differ by more than a certain amount. For example, perhaps we are going to calculate where a ball thrown in a computer game traveled, and we want to know whether it struck a bat. In this case, we certainly want to get "true" if the ball strikes the bat, and we want to get "false" if the ball is far from the bat, and we can accept an incorrect "true" answer if the ball in a mathematically exact simulation missed the bat but is within a millimeter of hitting the bat. In that case, we need to prove (or guess/estimate) that our calculation of the ball's position and the bat's position have a combined error of at most one millimeter (for all positions of interest). This would allow us to always return "false" if the ball and bat are more than a millimeter apart, to return "true" if they touch, and to return "true" if they are close enough to be acceptable.
So, how you decide what to return when comparing floating-point numbers depends very much on your specific situation.
As to how you go about proving error bounds for calculations, that can be a complicated subject. Any floating-point implementation using the IEEE 754 standard in round-to-nearest mode returns the floating-point number nearest to the exact result for any basic operation (notably multiplication, division, addition, subtraction, square root). (In case of tie, round so the low bit is even.) (Be particularly careful about square root and division; your language implementation might use methods that do not conform to IEEE 754 for those.) Because of this requirement, we know the error in a single result is at most 1/2 of the value of the least significant bit. (If it were more, the rounding would have gone to a different number that is within 1/2 the value.)
Going on from there gets substantially more complicated; the next step is performing an operation where one of the inputs already has some error. For simple expressions, these errors can be followed through the calculations to reach a bound on the final error. In practice, this is only done in a few situations, such as working on a high-quality mathematics library. And, of course, you need precise control over exactly which operations are performed. High-level languages often give the compiler a lot of slack, so you might not know in which order operations are performed.
There is much more that could be (and is) written about this topic, but I have to stop there. In summary, the answer is: There is no library routine for this comparison because there is no single solution that fits most needs that is worth putting into a library routine. (If comparing with a relative or absolute error interval suffices for you, you can do it simply without a library routine.)
math.isclose() has been added to Python 3.5 for that (source code). Here is a port of it to Python 2. It's difference from one-liner of Mark Ransom is that it can handle "inf" and "-inf" properly.
def isclose(a, b, rel_tol=1e-09, abs_tol=0.0):
'''
Python 2 implementation of Python 3.5 math.isclose()
https://github.com/python/cpython/blob/v3.5.10/Modules/mathmodule.c#L1993
'''
# sanity check on the inputs
if rel_tol < 0 or abs_tol < 0:
raise ValueError("tolerances must be non-negative")
# short circuit exact equality -- needed to catch two infinities of
# the same sign. And perhaps speeds things up a bit sometimes.
if a == b:
return True
# This catches the case of two infinities of opposite sign, or
# one infinity and one finite number. Two infinities of opposite
# sign would otherwise have an infinite relative tolerance.
# Two infinities of the same sign are caught by the equality check
# above.
if math.isinf(a) or math.isinf(b):
return False
# now do the regular computation
# this is essentially the "weak" test from the Boost library
diff = math.fabs(b - a)
result = (((diff <= math.fabs(rel_tol * b)) or
(diff <= math.fabs(rel_tol * a))) or
(diff <= abs_tol))
return result
I'm not aware of anything in the Python standard library (or elsewhere) that implements Dawson's AlmostEqual2sComplement function. If that's the sort of behaviour you want, you'll have to implement it yourself. (In which case, rather than using Dawson's clever bitwise hacks you'd probably do better to use more conventional tests of the form if abs(a-b) <= eps1*(abs(a)+abs(b)) + eps2 or similar. To get Dawson-like behaviour you might say something like if abs(a-b) <= eps*max(EPS,abs(a),abs(b)) for some small fixed EPS; this isn't exactly the same as Dawson, but it's similar in spirit.
If you want to use it in testing/TDD context, I'd say this is a standard way:
from nose.tools import assert_almost_equals
assert_almost_equals(x, y, places=7) # The default is 7
In terms of absolute error, you can just check
if abs(a - b) <= error:
print("Almost equal")
Some information of why float act weird in Python:
Python 3 Tutorial 03 - if-else, logical operators and top beginner mistakes
You can also use math.isclose for relative errors.
