[Numpy-discussion] (Value Based Promotion) Current Behaviour
Sebastian Berg
sebastian at sipsolutions.net
Thu Jun 13 18:37:03 EDT 2019
(this may be a bit thinking out loudly...)
On Thu, 2019-06-13 at 07:30 +0200, Hameer Abbasi wrote:
> Hi Sebastian,
>
> One way to avoid an ugly lookup table and special cases is to store
> the amount of sign bits, the amount of integer/mantissa bits and the
> amount of exponent bits for each numeric style. A safe cast can only
> happen if all three are exceeded or equal. Just a thought.
>
True, although I am not sure I like it as a general solution. Within
numpy, we do not have much problems in any case, and I do not mind the
table as such.
What I am worrying more about right now is what happens with user
dtypes and the "minimal dtype" logic for them. For example, if a user
creates an int24, an python integer could probably not get cast to it
automatically (since it is a user defined type). But for the sake of
"minimal dtype" should we try to support it?
I.e. should: `user_library.int24(1) - 2**20` have machinery to convert
2**20 to an int24 instead of an int32 as "minimal type"? But if a
second user dtype does not know about int24 (say rational), it may be
an invalid "minimal dtype" for that (at least unless numpy tries to
automagically fill holes in the casting table, and that seems like too
much magic to me)?
(Another example is a masked_int8 which uses -128 to mean NA)
Similarly, fixing the hierarchy by representing 0-127 with a uint7 is
problematic since current dtypes cannot register with it or have not
done so (instead it really means "uint8 or int8"). Of course you could
do that for them, but its just another complexity hoop to jump through.
A similar thing is Marten's thought about intermediate values of int16
casting safely to float16, requiring ever more fine grained value based
logic.
Just to be clear, I think we _can_ very much live with any hypothetical
inconsistencies when it comes to "minimal dtypes" for the time being;
they seem very much irrelevant!
My issue is that I would like to figure out a good final solution that
covers them (since I doubt we will get rid of them completely).
And if that means that we cannot cache them very well (because any
value _could_ behave differently) then maybe so be it.
Even if that means that `arr + python_int` is slower then `arr +
int32(python_int)` then that is maybe fine, right now for all we know
the difference, may even be small.
For such a scalar object instead what would seem necessary is to call a
`dtype.__coerce_pyvalue__(scalar, casting="safe")`, or a
`__can_coerce_pyvalue__` method/slot. It would replace the current
`PyArray_CanCastArrayTo`, which can only handle the current hardcoded
special "minimum value" rules. And of course it also means that
resolvers need to handle such scalar(?) objects, but in many cases they
do not need more than "can cast" anyway.
Best,
Sebastian
> Best Regards,
> Hameer Abbasi
>
> > On Wednesday, Jun 12, 2019 at 9:50 PM, Sebastian Berg <
> > sebastian at sipsolutions.net> wrote:
> > On Wed, 2019-06-12 at 12:03 -0500, Sebastian Berg wrote:
> > > On Tue, 2019-06-11 at 22:08 -0400, Marten van Kerkwijk wrote:
> > > > HI Sebastian,
> > > >
> > > > Thanks for the overview! In the value-based casting, what
> > > > perhaps
> > > > surprises me most is that it is done within a kind; it would
> > > > seem
> > > > an
> > > > improvement to check whether a given integer scalar is exactly
> > > > representable in a given float (your example of 1024 in
> > > > `float16`).
> > > > If we switch to the python-only scalar values idea, I would
> > > > suggest
> > > > to abandon this. That might make dealing with things like
> > > > `Decimal`
> > > > or `Fraction` easier as well.
> > > >
> > >
> > > Yeah, one can argue that since we have this "safe casting" based
> > > approach, we should go all the way for the value based logic. I
> > > think
> > > I
> > > tend to agree, but I am not quite sure right now to be honest.
> >
> > Just realized, one issue with this is that you get much more
> > "special
> > cases" if you think of it in terms of "minimal dtype". Because
> > suddenly, not just the unsigned/signed integers such as "< 128"
> > are
> > special, but even more values require special handling. An int16
> > "minimal dtype" may or may not be castable to float16.
