Two thumbs up! (packed into one float thumbs up) HP
On Fri, Feb 19, 2016 at 10:21 AM, Andrew Gartner <[email protected]> wrote: > Yep, I'm good with that. > > Thanks again for teasing this apart Larry/HP > > ~Andrew > > On Fri, Feb 19, 2016 at 10:09 AM, Larry Gritz <[email protected]> wrote: > >> OK, that sounds reasonable. So we have: >> >> * If you ask for a specific type, convert and return an array of that >> type. If you ask for HALF, the half bit pattern gets returned in a uint16 >> array, since there is no true half type. >> >> * If you ask for UNKNOWN (explicitly "give me raw data"), it returns an >> array of unsigned chars containing the raw data. >> >> Everybody can live with that? >> >> >> On Feb 19, 2016, at 8:08 AM, Haarm-Pieter Duiker <[email protected]> >> wrote: >> >> Of the options "pass half values disguised as an unsigned short array" >> feels the cleanest to me. You keep the right number of components in the >> array, if you have any checks for that, and the data to be convert to halfs >> is already grouped appropriately. >> >> Converting to halfs is also a one line call to numpy: >> np.frombuffer(np.getbuffer(np.uint16(uint16Value)), dtype=np.float16) >> Ex. https://github.com/hpd/CLF/blob/master/python/aces/clf/Common.py#L92 >> >> HP >> >> >> >> >> >> On Thu, Feb 18, 2016 at 4:06 PM, Larry Gritz <[email protected]> wrote: >> >>> In C++, asking for UNKNOWN just copies the native format data and leaves >>> it for you to sort out. But to C++, a buffer is a buffer, you're passing it >>> a void* in any case. >>> >>> In Python, it's dynamic typing, so read_image RETURNS an array, and it >>> has to be an array of some type. Which type? >>> >>> I think we all are coming to agree that if you ask for UNKNOWN, probably >>> the most analogous thing (to C++) is to return an unsigned char array, >>> filled with the raw data, and leave you to sort it out. That's as close to >>> "untyped raw buffer" as we can get. >>> >>> If you *ask* for HALF, it's nonsensical, because you can't make an >>> actual half array in Python. You could promote and convert it to float. Or >>> you could return raw values in unsigned char array (like if you'd passed >>> UNKNOWN). Or, yeah, another possibility is to pass half values disguised as >>> an unsigned short array? >>> >>> I'm not super fond of the last choice. >>> >>> Right now, we do something stupider than any of those -- which is to >>> pack raw half values into a buffer, but the buffer advertises itself as >>> being a float array. That clearly needs to change. It was never >>> intentional; I just never thought carefully about that case because I never >>> imagined anybody asking for a type that didn't exist in Python. >>> >>> So, current proposal on the table: >>> >>> * If you ask for a type that can be a valid Python array type, convert >>> and return an array of that type. >>> >>> * If you ask for UNKNOWN (explicitly "give me raw data") or HALF >>> (implicitly so, because it doesn't exist in Python), it returns an array of >>> unsigned chars containing the raw data. >>> >>> >>> >>> On Feb 18, 2016, at 3:36 PM, Haarm-Pieter Duiker < >>> [email protected]> wrote: >>> >>> Returning a series of unsigned 16 bit ints for a call with the type half >>> feels like a nice middle ground. The consumer will have to know that halfs >>> aren't natively supported in Python, and how to convert from unsigned short >>> to half, but that doesn't feel like a large burden. >>> >>> I can't speak to the expected behavior of the UNKNOWN in Python. I >>> haven't used that path in Python or C++. >>> >>> HP >>> >>> >>> >>> >>> >>> >>> >>> >>> On Thu, Feb 18, 2016 at 1:23 PM, Larry Gritz <[email protected]> wrote: >>> >>>> I don't have especially strong feelings about this one way or the other. >>>> >>>> Just returning a raw data byte array matches the C++ behavior more >>>> closely, no argument there. >>>> >>>> On the "con" side, perhaps I was thinking of compatibility? We're >>>> really talking about changing the meaning of oiio.UNKNOWN from "use >>>> spec.format" to "return raw data", which differ in the case of mixed >>>> channel types. >>>> >>>> Are there Python programs out there that pass UNKNOWN (or pass nothing, >>>> defaulting to UNKNOWN) and rely on getting the right kind of array back >>>> that matches spec.format? >>>> >>>> >>>> >>>> >>>> On Feb 18, 2016, at 12:58 PM, Andrew Gartner <[email protected]> >>>> wrote: >>>> >>>> "Second, I could collapse 2a and 2b, and just say that if you ask for >>>> UNKNOWN, you get an array of uint8 back