This is useful for the case where you want to make sure two numbers are the same 'up to precision', and there isn't any need to specify the tolerance:
Find minimum precision of the two numbers
Round both of them to minimum precision and compare
def isclose(a, b):
astr = str(a)
aprec = len(astr.split('.')[1]) if '.' in astr else 0
bstr = str(b)
bprec = len(bstr.split('.')[1]) if '.' in bstr else 0
prec = min(aprec, bprec)
return round(a, prec) == round(b, prec)
As written, it only works for numbers without the 'e' in their string representation (meaning 0.9999999999995e-4 < number <= 0.9999999999995e11)
Example:
>>> isclose(10.0, 10.049)
True
>>> isclose(10.0, 10.05)
False
For some of the cases where you can affect the source number representation, you can represent them as fractions instead of floats, using integer numerator and denominator. That way you can have exact comparisons.
See Fraction from fractions module for details.
I liked Sesquipedal's suggestion, but with modification (a special use case when both values are 0 returns False). In my case, I was on Python 2.7 and just used a simple function:
if f1 ==0 and f2 == 0:
return True
else:
return abs(f1-f2) < tol*max(abs(f1),abs(f2))
If you want to do it in a testing or TDD context using the pytest package, here's how:
import pytest
PRECISION = 1e-3
def assert_almost_equal():
obtained_value = 99.99
expected_value = 100.00
assert obtained_value == pytest.approx(expected_value, PRECISION)
I found the following comparison helpful:
str(f1) == str(f2)
To compare up to a given decimal without atol/rtol:
def almost_equal(a, b, decimal=6):
return '{0:.{1}f}'.format(a, decimal) == '{0:.{1}f}'.format(b, decimal)
print(almost_equal(0.0, 0.0001, decimal=5)) # False
print(almost_equal(0.0, 0.0001, decimal=4)) # True
This maybe is a bit ugly hack, but it works pretty well when you don't need more than the default float precision (about 11 decimals).
The round_to function uses the format method from the built-in str class to round up the float to a string that represents the float with the number of decimals needed, and then applies the eval built-in function to the rounded float string to get back to the float numeric type.
The is_close function just applies a simple conditional to the rounded up float.
def round_to(float_num, prec):
return eval("'{:." + str(int(prec)) + "f}'.format(" + str(float_num) + ")")
def is_close(float_a, float_b, prec):
if round_to(float_a, prec) == round_to(float_b, prec):
return True
return False
>>>a = 10.0
10.0
>>>b = 10.0001
10.0001
>>>print is_close(a, b, prec=3)
True
>>>print is_close(a, b, prec=4)
False
Update:
As suggested by #stepehjfox, a cleaner way to build a rount_to function avoiding "eval" is using nested formatting:
def round_to(float_num, prec):
return '{:.{precision}f}'.format(float_num, precision=prec)
Following the same idea, the code can be even simpler using the great new f-strings (Python 3.6+):
def round_to(float_num, prec):
return f'{float_num:.{prec}f}'
So, we could even wrap it up all in one simple and clean 'is_close' function:
def is_close(a, b, prec):
return f'{a:.{prec}f}' == f'{b:.{prec}f}'
If you want to compare floats, the options above are great, but in my case, I ended up using Enum's, since I only had few valid floats my use case was accepting.
from enum import Enum
class HolidayMultipliers(Enum):
EMPLOYED_LESS_THAN_YEAR = 2.0
EMPLOYED_MORE_THAN_YEAR = 2.5
Then running:
testable_value = 2.0
HolidayMultipliers(testable_value)
If the float is valid, it's fine, but otherwise it will just throw an ValueError.
Use == is a simple good way, if you don't care about tolerance precisely.
# Python 3.8.5
>>> 1.0000000000001 == 1
False
>>> 1.00000000000001 == 1
True
But watch out for 0:
>>> 0 == 0.00000000000000000000000000000000000000000001
False
The 0 is always the zero.
Use math.isclose if you want to control the tolerance.
The default a == b is equivalent to math.isclose(a, b, rel_tol=1e-16, abs_tol=0).
If you still want to use == with a self-defined tolerance:
>>> class MyFloat(float):
def __eq__(self, another):
return math.isclose(self, another, rel_tol=0, abs_tol=0.001)
>>> a == MyFloat(0)
>>> a
0.0
>>> a == 0.001
True
So far, I didn't find anywhere to config it globally for float. Besides, mock is also not working for float.__eq__.