> >
> > For `can_cast` that does not matter much, but if we use the same
> > logic
> > for promotion things may get uglier. Although, maybe it just gets
> > uglier implementation wise and is fairly logic on the user side...
> >
> > - Sebastian
> >
> >
> > > Fractions and Decimals are very interesting in that they raise
> > > the
> > > question what happens to user dtypes [0]. Although, you would
> > > still
> > > need a "no lower category" rule, since you do not want 1024. or
> > > 12/3
> > > be
> > > demoted to an integer.
> > >
> > > For me right now, what is most interesting is what we should do
> > > with
> > > ufunc calls, and if we can simplify them. I feel right now we
> > > have to
> > > types of ufuncs:
> > >
> > > 1. Ufuncs which use a "common type", where we can find the
> > > minimal
> > > type
> > > before dispatching.
> > >
> > > 2. More complex ufuncs, for which finding the minimal type is
> > > trickier
> > > [1]. And while I could not find any weird enough ufunc, I am not
> > > sure
> > > that blind promotion is a good idea for general ufuncs.
> > >
> > > Best,
> > >
> > > Sebastian
> > >
> > >
> > > [0] A python fraction could be converted to int64/int64 or
> > > int32/int32,
> > > etc. depending on the value, in principle. If we want such things
> > > to
> > > work in principle, we need machinery (although I expect one could
> > > tag
> > > that on later).
> > > [1] It is not impossible, but we need to insert non-existing
> > > types
> > > into
> > > the type hierarchy.
> > >
> > >
> > >
> > > PS: Another interesting issue is that if we try to move away
> > > from
> > > value
> > > based casting for numpy scalars, that initial `np.asarray(...)`
> > > call
> > > may lose the information that a python integer was passed in. So
> > > to
> > > support such things, we might need a whole new machinery.
> > >
> > >
> > >
> > >
> > > > All the best,
> > > >
> > > > Marten
> > > >
> > > > On Tue, Jun 11, 2019 at 8:46 PM Sebastian Berg <
> > > > sebastian at sipsolutions.net> wrote:
> > > > > Hi all,
> > > > >
> > > > > strange, something went wrong sending that email, but in any
> > > > > case...
> > > > >
> > > > > I tried to "summarize" the current behaviour of promotion
> > > > > and
> > > > > value
> > > > > based promotion in numpy (correcting a small error in what I
> > > > > wrote
> > > > > earlier). Since it got a bit long, you can find it here
> > > > > (also
> > > > > copy
> > > > > pasted at the end):
> > > > >
> > > > > https://hackmd.io/NF7Jz3ngRVCIQLU6IZrufA
> > > > >
> > > > > Allan's document which I link in there is also very
> > > > > interesting.
> > > > > One
> > > > > thing I had not really thought about before was the problem
> > > > > of
> > > > > commutativity.
> > > > >
> > > > > I do not have any specific points I want to discuss based on
> > > > > it
> > > > > (but
> > > > > those are likely to come up again later).
> > > > >
> > > > > All the Best,
> > > > >
> > > > > Sebastian
> > > > >
> > > > >
> > > > > -----------------------------
> > > > >
> > > > > PS: Below a copy of what I wrote:
> > > > >
> > > > > ---
> > > > > title: Numpy Value Based Promotion Rules
> > > > > author: Sebastian Berg
> > > > > ---
> > > > >
> > > > >
> > > > >
> > > > > NumPy Value Based Scalar Casting and Promotion
> > > > > ==============================================
> > > > >
> > > > > This document reviews some of the behaviours of the
> > > > > promotion
> > > > > rules
> > > > > within numpy. This is especially with respect to the
> > > > > promotion of
> > > > > scalars and 0D arrays which inspect the value to decide
> > > > > casting
> > > > > and
> > > > > promotion.
> > > > >
> > > > > Other documents discussing these things:
> > > > >
> > > > > * `from numpy.testing import print_coercion_tables` prints
> > > > > the
> > > > > current promotion tables including value based promotion for
> > > > > small
> > > > > positive/negative scalars.