with the native raw data" >>>> >>>> Just out of curiosity, what are the drawbacks to doing this? I admit I >>>> like having some way of getting at the raw data at any time (hence my >>>> original method of exposing the native calls). That allowed me to check my >>>> imagespec and regardless of whether I had a mixed format image or all half >>>> data I could get everything in one read call. Granted I'm used to keeping >>>> track of and manipulating the strides of those arrays in bytes just out of >>>> old habit (and C++ usage) so maybe I'm the minority opinion. >>>> >>>> Even so, your current thinking still works if that's where the >>>> consensus is I'm happy to use it as such. >>>> >>>> Thanks again >>>> >>>> ~Andrew >>>> >>>> On Thu, Feb 18, 2016 at 12:21 PM, Larry Gritz <[email protected]> >>>> wrote: >>>> >>>>> I think that the only format that we can encounter as pixel data, >>>>> which does not exist in Python arrays, is 'half'. >>>>> >>>>> So let me rephrase my current thinking: >>>>> >>>>> 1. If you ask for a specific type (except HALF), you'll get a Python >>>>> array of that type holding the converted values. >>>>> >>>>> 2. Otherwise (i.e., you ask for UNKNOWN or HALF), you will get the >>>>> native (raw) data. >>>>> (a) If all channels are the same data type and it's anything but half, >>>>> you'll get the data as a Python array of that type. >>>>> (b) Otherwise (half, or mixed channel types), you'll get the data as a >>>>> Python array of unsigned bytes. >>>>> >>>>> Note that (1) is the easy case to deal with: ask for the type you >>>>> want, let it do the conversion. If you go for option (2) by asking for >>>>> native data, you get a blob and it's up to you to figure out what to do >>>>> with it. >>>>> >>>>> There are two other choices we could make. I'm not inclined to at the >>>>> moment, but would be happy to do so if people think it's helpful. First, >>>>> if >>>>> you ask for HALF, I could have it return float. Second, I could collapse >>>>> 2a >>>>> and 2b, and just say that if you ask for UNKNOWN, you get an array of >>>>> uint8 >>>>> back with the native raw data, even if it happened to be all channels of >>>>> the same type, a type that you could have made into a Python array of the >>>>> right type. >>>>> >>>>> >>>>> On Feb 17, 2016, at 11:16 PM, Haarm-Pieter Duiker < >>>>> [email protected]> wrote: >>>>> >>>>> Picking this up a little later in the day. Sorry about that. Adding >>>>> quotes from earlier in the thread just so it's clear what I'm responding >>>>> to. >>>>> >>>>> The current status: >>>>> " >>>>> If you read_image(oiio.FLOAT) of a half image (on disk), you get >>>>> floats back? >>>>> " >>>>> Yes. >>>>> >>>>> " >>>>> But if you read_image(oiio.HALF) of a half image, you get what appears >>>>> to be an array of floats, but they are actually packed half values? >>>>> " >>>>> Yes. >>>>> >>>>> The proposal: >>>>> " >>>>> 1. If you ask for a (non-UNKNOWN) format that exists in Python, it >>>>> converts to and returns an array of that format. >>>>> " >>>>> This is the current behavior, no? >>>>> >>>>> " >>>>> 2. If you ask for UNKNOWN, or a format that doesn't exist, it returns >>>>> the raw data in an unsigned char array. >>>>> " >>>>> It feels like this is two proposals (Trying not to clash with your >>>>> earlier 2a and 2b): >>>>> 2c. If you ask for UNKNOWN, return raw data in an unsigned char array >>>>> 2d. If you ask for a format that doesn't exist, return raw data in an >>>>> unsigned char array >>>>> >>>>> 2c. feels right. It should work for the case of typical RGB or RGBA >>>>> images but also for multi-layer EXRs. The consumer can convert the >>>>> channels >>>>> to their intended types using methods from the ImageSpec. I'd suggest that >>>>> asking for UNKNOWN lead unequivocally to a raw unsigned char array. >>>>> Supporting the special cases described in the 2a and 2b listed earlier >>>>> would require additional logic on the consuming code side to account for >>>>> those cases. Feels like a recipe for lots of brittle special case logic. >>>>> >>>>> 2d. is less clear. How is the change in behavior from returning real >>>>> values for known types to returning raw char array data for unknown types >>>>> signaled to the consumer? Is this still something that programmers have to >>>>> just know a priori? How is this different from the current behavior? >>>>> >>>>> I suppose the list of types known to OIIO but not Python is finite and >>>>> likely to shrink over time. Having special cases like we have in that >>>>> example