I know that questions about rounding in python have been asked multiple times already, but the answers did not help me. I'm looking for a method that is rounding a float number half up and returns a float number. The method should also accept a parameter that defines the decimal place to round to. I wrote a method that implements this kind of rounding. However, I think it does not look elegant at all.
def round_half_up(number, dec_places):
s = str(number)
d = decimal.Decimal(s).quantize(
decimal.Decimal(10) ** -dec_places,
rounding=decimal.ROUND_HALF_UP)
return float(d)
I don't like it, that I have to convert float to a string (to avoid floating point inaccuracy) and then work with the decimal module.
Do you have any better solutions?
Edit: As pointed out in the answers below, the solution to my problem is not that obvious as correct rounding requires correct representation of numbers in the first place and this is not the case with float. So I would expect that the following code
def round_half_up(number, dec_places):
d = decimal.Decimal(number).quantize(
decimal.Decimal(10) ** -dec_places,
rounding=decimal.ROUND_HALF_UP)
return float(d)
(that differs from the code above just by the fact that the float number is directly converted into a decimal number and not to a string first) to return 2.18 when used like this: round_half_up(2.175, 2) But it doesn't because Decimal(2.175) will return Decimal('2.17499999999999982236431605997495353221893310546875'), the way the float number is represented by the computer.
Suprisingly, the first code returns 2.18 because the float number is converted to string first. It seems that the str() function conducts an implicit rounding to the number that was initially meant to be rounded. So there are two roundings taking place. Even though this is the result that I would expect, it is technically wrong.
Rounding is surprisingly hard to do right, because you have to handle floating-point calculations very carefully. If you are looking for an elegant solution (short, easy to understand), what you have like like a good starting point. To be correct, you should replace decimal.Decimal(str(number)) with creating the decimal from the number itself, which will give you a decimal version of its exact representation:
d = Decimal(number).quantize(...)...
Decimal(str(number)) effectively rounds twice, as formatting the float into the string representation performs its own rounding. This is because str(float value) won't try to print the full decimal representation of the float, it will only print enough digits to ensure that you get the same float back if you pass those exact digits to the float constructor.
If you want to retain correct rounding, but avoid depending on the big and complex decimal module, you can certainly do it, but you'll still need some way to implement the exact arithmetics needed for correct rounding. For example, you can use fractions:
import fractions, math
def round_half_up(number, dec_places=0):
sign = math.copysign(1, number)
number_exact = abs(fractions.Fraction(number))
shifted = number_exact * 10**dec_places
shifted_trunc = int(shifted)
if shifted - shifted_trunc >= fractions.Fraction(1, 2):
result = (shifted_trunc + 1) / 10**dec_places
else:
result = shifted_trunc / 10**dec_places
return sign * float(result)
assert round_half_up(1.49) == 1
assert round_half_up(1.5) == 2
assert round_half_up(1.51) == 2
assert round_half_up(2.49) == 2
assert round_half_up(2.5) == 3
assert round_half_up(2.51) == 3
Note that the only tricky part in the above code is the precise conversion of a floating-point to a fraction, and that can be off-loaded to the as_integer_ratio() float method, which is what both decimals and fractions do internally. So if you really want to remove the dependency on fractions, you can reduce the fractional arithmetic to pure integer arithmetic; you stay within the same line count at the expense of some legibility:
def round_half_up(number, dec_places=0):
sign = math.copysign(1, number)
exact = abs(number).as_integer_ratio()
shifted = (exact[0] * 10**dec_places), exact[1]
shifted_trunc = shifted[0] // shifted[1]
difference = (shifted[0] - shifted_trunc * shifted[1]), shifted[1]
if difference[0] * 2 >= difference[1]: # difference >= 1/2
shifted_trunc += 1
return sign * (shifted_trunc / 10**dec_places)
Note that testing these functions brings to spotlight the approximations performed when creating floating-point numbers. For example, print(round_half_up(2.175, 2)) prints 2.17 because the decimal number 2.175 cannot be represented exactly in binary, so it is replaced by an approximation that happens to be slightly smaller than the 2.175 decimal. The function receives that value, finds it smaller than the actual fraction corresponding to the 2.175 decimal, and decides to round it down. This is not a quirk of the implementation; the behavior derives from properties of floating-point numbers and is also present in the round built-in of Python 3 and 2.
