Perhaps worth mentioning that we've discussed this sort of API before, in https://github.com/numpy/numpy/pull/11897. Under that proposal, the api would be something like: * `copy=True` - always copy, like it is today * `copy=False` - copy if needed, like it is today * `copy=np.never_copy` - never copy, throw an exception if not possible I think the discussion stalled on the precise spelling of the third option. `__array__` was not discussed there, but it seems like adding the `copy` argument to `__array__` would be a perfectly reasonable extension. Eric On Fri, 24 Apr 2020 at 03:00, Juan Nunez-Iglesias wrote:

Hi everyone,

One bit of expressivity we would miss is “copy if necessary, but otherwise

...

don’t bother”, but there are workarounds to this.

After a side discussion with Stéfan van der Walt, we came up with `allow_copy=True`, which would express to the downstream library that we don’t mind waiting, but that zero-copy would also be ok.

This sounds like the sort of thing that is use case driven. If enough projects want to use it, then I have no objections to adding the keyword. OTOH, we need to be careful about adding too many interoperability tricks as they complicate the code and makes it hard for folks to determine the best solution. Interoperability is a hot topic and we need to be careful not put too leave behind too many experiments in the NumPy code. Do you have any other ideas of how to achieve the same effect?

Personally, I don’t have any other ideas, but would be happy to hear some!

My view regarding API/experiment creep is that `__array__` is the oldest and most basic of all the interop tricks and that this can be safely maintained for future generations. Currently it only takes `dtype=` as a keyword argument, so it is a very lean API. I think this particular use case is very natural and I’ve encountered the reluctance to implicitly copy twice, so I expect it is reasonably common.

Regarding difficulty in determining the best solution, I would be happy to contribute to the dispatch basics guide together with the new kwarg. I agree that the protocols are getting quite numerous and I couldn’t find a single place that gathers all the best practices together. But, to reiterate my point: `__array__` is the simplest of these and I think this keyword is pretty safe to add.

For ease of discussion, here are the API options discussed so far, as well as a few extra that I don’t like but might trigger other ideas:

np.asarray(my_duck_array, allow_copy=True) # default is False, or None -> leave it to the duck array to decide np.asarray(my_duck_array, copy=True) # always copies, but, if supported by the duck array, defers to it for the copy np.asarray(my_duck_array, copy=‘allow’) # could take values ‘allow’, ‘force’, ’no’, True(=‘force’), False(=’no’) np.asarray(my_duck_array, force_copy=False, allow_copy=True) # separate concepts, but unclear what force_copy=True, allow_copy=False means! np.asarray(my_duck_array, force=True)

Juan. _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

On Fri, Apr 24, 2020 at 6:31 AM Sebastian Berg wrote:

One thing to note is that `__array__` is actually asked to return a copy AFAIK.

The documentation on __array__ seems to quite limited, unfortunately. The most I can find are a few sentences here: https://numpy.org/doc/stable/reference/arrays.classes.html#numpy.class.__arr... I don't see anything about returning copies. My interpretation has always been that __array__ can return either a copy or a view, like the np.asarray() constructor.

I doubt it always does, but if it does not I assume the object should and could provide `__array_interface__`.

Objects like xarray.DataArray and pandas.Series sometimes directly wrap NumPy arrays and sometimes don't. They both implement __array__ but not __array_inferace__. It's very obvious how to implement a "forwarding" __array__ method (just call `np.asarray()` on an argument that might implement it). I guess something similar could be done for __array_interface__, but it's not clear to me that it's right to implement __array_interface__ when doing so might require a copy.

Under that assumption, it would be an opt-out right now since NumPy allows copies by default here. Defining things along copy does seem sensible, though I do not know how it would play with some of the current array-likes choosing to refuse `__array__`.

- Sebastian

...

Eric

On Fri, 24 Apr 2020 at 03:00, Juan Nunez-Iglesias wrote:

...

Hi everyone,

One bit of expressivity we would miss is “copy if necessary, but otherwise

...

don’t bother”, but there are workarounds to this.

After a side discussion with Stéfan van der Walt, we came up with `allow_copy=True`, which would express to the downstream library that we don’t mind waiting, but that zero-copy would also be ok.