> > > > > * Allan Haldane's thoughts on changing casting/promotion to
> > > > > be
> > > > > more
> > > > > C-like and discussing things such as here:
> > > > >
> > > > > https://gist.github.com/ahaldane/0f5ade49730e1a5d16ff6df4303f2e76
> > > > > * Discussion around the problem of uint64 and int64 being
> > > > > promoted to
> > > > > float64: https://github.com/numpy/numpy/issues/12525 (lists
> > > > > many
> > > > > related issues).
> > > > >
> > > > >
> > > > > Nomenclature and Defintions
> > > > > ---------------------------
> > > > >
> > > > > * **dtype/type**: The data type of an array or scalar:
> > > > > `float32`,
> > > > > `float64`, `int8`, …
> > > > >
> > > > > * **Category**: A category to which the data type belongs,
> > > > > in
> > > > > this
> > > > > context these are:
> > > > > 1. boolean
> > > > > 2. integer (unsigned and signed are not split up here, but
> > > > > are
> > > > > different "kinds")
> > > > > 3. floating point and complex (not split up here but are
> > > > > different
> > > > > "kinds")
> > > > > 5. All others
> > > > >
> > > > > * **Casting**: converting from one dtype to another. There
> > > > > are
> > > > > four
> > > > > different rules of casting:
> > > > > 1. *"safe"* casting: All values are representable in the new
> > > > > data
> > > > > type. I.e. no information is lost during the conversion.
> > > > > 2. *"same kind"* casting: data loss may occur, but only
> > > > > within
> > > > > the
> > > > > same "kind". For example a float64 can be converted to
> > > > > float32
> > > > > using
> > > > > "same kind" rules, an int64 can be converted to int16. This
> > > > > is
> > > > > although
> > > > > both lose precision or even produce incorrect values. Note
> > > > > that
> > > > > "kind"
> > > > > is different from "category" in that it distinguishes
> > > > > between
> > > > > signed
> > > > > and unsigned integers.
> > > > > 4. *"unsafe"* casting: Any conversion which can be defined,
> > > > > e.g.
> > > > > floating point to integer. For promotion this is fairly
> > > > > unimportant.
> > > > > (Some conversions such as string to integer, which not even
> > > > > work
> > > > > fall
> > > > > in this category, but could also be called coercions or
> > > > > conversions.)
> > > > >
> > > > > * **Promotion**: The general process of finding a new dtype
> > > > > for
> > > > > multiple input dtypes. Will be used here to also denote any
> > > > > kind
> > > > > of
> > > > > casting/promotion done before a specific function is called.
> > > > > This
> > > > > can
> > > > > be more complex, because in rare cases a functions can for
> > > > > example
> > > > > take
> > > > > floating point numbers and integers as input at the same
> > > > > time
> > > > > (i.e.
> > > > > `np.ldexp`).
> > > > >
> > > > > * **Common dtype**: A dtype which can represent all input
> > > > > data.
> > > > > In
> > > > > general this means that all inputs can be safely cast to
> > > > > this
> > > > > dtype.
> > > > > Within numpy this is the normal and simplest form of
> > > > > promotion.
> > > > >
> > > > > * **`type1, type2 -> type3`**: Defines a promotion or
> > > > > signature.
> > > > > For
> > > > > example adding two integers: `np.int32(5) + np.int32(3)`
> > > > > gives
> > > > > `np.int32(8)`. The dtype signature for that example would
> > > > > be:
> > > > > `int32,
> > > > > int32 -> int32`. A short form for this is also `ii->i` using
> > > > > C-
> > > > > like
> > > > > type codes, this can be found for example in
> > > > > `np.ldexp.types`
> > > > > (and
> > > > > any
> > > > > numpy ufunc).
> > > > >
> > > > > * **Scalar**: A numpy or python scalar or a **0-D array**. It
> > > > > is
> > > > > important to remember that zero dimensional arrays are
> > > > > treated
> > > > > just
> > > > > like scalars with respect to casting and promotion.
