Hi folks,
Someone on the wxPython list posted a nifty recarray example that I don't quite understand. The idea is to have an array for an RGBA image:
rgbarec = numpy.dtype({'r':(numpy.uint8,0), 'g':(numpy.uint8,1), 'b':(numpy.uint8,2), 'a':(numpy.uint8,3)})
A = numpy.zeros(shape, dtype=(numpy.uint32, rgbarec) )
what I don't understand is having BOTH numpy.uint32 and rgbrec as the dtype. How does that work?
Actually, with a bit of testing, it's pretty cool -- if you index like:
A[i,j] ==> uint32
A['r'][i,j] ==> uint8
pretty cool really.
it seems that numpy is treating it as both a recarray and a regular uint32 array, which makes some sense, but I'm still confused by the semantics.
also:
A
array([[4278190080, ... [4278190080, 4278190080, 4278190080, 4278190080, 4278190080]], dtype=uint32)
but:
A.dtype
dtype(('>u4', [('r', '|u1'), ('g', '|u1'), ('b', '|u1'), ('a', '|u1')]))
so what is the dtype?
Also, I see advantages and disadvantages to either way. If you do:
A = numpy.zeros(shape, dtype=(numpy.uint32, rgbarec) )
then you can do:
A[i,j]['r'] to get the red value of a pixel A[i,j] = (red, green, blue, alpha) to set a pixel A[:,:] = (red, green, blue, alpha) to make the whole image one color
With the "dual dtype" approach: A = numpy.zeros(shape, dtype=(numpy.uint32, rgbarec) )
You need to set the pixels separately: A['r'][i,j] = red B['r'][i,j] = green ... or construct uint32 values:
C = numpy.uint32(red) C += numpy.uint32(green) << 8 C += numpy.uint32(blue) << 16
Which is really ugly! (and I may not even have it right!). Is there a better way?
One more idea -- is there a way to have two arrays, each with a different dtype, that point to the same data? maybe:
RGBImage = numpy.zeros(shape, dtype=rgbarec )
IntImage = RGBImage.view() IntImage.dtype = numpy.uint32
That way you could work with it either way, whichever was easier in the context.
So -- what are folks' thoughts about how best to deal with images as numpy arrays? (I'll put it in a Wiki page if I get some good comments)
-Chris
Christopher Barker wrote:
Hi folks,
Someone on the wxPython list posted a nifty recarray example that I don't quite understand. The idea is to have an array for an RGBA image:
rgbarec = numpy.dtype({'r':(numpy.uint8,0), 'g':(numpy.uint8,1), 'b':(numpy.uint8,2), 'a':(numpy.uint8,3)})
A = numpy.zeros(shape, dtype=(numpy.uint32, rgbarec) )
what I don't understand is having BOTH numpy.uint32 and rgbrec as the dtype. How does that work?
Basically, the rgbarec defines the fields, but the "base-type" for the numpy array is numpy.uint32 (rather than VOID which is the default data-type for arrays with fields defined). This is why it prints the way it does.
I'm not sure what the real value is doing it that way as opposed to just having two views on the data: one as a uint32 and another as a "normal" recarray with a VOID data-type underlying it) like you suggest at the end.
I think it's really just the "architecture" showing through to the user layer.
-Travis
Hi Chris
2008/4/30 Christopher Barker Chris.Barker@noaa.gov:
Someone on the wxPython list posted a nifty recarray example that I don't quite understand. The idea is to have an array for an RGBA image:
rgbarec = numpy.dtype({'r':(numpy.uint8,0), 'g':(numpy.uint8,1), 'b':(numpy.uint8,2), 'a':(numpy.uint8,3)})
I find manipulating data with records arrays highly intuitive; I had the same approach in mind when I wrote
http://www.scipy.org/RecordArrays
One more idea -- is there a way to have two arrays, each with a different dtype, that point to the same data? maybe:
RGBImage = numpy.zeros(shape, dtype=rgbarec )
IntImage = RGBImage.view() IntImage.dtype = numpy.uint32
That way you could work with it either way, whichever was easier in the context.
That's the way, or just rgba_image.view(numpy.int32).
Cheers Stéfan
Stéfan van der Walt wrote:
That's the way, or just rgba_image.view(numpy.int32).
ah -- interestingly, I tried:
rgba_image.view(dtype=numpy.int32)
and got:
Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: view() takes no keyword arguments
Since it is optional, shouldn't it be keyword argument?
-Chris
2008/4/30 Christopher Barker Chris.Barker@noaa.gov:
Stéfan van der Walt wrote:
That's the way, or just rgba_image.view(numpy.int32).
ah -- interestingly, I tried:
rgba_image.view(dtype=numpy.int32)
and got:
Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: view() takes no keyword arguments
Since it is optional, shouldn't it be keyword argument?