This sounds like the sort of thing that is use case driven. If enough projects want to use it, then I have no objections to adding the keyword. OTOH, we need to be careful about adding too many interoperability tricks as they complicate the code and makes it hard for folks to determine the best solution. Interoperability is a hot topic and we need to be careful not put too leave behind too many experiments in the NumPy code. Do you have any other ideas of how to achieve the same effect?

Personally, I don’t have any other ideas, but would be happy to hear some!

My view regarding API/experiment creep is that `__array__` is the oldest and most basic of all the interop tricks and that this can be safely maintained for future generations. Currently it only takes `dtype=` as a keyword argument, so it is a very lean API. I think this particular use case is very natural and I’ve encountered the reluctance to implicitly copy twice, so I expect it is reasonably common.

Regarding difficulty in determining the best solution, I would be happy to contribute to the dispatch basics guide together with the new kwarg. I agree that the protocols are getting quite numerous and I couldn’t find a single place that gathers all the best practices together. But, to reiterate my point: `__array__` is the simplest of these and I think this keyword is pretty safe to add.

For ease of discussion, here are the API options discussed so far, as well as a few extra that I don’t like but might trigger other ideas:

np.asarray(my_duck_array, allow_copy=True) # default is False, or None -> leave it to the duck array to decide np.asarray(my_duck_array, copy=True) # always copies, but, if supported by the duck array, defers to it for the copy np.asarray(my_duck_array, copy=‘allow’) # could take values ‘allow’, ‘force’, ’no’, True(=‘force’), False(=’no’) np.asarray(my_duck_array, force_copy=False, allow_copy=True) # separate concepts, but unclear what force_copy=True, allow_copy=False means! np.asarray(my_duck_array, force=True)

Juan. _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

On Fri, Apr 24, 2020 at 12:35 PM Eric Wieser wrote:

Perhaps worth mentioning that we've discussed this sort of API before, in https://github.com/numpy/numpy/pull/11897.

Under that proposal, the api would be something like:

* `copy=True` - always copy, like it is today * `copy=False` - copy if needed, like it is today * `copy=np.never_copy` - never copy, throw an exception if not possible

There's a couple of issues I see with using `copy` for __array__: - copy is already weird (False doesn't mean no), and a [bool, some_obj_or_str] keyword isn't making that better - the behavior we're talking about can do more than copying, e.g. for PyTorch it would modify the autograd graph by adding detach(), and for sparse it's not just "make a copy" (which implies doubling memory use) but it densifies which can massively blow up the memory. - I'm -1 on adding things to the main namespace (never_copy) for something that can be handled differently (like a string, or a new keyword) tl;dr a new `force` keyword would be better Cheers, Ralf

I think the discussion stalled on the precise spelling of the third option.

`__array__` was not discussed there, but it seems like adding the `copy` argument to `__array__` would be a perfectly reasonable extension.

Eric

On Fri, 24 Apr 2020 at 03:00, Juan Nunez-Iglesias wrote:

...

Hi everyone,

One bit of expressivity we would miss is “copy if necessary, but

...

otherwise don’t bother”, but there are workarounds to this.

After a side discussion with Stéfan van der Walt, we came up with `allow_copy=True`, which would express to the downstream library that we don’t mind waiting, but that zero-copy would also be ok.

This sounds like the sort of thing that is use case driven. If enough projects want to use it, then I have no objections to adding the keyword. OTOH, we need to be careful about adding too many interoperability tricks as they complicate the code and makes it hard for folks to determine the best solution. Interoperability is a hot topic and we need to be careful not put too leave behind too many experiments in the NumPy code. Do you have any other ideas of how to achieve the same effect?

Personally, I don’t have any other ideas, but would be happy to hear some!

My view regarding API/experiment creep is that `__array__` is the oldest and most basic of all the interop tricks and that this can be safely maintained for future generations. Currently it only takes `dtype=` as a keyword argument, so it is a very lean API. I think this particular use case is very natural and I’ve encountered the reluctance to implicitly copy twice, so I expect it is reasonably common.