> > > > >
> > > > >
> > > > > Current Situation in Numpy
> > > > > --------------------------
> > > > >
> > > > > The current situation can be understand mostly in terms of
> > > > > safe
> > > > > casting
> > > > > which is defined based on the type hierarchy and is sensitive
> > > > > to
> > > > > values
> > > > > for scalars.
> > > > >
> > > > > This safe casting based approach is in contrast for example
> > > > > to
> > > > > promotion within C or Julia, which work based on category
> > > > > first.
> > > > > For
> > > > > example `int32` cannot be safely cast to `float32`, but C or
> > > > > Julia
> > > > > will
> > > > > use `int32, float32 -> float32` as the common type/promotion
> > > > > rule
> > > > > for
> > > > > example to decide on the output dtype for addition.
> > > > >
> > > > >
> > > > > ### Python Integers and Floats
> > > > >
> > > > > Note that python integers are handled exactly like numpy
> > > > > ones.
> > > > > They
> > > > > are, however, special in that they do not have a dtype
> > > > > associated
> > > > > with
> > > > > them explicitly. Value based logic, as described here, seems
> > > > > useful
> > > > > for
> > > > > python integers and floats to allow:
> > > > > ```
> > > > > arr = np.arange(10, dtype=np.int8)
> > > > > arr += 1
> > > > > # or:
> > > > > res = arr + 1
> > > > > res.dtype == np.int8
> > > > > ```
> > > > > which ensures that no upcast (for example with higher memory
> > > > > usage)
> > > > > occurs.
> > > > >
> > > > >
> > > > > ### Safe Casting
> > > > >
> > > > > Most safe casting is clearly defined based on whether or not
> > > > > any
> > > > > possible value is representable in the ouput dtype. Within
> > > > > numpy
> > > > > there
> > > > > is currently a single exception to this rule:
> > > > > `np.can_cast(np.int64,
> > > > > np.float64, casting="safe")` is considered to be true
> > > > > although
> > > > > float64
> > > > > cannot represent some large integer values exactly. In
> > > > > contrast,
> > > > > `np.can_cast(np.int32, np.float32, casting="safe")` is
> > > > > `False`
> > > > > and
> > > > > `np.float64` would have to be used if a "safe" cast is
> > > > > desired.
> > > > >
> > > > > This exception may be one thing that should be changed,
> > > > > however,
> > > > > concurrently the promotion rules have to be adapted to keep
> > > > > doing
> > > > > the
> > > > > same thing, or a larger behaviour change decided.
> > > > >
> > > > >
> > > > > #### Scalar based rules
> > > > >
> > > > > Unlike arrays, where inspection of all values is not
> > > > > feasable,
> > > > > for
> > > > > scalars (and 0-D arrays) the value is inspected. The casting
> > > > > becomes a
> > > > > two step process:
> > > > > 1. The minimal dtype capable of holding the value is found.
> > > > > 2. The normal casting rules are applied to the new dtype.
> > > > >
> > > > > The first step uses the following rules by finding the
> > > > > minimal
> > > > > dtype
> > > > > within its category:
> > > > >
> > > > > * Boolean: Dtype is already minimal
> > > > >
> > > > > * Integers:
> > > > > Casting is possible if output can hold the value. This
> > > > > includes
> > > > > uint8(127) casting to an int8.
> > > > >
> > > > > * Floats and Complex
> > > > > Scalars can be demoted based on value, roughly this avoids
> > > > > overflows:
> > > > > ```
> > > > > float16: -65000 < value < 65000
> > > > > float32: -3.4e38 < value < 3.4e38
> > > > > float64: -1.7e308 < value < 1.7e308
> > > > > float128 (largest type, does not apply).
> > > > > ```
> > > > > For complex, the logic is simply applied to both real and
> > > > > imaginary
> > > > > part. Complex numbers cannot be downcast to floating point.
> > > > >
> > > > > * Others: Dtype is not modified.