I don't like it, that I have to convert float to a string (to avoid
floating point inaccuracy) and then work with the decimal module. Do
you have any better solutions?
Yes; use Decimal to represent your numbers throughout your whole program, if you need to represent numbers such as 2.675 exactly and have them round to 2.68 instead of 2.67.
There is no other way. The floating point number which is shown on your screen as 2.675 is not the real number 2.675; in fact, it is very slightly less than 2.675, which is why it gets rounded down to 2.67:
>>> 2.675 - 2
0.6749999999999998
It only shows in string form as '2.675' because that happens to be the shortest string such that float(s) == 2.6749999999999998. Note that this longer representation (with lots of 9s) isn't exact either.
However you write your rounding function, it is not possible for my_round(2.675, 2) to round up to 2.68 and also for my_round(2 + 0.6749999999999998, 2) to round down to 2.67; because the inputs are actually the same floating point number.
So if your number 2.675 ever gets converted to a float and back again, you have already lost the information about whether it should round up or down. The solution is not to make it float in the first place.
After trying for a very long time to produce an elegant one-line function, I ended up getting something that is comparable to a dictionary in size.
I would say the simplest way to do this is just to
def round_half_up(inp,dec_places):
return round(inp+0.0000001,dec_places)
i would acknowledge that this is not accurate in every cases, but should work if you just want a simple quick workaround.
It's well known that comparing floats for equality is a little fiddly due to rounding and precision issues.
For example: Comparing Floating Point Numbers, 2012 Edition
What is the recommended way to deal with this in Python?
Is a standard library function for this somewhere?
Python 3.5 adds the math.isclose and cmath.isclose functions as described in PEP 485.
If you're using an earlier version of Python, the equivalent function is given in the documentation.
def isclose(a, b, rel_tol=1e-09, abs_tol=0.0):
return abs(a-b) <= max(rel_tol * max(abs(a), abs(b)), abs_tol)
rel_tol is a relative tolerance, it is multiplied by the greater of the magnitudes of the two arguments; as the values get larger, so does the allowed difference between them while still considering them equal.
abs_tol is an absolute tolerance that is applied as-is in all cases. If the difference is less than either of those tolerances, the values are considered equal.
Something as simple as the following may be good enough:
return abs(f1 - f2) <= allowed_error
I would agree that Gareth's answer is probably most appropriate as a lightweight function/solution.
But I thought it would be helpful to note that if you are using NumPy or are considering it, there is a packaged function for this.
numpy.isclose(a, b, rtol=1e-05, atol=1e-08, equal_nan=False)
A little disclaimer though: installing NumPy can be a non-trivial experience depending on your platform.
Use Python's decimal module, which provides the Decimal class.
From the comments:
It is worth noting that if you're
doing math-heavy work and you don't
absolutely need the precision from
decimal, this can really bog things
down. Floats are way, way faster to
deal with, but imprecise. Decimals are
extremely precise but slow.
The common wisdom that floating-point numbers cannot be compared for equality is inaccurate. Floating-point numbers are no different from integers: If you evaluate "a == b", you will get true if they are identical numbers and false otherwise (with the understanding that two NaNs are of course not identical numbers).
The actual problem is this: If I have done some calculations and am not sure the two numbers I have to compare are exactly correct, then what? This problem is the same for floating-point as it is for integers. If you evaluate the integer expression "7/3*3", it will not compare equal to "7*3/3".
So suppose we asked "How do I compare integers for equality?" in such a situation. There is no single answer; what you should do depends on the specific situation, notably what sort of errors you have and what you want to achieve.
Here are some possible choices.
If you want to get a "true" result if the mathematically exact numbers would be equal, then you might try to use the properties of the calculations you perform to prove that you get the same errors in the two numbers. If that is feasible, and you compare two numbers that result from expressions that would give equal numbers if computed exactly, then you will get "true" from the comparison. Another approach is that you might analyze the properties of the calculations and prove that the error never exceeds a certain amount, perhaps an absolute amount or an amount relative to one of the inputs or one of the outputs. In that case, you can ask whether the two calculated numbers differ by at most that amount, and return "true" if they are within the interval. If you cannot prove an error bound, you might guess and hope for the best. One way of guessing is to evaluate many random samples and see what sort of distribution you get in the results.