Regarding difficulty in determining the best solution, I would be happy to contribute to the dispatch basics guide together with the new kwarg. I agree that the protocols are getting quite numerous and I couldn’t find a single place that gathers all the best practices together. But, to reiterate my point: `__array__` is the simplest of these and I think this keyword is pretty safe to add.

For ease of discussion, here are the API options discussed so far, as well as a few extra that I don’t like but might trigger other ideas:

np.asarray(my_duck_array, allow_copy=True) # default is False, or None -> leave it to the duck array to decide np.asarray(my_duck_array, copy=True) # always copies, but, if supported by the duck array, defers to it for the copy np.asarray(my_duck_array, copy=‘allow’) # could take values ‘allow’, ‘force’, ’no’, True(=‘force’), False(=’no’) np.asarray(my_duck_array, force_copy=False, allow_copy=True) # separate concepts, but unclear what force_copy=True, allow_copy=False means! np.asarray(my_duck_array, force=True)

Juan. _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

On Sat, 2020-04-25 at 10:52 -0700, Stephan Hoyer wrote:

On Sat, Apr 25, 2020 at 10:40 AM Ralf Gommers

...

wrote:

...

On Fri, Apr 24, 2020 at 12:35 PM Eric Wieser < wieser.eric+numpy@gmail.com> wrote:

...

Perhaps worth mentioning that we've discussed this sort of API before, in https://github.com/numpy/numpy/pull/11897.

Under that proposal, the api would be something like:

* `copy=True` - always copy, like it is today * `copy=False` - copy if needed, like it is today * `copy=np.never_copy` - never copy, throw an exception if not possible

There's a couple of issues I see with using `copy` for __array__: - copy is already weird (False doesn't mean no), and a [bool, some_obj_or_str] keyword isn't making that better - the behavior we're talking about can do more than copying, e.g. for PyTorch it would modify the autograd graph by adding detach(), and for sparse it's not just "make a copy" (which implies doubling memory use) but it densifies which can massively blow up the memory. - I'm -1 on adding things to the main namespace (never_copy) for something that can be handled differently (like a string, or a new keyword)

tl;dr a new `force` keyword would be better

I agree, “copy” is not a good description of this desired coercion behavior.

A new keyword argument like “force” would be much clearer.

That seems fine and practical. But, in the end it seems to me that the `force=` keyword, just means that some projects want to teach their users that: 1. `np.asarray()` can be expensive (and may always copy) 2. `np.asarray()` always loses type properties while others do not choose to teach about it. There seems very little or even no "promise" attached to either `force=True` or `force=False`. In the end, the question is whether sparse will actually want to implement `force=True` if the main reason we add is for library use. There is no difference between a visualization library or numpy. In both cases the users memory will blow up just the same. As for PyTorch, is `.detach()` even a good reason? Maybe I am missing things, but:

...

...

torch.ones(10, requires_grad=True) + np.arange(10) # RuntimeError: Can't call numpy() on Variable that requires grad. Use var.detach().numpy() instead.

So arguably, there is no type-safety concern due to `.detach()`. There is an (obvious) general loss of type information that always occurs with an `np.asarray` call. But I do not see that creating any openings for bugs here, due to the wisdom of not allowing the above operation. In fact, it actually seems much worse for for xarray, or pandas. They do support the above operation and will potentially mess up if the arange was previously an xarray with a matching index, but different order. I am very much in favor of adding such things, but I still lack a bit of clarity as to whom we would be helping? If end-users will actually use `np.asarray(..., force=True)` over special methods, then great! But I am currently not sure the type- safety argument is all that big of a point. And the performance or memory-blowup argument remains true even for visualization libraries (where the array is purely input and never output as such). But yes, "never copy" is a somewhat different extension to `__array__` and `np.asarray`. It guarantees high speed and in-place behaviour which is useful for different settings. - Sebastian

...

Cheers, Ralf

...

I think the discussion stalled on the precise spelling of the third option.

`__array__` was not discussed there, but it seems like adding the `copy` argument to `__array__` would be a perfectly reasonable extension.

Eric

On Fri, 24 Apr 2020 at 03:00, Juan Nunez-Iglesias < jni@fastmail.com> wrote:

...

Hi everyone,

One bit of expressivity we would miss is “copy if necessary, but

...

otherwise don’t bother”, but there are workarounds to this.