code, isn't such a big deal in the mean time, but then that's just >>>>> saying the the current behavior is fine. >>>>> >>>>> Hope that's helpful in some way. Aside from agreeing that adding an >>>>> UNKNOWN option is a good idea, we're still left without a good way to >>>>> consume half data without accounting for it explicitly. >>>>> >>>>> HP >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> On Wed, Feb 17, 2016 at 4:35 PM, Andrew Gartner < >>>>> [email protected]> wrote: >>>>> >>>>>> That would certainly take care of things for me. Hopefully not too >>>>>> much of an impact on others as well. >>>>>> >>>>>> ~Andrew >>>>>> >>>>>> >>>>>> >>>>>> On Wed, Feb 17, 2016 at 4:20 PM, Larry Gritz <[email protected]> >>>>>> wrote: >>>>>> >>>>>>> So I'm proposing: >>>>>>> >>>>>>> 1. If you ask for a (non-UNKNOWN) format that exists in Python, it >>>>>>> converts to and returns an array of that format. >>>>>>> >>>>>>> 2. If you ask for UNKNOWN, or a format that doesn't exist, it >>>>>>> returns the raw data in an unsigned char array. >>>>>>> >>>>>>> >>>>>>> There is a variation: >>>>>>> >>>>>>> 2a. If you ask for UNKNOWN, and all channels are the same format and >>>>>>> it's a type that exists in Python, return that type. >>>>>>> 2b. If you ask for UNKNOWN and it's a "mixed type" file, or a single >>>>>>> type but one that doesn't exist in Python, or the type you ask for >>>>>>> doesn't >>>>>>> exist in Python, return raw data packed into an unsigned char array. >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Feb 17, 2016, at 4:10 PM, Andrew Gartner <[email protected]> >>>>>>> wrote: >>>>>>> >>>>>>> Yea the C++ implementation works well with oiio.UNKNOWN, I kinda >>>>>>> miss that in the python side to be honest. Right now it looks like >>>>>>> things >>>>>>> revert back to spec.format if oiio.UNKNOWN is supplied to >>>>>>> read_scanlines, >>>>>>> that can be problematic if you have multiple formats in a single image >>>>>>> so >>>>>>> I've avoided it. >>>>>>> >>>>>>> @Larry, to you question about returning an unsigned char array, I >>>>>>> like the idea on principle in that it preserves the decoupling as you >>>>>>> said. >>>>>>> I'm wondering if there would be any weirdness if you had to grab >>>>>>> multiple >>>>>>> channels of an image that had different data types one of which isn't >>>>>>> representable in python? Would it default to just unsigned char yet >>>>>>> again >>>>>>> in that case? >>>>>>> >>>>>>> @Haarm: interesting, I didn't realize they were concatenated/packed >>>>>>> like that! I just saw the 'f' in the python array and assumed I was >>>>>>> seeing >>>>>>> promoted values :) I'm still scratching my head over the multiple format >>>>>>> reads though, same as for Larry's idea. >>>>>>> >>>>>>> Thanks for the replies, Cheers, >>>>>>> >>>>>>> ~Andrew >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Wed, Feb 17, 2016 at 3:49 PM, Haarm-Pieter Duiker < >>>>>>> [email protected]> wrote: >>>>>>> >>>>>>>> If you're up for using numpy, this will get you the half float >>>>>>>> values without too much extra work: >>>>>>>> oiioFloats = inputImage.read_image(oiio.HALF) >>>>>>>> oiioHalfs = np.frombuffer(np.getbuffer(np.float32(oiioFloats)), >>>>>>>> dtype=np.float16) >>>>>>>> >>>>>>>> One note, the current OIIO Python implementation doesn't promote >>>>>>>> the halfs to float on read. The 'float' values in the returned buffer >>>>>>>> are >>>>>>>> actually each two concatenated half values, and the float buffer will >>>>>>>> have >>>>>>>> half as many entries as you would expect. >>>>>>>> >>>>>>>> Example usage for reading here: >>>>>>>> >>>>>>>> https://github.com/hpd/CLF/blob/master/python/aces/filterImageWithCLF.py#L126 >>>>>>>> and the reverse for writing: >>>>>>>> >>>>>>>> https://github.com/hpd/CLF/blob/master/python/aces/filterImageWithCLF.py#L193 >>>>>>>> >>>>>>>> HP >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On Wed, Feb 17, 2016 at 3:25 PM, Larry Gritz <[email protected]> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> In C++, you can just call read_scanlines and pass format=UNKNOWN >>>>>>>>> to get back the raw data in its original format. >>>>>>>>> >>>>>>>>> The problem is that in Python, there is no 'half' so it's not >>>>>>>>> quite sure what to return. >>>>>>>>> >>>>>>>>> I kinda like the decoupling of the raw reads (read_native_*) which >>>>>>>>> are the part overloaded by the individual format readers, from the >>>>>>>>> app-callable read_*. So perhaps rather than exposing read_native_*, we >>>>>>>>> should just modify the