Of course, since we only set the requirement that you get "true" if the mathematically exact results are equal, we left open the possibility that you get "true" even if they are unequal. (In fact, we can satisfy the requirement by always returning "true". This makes the calculation simple but is generally undesirable, so I will discuss improving the situation below.)
If you want to get a "false" result if the mathematically exact numbers would be unequal, you need to prove that your evaluation of the numbers yields different numbers if the mathematically exact numbers would be unequal. This may be impossible for practical purposes in many common situations. So let us consider an alternative.
A useful requirement might be that we get a "false" result if the mathematically exact numbers differ by more than a certain amount. For example, perhaps we are going to calculate where a ball thrown in a computer game traveled, and we want to know whether it struck a bat. In this case, we certainly want to get "true" if the ball strikes the bat, and we want to get "false" if the ball is far from the bat, and we can accept an incorrect "true" answer if the ball in a mathematically exact simulation missed the bat but is within a millimeter of hitting the bat. In that case, we need to prove (or guess/estimate) that our calculation of the ball's position and the bat's position have a combined error of at most one millimeter (for all positions of interest). This would allow us to always return "false" if the ball and bat are more than a millimeter apart, to return "true" if they touch, and to return "true" if they are close enough to be acceptable.
So, how you decide what to return when comparing floating-point numbers depends very much on your specific situation.
As to how you go about proving error bounds for calculations, that can be a complicated subject. Any floating-point implementation using the IEEE 754 standard in round-to-nearest mode returns the floating-point number nearest to the exact result for any basic operation (notably multiplication, division, addition, subtraction, square root). (In case of tie, round so the low bit is even.) (Be particularly careful about square root and division; your language implementation might use methods that do not conform to IEEE 754 for those.) Because of this requirement, we know the error in a single result is at most 1/2 of the value of the least significant bit. (If it were more, the rounding would have gone to a different number that is within 1/2 the value.)
Going on from there gets substantially more complicated; the next step is performing an operation where one of the inputs already has some error. For simple expressions, these errors can be followed through the calculations to reach a bound on the final error. In practice, this is only done in a few situations, such as working on a high-quality mathematics library. And, of course, you need precise control over exactly which operations are performed. High-level languages often give the compiler a lot of slack, so you might not know in which order operations are performed.
There is much more that could be (and is) written about this topic, but I have to stop there. In summary, the answer is: There is no library routine for this comparison because there is no single solution that fits most needs that is worth putting into a library routine. (If comparing with a relative or absolute error interval suffices for you, you can do it simply without a library routine.)
math.isclose() has been added to Python 3.5 for that (source code). Here is a port of it to Python 2. It's difference from one-liner of Mark Ransom is that it can handle "inf" and "-inf" properly.
def isclose(a, b, rel_tol=1e-09, abs_tol=0.0):
'''
Python 2 implementation of Python 3.5 math.isclose()
https://github.com/python/cpython/blob/v3.5.10/Modules/mathmodule.c#L1993
'''
# sanity check on the inputs
if rel_tol < 0 or abs_tol < 0:
raise ValueError("tolerances must be non-negative")
# short circuit exact equality -- needed to catch two infinities of
# the same sign. And perhaps speeds things up a bit sometimes.
if a == b:
return True
# This catches the case of two infinities of opposite sign, or
# one infinity and one finite number. Two infinities of opposite
# sign would otherwise have an infinite relative tolerance.
# Two infinities of the same sign are caught by the equality check
# above.
if math.isinf(a) or math.isinf(b):
return False
# now do the regular computation
# this is essentially the "weak" test from the Boost library
diff = math.fabs(b - a)
result = (((diff <= math.fabs(rel_tol * b)) or
(diff <= math.fabs(rel_tol * a))) or
(diff <= abs_tol))
return result
I'm not aware of anything in the Python standard library (or elsewhere) that implements Dawson's AlmostEqual2sComplement function. If that's the sort of behaviour you want, you'll have to implement it yourself. (In which case, rather than using Dawson's clever bitwise hacks you'd probably do better to use more conventional tests of the form if abs(a-b) <= eps1*(abs(a)+abs(b)) + eps2 or similar. To get Dawson-like behaviour you might say something like if abs(a-b) <= eps*max(EPS,abs(a),abs(b)) for some small fixed EPS; this isn't exactly the same as Dawson, but it's similar in spirit.