After a side discussion with Stéfan van der Walt, we came up with `allow_copy=True`, which would express to the downstream library that we don’t mind waiting, but that zero-copy would also be ok.

This sounds like the sort of thing that is use case driven. If enough projects want to use it, then I have no objections to adding the keyword. OTOH, we need to be careful about adding too many interoperability tricks as they complicate the code and makes it hard for folks to determine the best solution. Interoperability is a hot topic and we need to be careful not put too leave behind too many experiments in the NumPy code. Do you have any other ideas of how to achieve the same effect?

Personally, I don’t have any other ideas, but would be happy to hear some!

My view regarding API/experiment creep is that `__array__` is the oldest and most basic of all the interop tricks and that this can be safely maintained for future generations. Currently it only takes `dtype=` as a keyword argument, so it is a very lean API. I think this particular use case is very natural and I’ve encountered the reluctance to implicitly copy twice, so I expect it is reasonably common.

Regarding difficulty in determining the best solution, I would be happy to contribute to the dispatch basics guide together with the new kwarg. I agree that the protocols are getting quite numerous and I couldn’t find a single place that gathers all the best practices together. But, to reiterate my point: `__array__` is the simplest of these and I think this keyword is pretty safe to add.

For ease of discussion, here are the API options discussed so far, as well as a few extra that I don’t like but might trigger other ideas:

np.asarray(my_duck_array, allow_copy=True) # default is False, or None -> leave it to the duck array to decide np.asarray(my_duck_array, copy=True) # always copies, but, if supported by the duck array, defers to it for the copy np.asarray(my_duck_array, copy=‘allow’) # could take values ‘allow’, ‘force’, ’no’, True(=‘force’), False(=’no’) np.asarray(my_duck_array, force_copy=False, allow_copy=True) # separate concepts, but unclear what force_copy=True, allow_copy=False means! np.asarray(my_duck_array, force=True)

Juan. _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

Hi! Yes, I would advocate for a `force=` kwarg but personally I don't think it's explicit enough, but probably as explicit as can be given NumPy's API. Personally, I'd also raise a warning within PyData/Sparse and I hope it's in big bold letters in the docs in NumPy to be careful with this. Get Outlook for iOShttps://aka.ms/o0ukef ________________________________ From: NumPy-Discussion on behalf of Ralf Gommers Sent: Tuesday, April 28, 2020 11:51:13 AM To: Discussion of Numerical Python Subject: Re: [Numpy-discussion] Proposal: add `force=` or `copy=` kwarg to `__array__` interface On Mon, Apr 27, 2020 at 12:10 AM Sebastian Berg mailto:sebastian@sipsolutions.net> wrote: On Sat, 2020-04-25 at 10:52 -0700, Stephan Hoyer wrote:

On Sat, Apr 25, 2020 at 10:40 AM Ralf Gommers mailto:ralf.gommers@gmail.com

...

wrote:

...

On Fri, Apr 24, 2020 at 12:35 PM Eric Wieser < wieser.eric+numpy@gmail.commailto:wieser.eric%2Bnumpy@gmail.com> wrote:

...

Perhaps worth mentioning that we've discussed this sort of API before, in https://github.com/numpy/numpy/pull/11897.

Under that proposal, the api would be something like:

* `copy=True` - always copy, like it is today * `copy=False` - copy if needed, like it is today * `copy=np.never_copy` - never copy, throw an exception if not possible

There's a couple of issues I see with using `copy` for __array__: - copy is already weird (False doesn't mean no), and a [bool, some_obj_or_str] keyword isn't making that better - the behavior we're talking about can do more than copying, e.g. for PyTorch it would modify the autograd graph by adding detach(), and for sparse it's not just "make a copy" (which implies doubling memory use) but it densifies which can massively blow up the memory. - I'm -1 on adding things to the main namespace (never_copy) for something that can be handled differently (like a string, or a new keyword)

tl;dr a new `force` keyword would be better

I agree, “copy” is not a good description of this desired coercion behavior.

A new keyword argument like “force” would be much clearer.