Thanks, fixed in r5115.
Cheers Stéfan
Stéfan van der Walt wrote:
2008/4/30 Christopher Barker Chris.Barker@noaa.gov:
Stéfan van der Walt wrote:
That's the way, or just rgba_image.view(numpy.int32).
ah -- interestingly, I tried:
rgba_image.view(dtype=numpy.int32)
and got:
Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: view() takes no keyword arguments
Since it is optional, shouldn't it be keyword argument?
Thanks, fixed in r5115.
This was too hasty. I had considered this before.
The problem with this is that the object can be either a type object or a data-type object. You can use view to both re-cast a numpy array as another subtype or as another data-type.
So, please revert the change until a better solution is posted.
-Travis
Travis E. Oliphant wrote:
Stéfan van der Walt wrote:
2008/4/30 Christopher Barker Chris.Barker@noaa.gov:
Since it is optional, shouldn't it be keyword argument?
Thanks, fixed in r5115.
This was too hasty. I had considered this before.
The problem with this is that the object can be either a type object or a data-type object. You can use view to both re-cast a numpy array as another subtype or as another data-type.
Is the issue here that this is a slightly different meaning than the "dtype" argument everywhere else? Frankly, that doesn't bother me -- it is a superset of the functionality, is it not?
Though I guess I don't really understand quite what the difference is between a subtype and a data-type. Or a type object vs. a datatype object.
-Chris
2008/5/1 Travis E. Oliphant oliphant@enthought.com:
Stéfan van der Walt wrote:
2008/4/30 Christopher Barker Chris.Barker@noaa.gov:
Stéfan van der Walt wrote:
That's the way, or just rgba_image.view(numpy.int32).
ah -- interestingly, I tried:
rgba_image.view(dtype=numpy.int32)
and got:
Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: view() takes no keyword arguments
Since it is optional, shouldn't it be keyword argument?
Thanks, fixed in r5115.
This was too hasty. I had considered this before.
The problem with this is that the object can be either a type object or a data-type object. You can use view to both re-cast a numpy array as another subtype or as another data-type.
So, please revert the change until a better solution is posted.
If we're going to support keyword arguments, shouldn't those two options be *different* keywords (dtype and ndarray_subclass, say)? Then a single non-keyword argument tells numpy to guess which one you wanted...
Anne
2008/5/1 Travis E. Oliphant oliphant@enthought.com:
Stéfan van der Walt wrote:
2008/4/30 Christopher Barker Chris.Barker@noaa.gov:
Stéfan van der Walt wrote:
That's the way, or just rgba_image.view(numpy.int32).
ah -- interestingly, I tried:
rgba_image.view(dtype=numpy.int32)
and got:
Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: view() takes no keyword arguments
Since it is optional, shouldn't it be keyword argument?
Thanks, fixed in r5115.
This was too hasty. I had considered this before.
The problem with this is that the object can be either a type object or a data-type object. You can use view to both re-cast a numpy array as another subtype or as another data-type.
So, please revert the change until a better solution is posted.
OK, I see your point. I'm working on a patch that does the following:
def view(type_or_dtype=None, dtype=None, type=None): if type_or_dtype: if dtype: raise ValueError("Cannot specify dtype twice") if type: raise ValueError("Cannot specify type twice")
if isinstance(type_or_dtype,py_type): type = type_or_dtype
if isinstance(type_or_dtype,numpy_dtype): dtype = type_or_dtype
return x.view(type=type).view(dtype=dtype)
Would that be a satisfying solution? I'll be back around 21:00 SAST to attend to the matter.
Regards Stéfan
Stéfan van der Walt wrote:
2008/5/1 Travis E. Oliphant oliphant@enthought.com:
Stéfan van der Walt wrote:
2008/4/30 Christopher Barker Chris.Barker@noaa.gov:
Stéfan van der Walt wrote:
That's the way, or just rgba_image.view(numpy.int32).
ah -- interestingly, I tried:
rgba_image.view(dtype=numpy.int32)
and got:
Traceback (most recent call last): File "<stdin>", line 1, in <module> TypeError: view() takes no keyword arguments
Since it is optional, shouldn't it be keyword argument?
Thanks, fixed in r5115.
This was too hasty. I had considered this before.
The problem with this is that the object can be either a type object or a data-type object. You can use view to both re-cast a numpy array as another subtype or as another data-type.
So, please revert the change until a better solution is posted.