If you want to use it in testing/TDD context, I'd say this is a standard way:
from nose.tools import assert_almost_equals
assert_almost_equals(x, y, places=7) # The default is 7
In terms of absolute error, you can just check
if abs(a - b) <= error:
print("Almost equal")
Some information of why float act weird in Python:
Python 3 Tutorial 03 - if-else, logical operators and top beginner mistakes
You can also use math.isclose for relative errors.
This is useful for the case where you want to make sure two numbers are the same 'up to precision', and there isn't any need to specify the tolerance:
Find minimum precision of the two numbers
Round both of them to minimum precision and compare
def isclose(a, b):
astr = str(a)
aprec = len(astr.split('.')[1]) if '.' in astr else 0
bstr = str(b)
bprec = len(bstr.split('.')[1]) if '.' in bstr else 0
prec = min(aprec, bprec)
return round(a, prec) == round(b, prec)
As written, it only works for numbers without the 'e' in their string representation (meaning 0.9999999999995e-4 < number <= 0.9999999999995e11)
Example:
>>> isclose(10.0, 10.049)
True
>>> isclose(10.0, 10.05)
False
For some of the cases where you can affect the source number representation, you can represent them as fractions instead of floats, using integer numerator and denominator. That way you can have exact comparisons.
See Fraction from fractions module for details.
I liked Sesquipedal's suggestion, but with modification (a special use case when both values are 0 returns False). In my case, I was on Python 2.7 and just used a simple function:
if f1 ==0 and f2 == 0:
return True
else:
return abs(f1-f2) < tol*max(abs(f1),abs(f2))
If you want to do it in a testing or TDD context using the pytest package, here's how:
import pytest
PRECISION = 1e-3
def assert_almost_equal():
obtained_value = 99.99
expected_value = 100.00
assert obtained_value == pytest.approx(expected_value, PRECISION)
I found the following comparison helpful:
str(f1) == str(f2)
To compare up to a given decimal without atol/rtol:
def almost_equal(a, b, decimal=6):
return '{0:.{1}f}'.format(a, decimal) == '{0:.{1}f}'.format(b, decimal)
print(almost_equal(0.0, 0.0001, decimal=5)) # False
print(almost_equal(0.0, 0.0001, decimal=4)) # True
This maybe is a bit ugly hack, but it works pretty well when you don't need more than the default float precision (about 11 decimals).
The round_to function uses the format method from the built-in str class to round up the float to a string that represents the float with the number of decimals needed, and then applies the eval built-in function to the rounded float string to get back to the float numeric type.
The is_close function just applies a simple conditional to the rounded up float.
def round_to(float_num, prec):
return eval("'{:." + str(int(prec)) + "f}'.format(" + str(float_num) + ")")
def is_close(float_a, float_b, prec):
if round_to(float_a, prec) == round_to(float_b, prec):
return True
return False
>>>a = 10.0
10.0
>>>b = 10.0001
10.0001
>>>print is_close(a, b, prec=3)
True
>>>print is_close(a, b, prec=4)
False
Update:
As suggested by #stepehjfox, a cleaner way to build a rount_to function avoiding "eval" is using nested formatting:
def round_to(float_num, prec):
return '{:.{precision}f}'.format(float_num, precision=prec)
Following the same idea, the code can be even simpler using the great new f-strings (Python 3.6+):
def round_to(float_num, prec):
return f'{float_num:.{prec}f}'
So, we could even wrap it up all in one simple and clean 'is_close' function:
def is_close(a, b, prec):
return f'{a:.{prec}f}' == f'{b:.{prec}f}'
If you want to compare floats, the options above are great, but in my case, I ended up using Enum's, since I only had few valid floats my use case was accepting.
from enum import Enum
class HolidayMultipliers(Enum):
EMPLOYED_LESS_THAN_YEAR = 2.0
EMPLOYED_MORE_THAN_YEAR = 2.5
Then running:
testable_value = 2.0
HolidayMultipliers(testable_value)
If the float is valid, it's fine, but otherwise it will just throw an ValueError.
Use == is a simple good way, if you don't care about tolerance precisely.