That seems fine and practical. But, in the end it seems to me that the `force=` keyword, just means that some projects want to teach their users that: 1. `np.asarray()` can be expensive (and may always copy) 2. `np.asarray()` always loses type properties while others do not choose to teach about it. There seems very little or even no "promise" attached to either `force=True` or `force=False`. In the end, the question is whether sparse will actually want to implement `force=True` if the main reason we add is for library use. That's for PyData Sparse and scipy.sparse devs to answer. Maybe Hameer can answer for the former here. For SciPy that should be decided on the scipy-dev list, but my opinion would be: yes to adding __array__ that raises TypeError by default, and converts with `force=True`. There is no difference between a visualization library or numpy. In both cases the users memory will blow up just the same. As for PyTorch, is `.detach()` even a good reason? Maybe I am missing things, but:

...

...

torch.ones(10, requires_grad=True) + np.arange(10) # RuntimeError: Can't call numpy() on Variable that requires grad. Use var.detach().numpy() instead.

So arguably, there is no type-safety concern due to `.detach()`. I'm not sure what the question is here; no one mentioned type-safety. The PyTorch maintainers have already said they're fine with adding a force keyword. There is an (obvious) general loss of type information that always occurs with an `np.asarray` call. But I do not see that creating any openings for bugs here, due to the wisdom of not allowing the above operation. In fact, it actually seems much worse for for xarray, or pandas. They do support the above operation and will potentially mess up if the arange was previously an xarray with a matching index, but different order. I am very much in favor of adding such things, but I still lack a bit of clarity as to whom we would be helping? See Juan's first email. I personally am ambivalent on this proposal, but if Juan and the Napari devs really want it, that's good enough for me. Cheers, Ralf If end-users will actually use `np.asarray(..., force=True)` over special methods, then great! But I am currently not sure the type- safety argument is all that big of a point. And the performance or memory-blowup argument remains true even for visualization libraries (where the array is purely input and never output as such). But yes, "never copy" is a somewhat different extension to `__array__` and `np.asarray`. It guarantees high speed and in-place behaviour which is useful for different settings. - Sebastian

...

Cheers, Ralf

...

I think the discussion stalled on the precise spelling of the third option.

`__array__` was not discussed there, but it seems like adding the `copy` argument to `__array__` would be a perfectly reasonable extension.

Eric

On Fri, 24 Apr 2020 at 03:00, Juan Nunez-Iglesias < jni@fastmail.commailto:jni@fastmail.com> wrote:

...

Hi everyone,

One bit of expressivity we would miss is “copy if necessary, but

...

otherwise don’t bother”, but there are workarounds to this.

After a side discussion with Stéfan van der Walt, we came up with `allow_copy=True`, which would express to the downstream library that we don’t mind waiting, but that zero-copy would also be ok.

This sounds like the sort of thing that is use case driven. If enough projects want to use it, then I have no objections to adding the keyword. OTOH, we need to be careful about adding too many interoperability tricks as they complicate the code and makes it hard for folks to determine the best solution. Interoperability is a hot topic and we need to be careful not put too leave behind too many experiments in the NumPy code. Do you have any other ideas of how to achieve the same effect?

Personally, I don’t have any other ideas, but would be happy to hear some!

My view regarding API/experiment creep is that `__array__` is the oldest and most basic of all the interop tricks and that this can be safely maintained for future generations. Currently it only takes `dtype=` as a keyword argument, so it is a very lean API. I think this particular use case is very natural and I’ve encountered the reluctance to implicitly copy twice, so I expect it is reasonably common.

Regarding difficulty in determining the best solution, I would be happy to contribute to the dispatch basics guide together with the new kwarg. I agree that the protocols are getting quite numerous and I couldn’t find a single place that gathers all the best practices together. But, to reiterate my point: `__array__` is the simplest of these and I think this keyword is pretty safe to add.