> > > > >
> > > > >
> > > > > This two step process means that
> > > > > `np.can_cast(np.int16(1024),
> > > > > np.float16)` is `False` even though float16 is capable of
> > > > > exactly
> > > > > representing the value 1024, since value based "demotion" to
> > > > > a
> > > > > lower
> > > > > dtype is used only within each category.
> > > > >
> > > > >
> > > > >
> > > > > ### Common Type Promotion
> > > > >
> > > > > For most operations in numpy the output type is just the
> > > > > common
> > > > > type of
> > > > > the inputs, this holds for example for concatenation, as well
> > > > > as
> > > > > almost
> > > > > all math funcions (e.g. addition and multiplication have two
> > > > > identical
> > > > > inputs and need one ouput dtype). This operation is exposed
> > > > > as
> > > > > `np.result_type` which includes value based logic, and
> > > > > `np.promote_types` which only accepts dtypes as input.
> > > > >
> > > > > Normal type promotion without value based/scalar logic finds
> > > > > the
> > > > > smallest type which both inputs can cast to safely. This will
> > > > > be
> > > > > the
> > > > > largest "kind" (bool < unsigned < integer < float < complex
> > > > > <
> > > > > other).
> > > > >
> > > > > Note that type promotion is handled in a "reduce" manner
> > > > > from
> > > > > left
> > > > > to
> > > > > right. In rare cases this means it is not associatetive:
> > > > > `float32,
> > > > > uint16, int16 -> float32`, but `float32, (uint16, int16) ->
> > > > > float64`.
> > > > >
> > > > > #### Scalar based rule
> > > > >
> > > > > When there is a mix of scalars and arrays, numpy will
> > > > > usually
> > > > > allow
> > > > > the
> > > > > scalars to be handled in the same fashion as for "safe"
> > > > > casting
> > > > > rules.
> > > > >
> > > > > The rules are as follows:
> > > > >
> > > > > 1. Value based logic is only applied if the "category" of
> > > > > any
> > > > > array
> > > > > is
> > > > > larger or equal to the category of all scalars. If this is
> > > > > not
> > > > > the
> > > > > case, the typical rules are used.
> > > > > * Specifically, this means: `np.array([1, 2, 3],
> > > > > dtype=np.uint8) +
> > > > > np.float64(12.)` gives a `float64` result, because the
> > > > > `np.float64(12.)` is not considered for being demoted.
> > > > >
> > > > > 2. Promotion is applied as normally, however, instead of the
> > > > > original
> > > > > dtype, the minimal dtype is used. In the case where the
> > > > > minimal
> > > > > data
> > > > > type is unsigned (say uint8) but the value is small enough,
> > > > > the
> > > > > minimal
> > > > > type may in fact be either `uint8` or `int8` (127 can be
> > > > > both).
> > > > > This
> > > > > promotion is also applied in pairs (reduction-like) from left
> > > > > to
> > > > > right.
> > > > >
> > > > >
> > > > > ### General Promotion during Function Execution
> > > > >
> > > > > General functions (read "ufuncs" such as `np.add`) may have
> > > > > a
> > > > > specific
> > > > > dtype signature which is (for most dtypes) stored e.g. as
> > > > > `np.add.types`. For many of these functions the common type
> > > > > promotion
> > > > > is used unchanged.
> > > > >
> > > > > However, some functions will employ a slightly different
> > > > > method
> > > > > (which
> > > > > should be equivalent in most cases). They will loop through
> > > > > all
> > > > > loops
> > > > > listed in `np.add.types` in order and find the first one to
> > > > > which
> > > > > all
> > > > > inputs can be safely cast:
> > > > > ```
> > > > > np.divide.types = ['ee->e', 'ff->f', 'dd->d', ...]
> > > > > ```
> > > > > Thus, `np.divide(np.int16(4), np.float16(3)` will refuse the
> > > > > first
> > > > > `float16, float16 -> float16` (`'ee->e'`) loop because
> > > > > `int16`
> > > > > cannot
> > > > > be cast safely, and then pick the float32 (`'ff->f'`) one.