# Python 3.8.5
>>> 1.0000000000001 == 1
False
>>> 1.00000000000001 == 1
True
But watch out for 0:
>>> 0 == 0.00000000000000000000000000000000000000000001
False
The 0 is always the zero.
Use math.isclose if you want to control the tolerance.
The default a == b is equivalent to math.isclose(a, b, rel_tol=1e-16, abs_tol=0).
If you still want to use == with a self-defined tolerance:
>>> class MyFloat(float):
def __eq__(self, another):
return math.isclose(self, another, rel_tol=0, abs_tol=0.001)
>>> a == MyFloat(0)
>>> a
0.0
>>> a == 0.001
True
So far, I didn't find anywhere to config it globally for float. Besides, mock is also not working for float.__eq__.
As a part of some unit testing code that I'm writing, I wrote the following function. The purpose of which is to determine if 'a' could be rounded to 'b', regardless of how accurate 'a' or 'b' are.
def couldRoundTo(a,b):
"""Can you round a to some number of digits, such that it equals b?"""
roundEnd = len(str(b))
if a == b:
return True
for x in range(0,roundEnd):
if round(a,x) == b:
return True
return False
Here's some output from the function:
>>> couldRoundTo(3.934567892987, 3.9)
True
>>> couldRoundTo(3.934567892987, 3.3)
False
>>> couldRoundTo(3.934567892987, 3.93)
True
>>> couldRoundTo(3.934567892987, 3.94)
False
As far as I can tell, it works. However, I'm scared of relying on it considering I don't have a perfect grasp of issues concerning floating point accuracy. Could someone tell me if this is an appropriate way to implement this function? If not, how could I improve it?
Could someone tell me if this is an appropriate way to implement this function?
It depends. The given function will behave surprisingly if b isn't precisely equal to a value that would normally be obtained directly from decimal-to-binary-float conversion.
For example:
>>> print(0.1, 0.2/2, 0.3/3)
0.1 0.1 0.1
>>> couldRoundTo(0.123, 0.1)
True
>>> couldRoundTo(0.123, 0.2/2)
True
>>> couldRoundTo(0.123, 0.3/3)
False
This fails because the calculation of 0.3 / 3 results in a slightly different representation than 0.1 and 0.2 / 2 (and round(0.123, 1)).
If not, how could I improve it?
Rule of thumb: if your calculation specifically involves decimal digits in any way, just use Decimal, to avoid all the lossy base-2 round-tripping.
In particular, Decimal includes a helper called quantize that makes this problem trivially easy:
from decimal import Decimal
def roundable(a, b):
a = Decimal(str(a))
b = Decimal(str(b))
return a.quantize(b) == b
One way to do it:
def could_round_to(a, b):
(x, y) = map(len, str(b).split('.'))
round_format = "%" + "%d.%df"%(x, y)
return round_format%a == str(b)
First, we take the number of digits before and after the decimal in x and y. Then, we construct a format such as %x.yf. Then, we supply a to the format string.
>>> "%2.2f"%123.1234
'123.12'
>>> "%2.2f"%123.1264
'123.13'
>>> "%3.2f"%000.001
'0.00'
Now, all that's left is comparing the strings.
The only point that I'm afraid of is the conversion from strings to floating points when interpreting floating-point literals (as in http://docs.python.org/reference/lexical_analysis.html#floating-point-literals). I don't know if there is any guarantee that a floating-point literal will evaluate to the floating-point number that is closest to the given string. This mentioned section is the place in the specification where I would expect such a guarantee.
For example, Java is much more specific about what to expect from a string literal. From the documentation of Double.valueOf(String):
[...] [the argument] is regarded as representing an exact decimal value in the usual "computerized scientific notation" or as an exact hexadecimal value; this exact numerical value is then conceptually converted to an "infinitely precise" binary value that is then rounded to type double by the usual round-to-nearest rule of IEEE 754 floating-point arithmetic [...]
Unless you can find such a guarantee anywhere in the Python documentation, you can be just lucky, because some earlier floating-point libraries (on which Python might rely) convert a string just to a floating-point number nearby, not to the best available.
Unfortunately, it seems to me that neither round, nor float, nor the specification for floating-point literaly give you any usable guarantee.
If you purpose is to test if round function will round to the target, then you are correct. Otherwise (what else is the purpose?) if you are in doubt , you should use decimal module