For ease of discussion, here are the API options discussed so far, as well as a few extra that I don’t like but might trigger other ideas:

np.asarray(my_duck_array, allow_copy=True) # default is False, or None -> leave it to the duck array to decide np.asarray(my_duck_array, copy=True) # always copies, but, if supported by the duck array, defers to it for the copy np.asarray(my_duck_array, copy=‘allow’) # could take values ‘allow’, ‘force’, ’no’, True(=‘force’), False(=’no’) np.asarray(my_duck_array, force_copy=False, allow_copy=True) # separate concepts, but unclear what force_copy=True, allow_copy=False means! np.asarray(my_duck_array, force=True)

Juan. _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.orgmailto:NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.orgmailto:NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.orgmailto:NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.orgmailto:NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.orgmailto:NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

On Tue, 2020-04-28 at 09:58 -0500, Sebastian Berg wrote:

On Tue, 2020-04-28 at 11:51 +0200, Ralf Gommers wrote: <snip>

...

...

So arguably, there is no type-safety concern due to `.detach()`.

I'm not sure what the question is here; no one mentioned type- safety. The PyTorch maintainers have already said they're fine with adding a force keyword.

But type-safety is the reason to distinguish between:

* np.asarrau(tensor) * np.asarray(tensor, force=True)

Similar to:

* operator.index(obj) * int(obj) # convert less type-safe (strings, floats)!

I actually mentioned 3 reasons in my email:

1. Teach and Inform users (about the next two mainly) 2. Type-safety 3. Expensive conversion

And only type-safety is related to `.detach()` mentioning that there may not be clear story about the usage in that case.

(Sorry something got broken here) The question is what PyTorch's reasons are to feel `np.asarray(tensor)` should not work generally. I for one thought it was type-safety with regard to `.detach()`. And then I was surprised to realize that type-safety might not be a great reason to reject an implicit `.detach()` within `np.asarray(tensor)`. In any case, all the long talk is simply that I first want to be clear on what the concerns are why libraries reject `np.asarray(tensor)`. And then, I want to be clear that adding `force=True` will actually solves those concerns. And I was surprised myself that this became very much unclear to me. Again, one reason for it being not clear to me is half the ecosystem could potentially can just always use `force=True`. So there must be some "good usage" and some "bad usage" and I would like to know what that is. - Sebastian

(continued below)

<snip>

...

...

I am very much in favor of adding such things, but I still lack a bit of clarity as to whom we would be helping?

See Juan's first email. I personally am ambivalent on this proposal, but if Juan and the Napari devs really want it, that's good enough for me.

Of course I read it, twice, but it is only good enough for me if we actually *solve the issue*, and for that I want to know which issue we are solving :), it seems obvious, but I am not so sure...

That brings us to the other two reasons:

Teaching and Informing users:

If Napari uses `force=True` indiscriminately, it is not very clear to the user about whether or not the operation is expensive. I.e. the user can learn it is when using `np.asarray(sparse_arr)` with other libraries. But they are not notified that `napari.vis_func(sparse_arr)` might kill their computer.

So the "Teaching" part can still partially work, but it does not inform the user well anymore on whether or not a function will blow-up memory.

Expensive Conversion:

If the main reason is expensive conversions, however, than, as a library I would probably just use it for half my API, since copying from GPU to CPU will still be much faster than my own function.

Generally:

I want to help Napari, but it seems like there may be more to this, and it may be good to finish these thoughts before making a call.

E.g. Napari wants to use it, but do the array-providers want Napari to use it?

For sparse Hameer just mentioned that he still would want big warnings both during the operation and in the `np.asarray` documentation. If we put such big warnings there, we should have an idea of who we want to ignore that warning? (Napari yes, sklearn sometimes, ...?)

-> Is "whatever the library feels right" good enough?

And if the conversion still gives warnings for some array-objects, have we actually gained much?

-> Maybe we do, end-users may be happy to ignore those warnings...

The one clear use-case for `force=True` is the end-user. Just like no library uses `int(obj)`, but end-users can use it very nicely. I am happy to help the end-user in this case, but if that is the target audience we may want to _discourage_ Napari from using `force=True` and encourage sparse not to put any RuntimeWarnings on it!

- Sebastian

...

Cheers, Ralf

...

If end-users will actually use `np.asarray(..., force=True)` over special methods, then great! But I am currently not sure the type- safety argument is all that big of a point. And the performance or memory-blowup argument remains true even for visualization libraries (where the array is purely input and never output as such).

But yes, "never copy" is a somewhat different extension to `__array__` and `np.asarray`. It guarantees high speed and in-place behaviour which is useful for different settings.