> > > > >
> > > > > For simple functions, which commonly have two identical
> > > > > inputs,
> > > > > this
> > > > > should be identical, since normally a clear order exists for
> > > > > the
> > > > > dtypes
> > > > > (it does require checking int8 before uint8, etc.).
> > > > >
> > > > > #### Scalar based rule
> > > > >
> > > > > When scalars are involved, the "safe" cast logic based on
> > > > > values
> > > > > is
> > > > > applied *if and only if* rule 1. applies as before: That is
> > > > > there
> > > > > must
> > > > > be an array with a higher or equal category as all of the
> > > > > scalars.
> > > > >
> > > > > In the above `np.divide` example, this means that
> > > > > `np.divide(np.int16(4), np.array([3], dtype=np.float16))`
> > > > > *will*
> > > > > use
> > > > > the `'ee->e'` loop, because the scalar `4` is of a lower or
> > > > > equal
> > > > > category than the array (integer <= float or complex). While
> > > > > checking,
> > > > > 4 is found to be safely castable to float16, since `(u)int8`
> > > > > is
> > > > > sufficient to hold 4 and that can be safely cast to
> > > > > `float16`.
> > > > > However, `np.divide(np.int16(4), np.int16(3))` would use
> > > > > `float32`
> > > > > because both are scalars and thus value based logic is not
> > > > > used
> > > > > (Note
> > > > > that in reality numpy forces double output for an all
> > > > > integer
> > > > > input
> > > > > in
> > > > > divide).
> > > > >
> > > > > In it is possible for ufuncs to have mixed type signatures
> > > > > (this
> > > > > is
> > > > > very rare within numy) and arbitrary inputs. In this case,
> > > > > in
> > > > > principle, the question is whether or not a clear ordering
> > > > > exists
> > > > > and
> > > > > if the rule of using value based logic is always clear. This
> > > > > is
> > > > > rather
> > > > > academical (I could not find any such function in numpy or
> > > > > `scipy.special` [^scipy-ufuncs]). But consider:
> > > > > ```
> > > > > imaginary_ufunc.types:
> > > > > int32, float32 -> int32, float32
> > > > > int64, float32 -> int64, float32
> > > > > ...
> > > > > ```
> > > > > it is not clear that `np.int64(5) + np.float32(3.)` should
> > > > > be
> > > > > able
> > > > > to
> > > > > demote the `5`. This is very theoretical of course
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > Footnotes
> > > > > ---------
> > > > >
> > > > > [^scipy-ufuncs]: See for example these functions:
> > > > > ```python
> > > > > import scipy.special
> > > > > for n, func in scipy.special.__dict__.items():
> > > > > if not isinstance(func, np.ufunc):
> > > > > continue
> > > > >
> > > > > if func.nin == 1:
> > > > > # a single input is not interesting
> > > > > continue
> > > > >
> > > > > # check if the signature is not uniform
> > > > > for types in func.types:
> > > > > if len(set(types[:func.nin])) != 1:
> > > > > break
> > > > > else:
> > > > > continue
> > > > > print(func, func.types)
> > > > > ```
> > > > > _______________________________________________
> > > > > NumPy-Discussion mailing list
> > > > > NumPy-Discussion at python.org
> > > > > https://mail.python.org/mailman/listinfo/numpy-discussion
> > > >
> > > > _______________________________________________
> > > > NumPy-Discussion mailing list
> > > > NumPy-Discussion at python.org
> > > > https://mail.python.org/mailman/listinfo/numpy-discussion
> > >
> > > _______________________________________________
> > > NumPy-Discussion mailing list
> > > NumPy-Discussion at python.org
> > > https://mail.python.org/mailman/listinfo/numpy-discussion
> > _______________________________________________
> > NumPy-Discussion mailing list
> > NumPy-Discussion at python.org
> > https://mail.python.org/mailman/listinfo/numpy-discussion
>
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at python.org
> https://mail.python.org/mailman/listinfo/numpy-discussion
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: This is a digitally signed message part
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20190613/2a256968/attachment-0001.sig>
More information about the NumPy-Discussion
mailing list