OK, I see your point. I'm working on a patch that does the following:
def view(type_or_dtype=None, dtype=None, type=None): if type_or_dtype: if dtype: raise ValueError("Cannot specify dtype twice") if type: raise ValueError("Cannot specify type twice")
if isinstance(type_or_dtype,py_type): type = type_or_dtype if isinstance(type_or_dtype,numpy_dtype): dtype = type_or_dtype return x.view(type=type).view(dtype=dtype)
Would that be a satisfying solution? I'll be back around 21:00 SAST to attend to the matter.
Yes, I think that would work. You need to do some checking for type=None and dtype=None as well, though.
That way, the first argument would continue to work as now but be labeled correctly, but it would also support dtype= and type= keywords.
-Travis
Travis E. Oliphant wrote:
def view(type_or_dtype=None, dtype=None, type=None):
Yes, I think that would work.
Is there a way to deprecate this for future API-incompatible versions? It's better than non keywords, but a bit ugly.
Maybe we should have a Wiki page for "stuff we'd like to change, but won't until major API breakage is otherwise occurring"
-Chris
On Thu, 01 May 2008, Christopher Barker apparently wrote:
Maybe we should have a Wiki page for "stuff we'd like to change, but won't until major API breakage is otherwise occurring"
Perhaps URL:http://www.scipy.org/ProposedEnhancements would suffice?
Cheers, Alan
2008/5/1 Travis E. Oliphant oliphant@enthought.com:
OK, I see your point. I'm working on a patch that does the following:
def view(type_or_dtype=None, dtype=None, type=None): if type_or_dtype: if dtype: raise ValueError("Cannot specify dtype twice") if type: raise ValueError("Cannot specify type twice")
if isinstance(type_or_dtype,py_type): type = type_or_dtype if isinstance(type_or_dtype,numpy_dtype): dtype = type_or_dtype return x.view(type=type).view(dtype=dtype)
Would that be a satisfying solution? I'll be back around 21:00 SAST to attend to the matter.
Yes, I think that would work. You need to do some checking for type=None and dtype=None as well, though.
That way, the first argument would continue to work as now but be labeled correctly, but it would also support dtype= and type= keywords.
Please review http://projects.scipy.org/scipy/numpy/changeset/5117.
Thanks Stéfan
Please review http://projects.scipy.org/scipy/numpy/changeset/5117.
Stefan,
I don't think we really need the dtype_or_type keyword. It seems that we could just check the first argument (dtype) to see if it is a subtype of the ndarray and assume that it is type= in that case.
-Travis
Stéfan van der Walt wrote:
2008/5/1 Travis E. Oliphant oliphant@enthought.com:
OK, I see your point. I'm working on a patch that does the following:
def view(type_or_dtype=None, dtype=None, type=None): if type_or_dtype: if dtype: raise ValueError("Cannot specify dtype twice") if type: raise ValueError("Cannot specify type twice")
if isinstance(type_or_dtype,py_type): type = type_or_dtype if isinstance(type_or_dtype,numpy_dtype): dtype = type_or_dtype return x.view(type=type).view(dtype=dtype)
Would that be a satisfying solution? I'll be back around 21:00 SAST to attend to the matter.
Yes, I think that would work. You need to do some checking for type=None and dtype=None as well, though.
That way, the first argument would continue to work as now but be labeled correctly, but it would also support dtype= and type= keywords.
Please review http://projects.scipy.org/scipy/numpy/changeset/5117.
Check out
http://projects.scipy.org/scipy/numpy/changeset/5119
-Travis
2008/5/1 Travis E. Oliphant oliphant@enthought.com:
Stéfan van der Walt wrote:
2008/5/1 Travis E. Oliphant oliphant@enthought.com:
OK, I see your point. I'm working on a patch that does the following:
def view(type_or_dtype=None, dtype=None, type=None): if type_or_dtype: if dtype: raise ValueError("Cannot specify dtype twice") if type: raise ValueError("Cannot specify type twice")
if isinstance(type_or_dtype,py_type): type = type_or_dtype if isinstance(type_or_dtype,numpy_dtype): dtype = type_or_dtype return x.view(type=type).view(dtype=dtype)
Would that be a satisfying solution? I'll be back around 21:00 SAST to attend to the matter.
Yes, I think that would work. You need to do some checking for type=None and dtype=None as well, though.
That way, the first argument would continue to work as now but be labeled correctly, but it would also support dtype= and type= keywords.
Please review http://projects.scipy.org/scipy/numpy/changeset/5117.
Check out
I think that's fine. It doesn't support weird combinations like
x.view(np.matrix,dtype=np.int32)
but people probably shouldn't do that anyway.
Cheers Stéfan