- Sebastian

...

...

Cheers, Ralf

...

I think the discussion stalled on the precise spelling of the third option.

`__array__` was not discussed there, but it seems like adding the `copy` argument to `__array__` would be a perfectly reasonable extension.

Eric

On Fri, 24 Apr 2020 at 03:00, Juan Nunez-Iglesias < jni@fastmail.com> wrote:

> Hi everyone, > > One bit of expressivity we would miss is “copy if > necessary, > but > > otherwise don’t bother”, but there are workarounds to > > this. > > > > After a side discussion with Stéfan van der Walt, we came > up > with > `allow_copy=True`, which would express to the downstream > library that we > don’t mind waiting, but that zero-copy would also be ok. > > This sounds like the sort of thing that is use case > driven. > If > enough > projects want to use it, then I have no objections to > adding > the keyword. > OTOH, we need to be careful about adding too many > interoperability tricks > as they complicate the code and makes it hard for folks > to > determine the > best solution. Interoperability is a hot topic and we > need > to > be careful > not put too leave behind too many experiments in the > NumPy > code. Do you > have any other ideas of how to achieve the same effect? > > > Personally, I don’t have any other ideas, but would be > happy to > hear > some! > > My view regarding API/experiment creep is that > `__array__` > is > the oldest > and most basic of all the interop tricks and that this > can > be > safely > maintained for future generations. Currently it only > takes > `dtype=` as a > keyword argument, so it is a very lean API. I think this > particular use > case is very natural and I’ve encountered the reluctance > to > implicitly copy > twice, so I expect it is reasonably common. > > Regarding difficulty in determining the best solution, I > would > be happy > to contribute to the dispatch basics guide together with > the > new kwarg. I > agree that the protocols are getting quite numerous and I > couldn’t find a > single place that gathers all the best practices > together. > But, > to > reiterate my point: `__array__` is the simplest of these > and I > think this > keyword is pretty safe to add. > > For ease of discussion, here are the API options > discussed > so > far, as > well as a few extra that I don’t like but might trigger > other > ideas: > > np.asarray(my_duck_array, allow_copy=True) # default is > False, > or None > -> leave it to the duck array to decide > np.asarray(my_duck_array, copy=True) # always copies, > but, > if > supported > by the duck array, defers to it for the copy > np.asarray(my_duck_array, copy=‘allow’) # could take > values > ‘allow’, > ‘force’, ’no’, True(=‘force’), False(=’no’) > np.asarray(my_duck_array, force_copy=False, > allow_copy=True) # > separate > concepts, but unclear what force_copy=True, > allow_copy=False > means! > np.asarray(my_duck_array, force=True) > > Juan. > _______________________________________________ > NumPy-Discussion mailing list > NumPy-Discussion@python.org > https://mail.python.org/mailman/listinfo/numpy-discussion > _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion

Hi everyone, and thank you Ralf for carrying the flag in my absence. =D Sebastian, the *primary* motivation behind avoiding detach() in PyTorch is listed in original post of the PyTorch issue:

People not very familiar with `requires_grad` and cpu/gpu Tensors might go back and forth with numpy. For example doing pytorch -> numpy -> pytorch and backward on the last Tensor. This will backward without issue but not all the way to the first part of the code and won’t raise any error.

The PyTorch team are concerned that they will be overwhelmed with help requests if np.array() silently succeeds on a tensor with gradients. I definitely get that. Avoiding .gpu() is more straightforwardly about avoiding implicit expensive computation.

while others do not choose to teach about it. There seems very little or even no "promise" attached to either `force=True` or `force=False`.

NumPy can set a precedent through policy. The *only* reason client libraries would implement `__array__` is to play well with NumPy, so if NumPy documents that `force=True` should *always* succeed, we can expect client libraries to follow suit. At least the PyTorch devs have indicated that they would be open to this.

E.g. Napari wants to use it, but do the array-providers want Napari to use it?

As Ralf pointed out, the PyTorch devs have already agreed to it.

From the napari perspective, we'd be ok with leaving the decision on warnings to client libraries. We may or may not suppress them depending on user requests. ;) But the point is to have a way of saying "give me a NumPy array DAMMIT" without having to know about all the possible array libraries. Which are numerous and getting numerouser.