Python bindings for read_* to notice that if >>>>>>>>> the >>>>>>>>> native raw data is not a type representable in Python, to return it >>>>>>>>> as an >>>>>>>>> unsigned character array? >>>>>>>>> >>>>>>>>> >>>>>>>>> > On Feb 17, 2016, at 2:55 PM, Andrew Gartner < >>>>>>>>> [email protected]> wrote: >>>>>>>>> > >>>>>>>>> > Hey all, >>>>>>>>> > >>>>>>>>> > Apologies if this has come up before, but I'm curious if anyone >>>>>>>>> had considered exposing ImageInput.read_native_scanlines() on the >>>>>>>>> python >>>>>>>>> side before. The reason I ask is mainly because the half datatype >>>>>>>>> doesn't >>>>>>>>> exist in the native python array class which OIIO uses for python >>>>>>>>> reads. >>>>>>>>> Currently the python array will punt and for anything to float (which >>>>>>>>> I'd >>>>>>>>> rather avoid). >>>>>>>>> > >>>>>>>>> > I had put together an implementation in OIIO 1.5 that simply >>>>>>>>> took the pixel size as a parameter and exposed read_native_scanlines >>>>>>>>> that >>>>>>>>> way and that allowed me to get the right data properly into either >>>>>>>>> numpy or >>>>>>>>> a raw char python array. However, I'd rather not be forked off like >>>>>>>>> that as >>>>>>>>> it's a headache trying to remain current with the mainline, plus >>>>>>>>> others may >>>>>>>>> find it useful. >>>>>>>>> > >>>>>>>>> > Does anyone think exposing the function in general makes sense? >>>>>>>>> I'm happy to send the implementation if anyone cares to see it as >>>>>>>>> well. >>>>>>>>> > >>>>>>>>> > Cheers, >>>>>>>>> > >>>>>>>>> > ~Andrew >>>>>>>>> > >>>>>>>>> >>>>>>>>> -- >>>>>>>>> Larry Gritz >>>>>>>>> [email protected] >>>>>>>>> >>>>>>>>> >>>>>>>>> _______________________________________________ >>>>>>>>> Oiio-dev mailing list >>>>>>>>> [email protected] >>>>>>>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> _______________________________________________ >>>>>>>> Oiio-dev mailing list >>>>>>>> [email protected] >>>>>>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>>>>>> >>>>>>>> >>>>>>> _______________________________________________ >>>>>>> Oiio-dev mailing list >>>>>>> [email protected] >>>>>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>>>>> >>>>>>> >>>>>>> -- >>>>>>> Larry Gritz >>>>>>> [email protected] >>>>>>> >>>>>>> >>>>>>> >>>>>>> _______________________________________________ >>>>>>> Oiio-dev mailing list >>>>>>> [email protected] >>>>>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>>>>> >>>>>>> >>>>>> >>>>>> _______________________________________________ >>>>>> Oiio-dev mailing list >>>>>> [email protected] >>>>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>>>> >>>>>> >>>>> _______________________________________________ >>>>> Oiio-dev mailing list >>>>> [email protected] >>>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>>> >>>>> >>>>> -- >>>>> Larry Gritz >>>>> [email protected] >>>>> >>>>> >>>>> >>>>> _______________________________________________ >>>>> Oiio-dev mailing list >>>>> [email protected] >>>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>>> >>>>> >>>> _______________________________________________ >>>> Oiio-dev mailing list >>>> [email protected] >>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>> >>>> >>>> -- >>>> Larry Gritz >>>> [email protected] >>>> >>>> >>>> >>>> _______________________________________________ >>>> Oiio-dev mailing list >>>> [email protected] >>>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>>> >>>> >>> _______________________________________________ >>> Oiio-dev mailing list >>> [email protected] >>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>> >>> >>> -- >>> Larry Gritz >>> [email protected] >>> >>> >>> >>> _______________________________________________ >>> Oiio-dev mailing list >>> [email protected] >>> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >>> >>> >> _______________________________________________ >> Oiio-dev mailing list >> [email protected] >> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >> >> >> -- >> Larry Gritz >> [email protected] >> >> >> >> _______________________________________________ >> Oiio-dev mailing list >> [email protected] >> http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org >> >> > > _______________________________________________ > Oiio-dev mailing list > [email protected] > http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org > >
_______________________________________________ Oiio-dev mailing list [email protected] http://lists.openimageio.org/listinfo.cgi/oiio-dev-openimageio.org