Ralf, you said you don't want warnings — even for sparse arrays? That was an area of concern for you on the PyTorch discussion side.

And if the conversion still gives warnings for some array-objects, have we actually gained much?

Yes. Hameer,

I would advocate for a `force=` kwarg but personally I don't think it's explicit enough, but probably as explicit as can be given NumPy's API.

Yeah, I agree that force is kind of vague, which is why I was looking for things like `allow_copy`. But it is hard to be general enough here: sparse requires an expensive instantiation, cupy requires copying from gpu to cpu, dask requires arbitrary computation, xarray requires information loss... I'm inclined to agree with Ralf that force= is the only generic-enough term, but I'm happy to entertain other options! Juan.

On Wed, 2020-04-29 at 05:26 -0500, Juan Nunez-Iglesias wrote:

Hi everyone, and thank you Ralf for carrying the flag in my absence. =D

Sebastian, the *primary* motivation behind avoiding detach() in PyTorch is listed in original post of the PyTorch issue:

...

People not very familiar with `requires_grad` and cpu/gpu Tensors might go back and forth with numpy. For example doing pytorch -> numpy -> pytorch and backward on the last Tensor. This will backward without issue but not all the way to the first part of the code and won’t raise any error.

The PyTorch team are concerned that they will be overwhelmed with help requests if np.array() silently succeeds on a tensor with gradients. I definitely get that.

Sorry for playing advocatus diaboli... I guess it is simply that before the end, it would be nice to have a short list with projects: * Napari, matplotlib on the "user" side * PyTorch, ...? on the "provider" side And maybe what their expectations on `force=True` are, to make sure they roughly align. The best definition for when to use `force=True` at this time seems to be "end-point" users (such as visualization or maybe writing to disk?). I still think performance can be just as valid of an issue there. For example it may be better to convert to a numpy array earlier in the computation. Or someone could be surprised that saving their gpu array to an hdf5 file is by far the slowest part of the computation. Maybe I have the feeling the definition we want is actually: There is definitely no way to do this computation faster or better than by converting it to a NumPy array. Since currently the main reason to reject it seems a bit to be: Wait, are you sure there is not a much better way than using NumPy arrays, be careful! And while that distinction is clear for PyTorch + visualization, I am not quite sure yet, that it is clear for various combinations of `force=True` and array-like users. Maybe CuPy does not want h5py to use `force=True`, because cupy has its own super fast "stream-to-file" functionality... But it wants to to do it for napari. - Sebastian

Avoiding .gpu() is more straightforwardly about avoiding implicit expensive computation.

...

while others do not choose to teach about it. There seems very little or even no "promise" attached to either `force=True` or `force=False`.

NumPy can set a precedent through policy. The *only* reason client libraries would implement `__array__` is to play well with NumPy, so if NumPy documents that `force=True` should *always* succeed, we can expect client libraries to follow suit. At least the PyTorch devs have indicated that they would be open to this.

...

E.g. Napari wants to use it, but do the array-providers want Napari to use it?

As Ralf pointed out, the PyTorch devs have already agreed to it.

From the napari perspective, we'd be ok with leaving the decision on warnings to client libraries. We may or may not suppress them depending on user requests. ;) But the point is to have a way of saying "give me a NumPy array DAMMIT" without having to know about all the possible array libraries. Which are numerous and getting numerouser.

Ralf, you said you don't want warnings — even for sparse arrays? That was an area of concern for you on the PyTorch discussion side.

...

And if the conversion still gives warnings for some array-objects, have we actually gained much?

Yes.

Hameer,

...

I would advocate for a `force=` kwarg but personally I don't think it's explicit enough, but probably as explicit as can be given NumPy's API.

Yeah, I agree that force is kind of vague, which is why I was looking for things like `allow_copy`. But it is hard to be general enough here: sparse requires an expensive instantiation, cupy requires copying from gpu to cpu, dask requires arbitrary computation, xarray requires information loss... I'm inclined to agree with Ralf that force= is the only generic-enough term, but I'm happy to entertain other options!

Juan. _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@python.org https://mail.python.org/mailman/listinfo/numpy-discussion