Support the NumPy pad fct. #1216
Comments
|
It need at least a rebase. Otherwise, there was suggested update, to call our other ops when possible. If you continue this, start by a rebase and make a new PR. Maybe there will On Fri, Jan 15, 2016 at 12:48 AM, Saatvik Shah notifications@github.com
|
|
Hello nouiz, I would like to work on this. #1957 has implemented part of this (I guess). What are the modifications that are supposed to be carried out further ? |
|
It need a rebase. The main problem is that this op don't implement a grad
and don't work on the GPU.
We will also need to modify the pad function to return a graph that use
alloc and set_subtensor for the case this work. This will make those case
work on the GPU and support grad.
With that, I think it would be good to merge. The padding mode constant can
be implemented as the above.
Also, it will need some update to use the newer Op interface. But when the
rebase is done, I can comment on the new PR about the change to do.
thanks
…On Tue, Jan 17, 2017 at 9:04 AM, Kv Manohar ***@***.***> wrote:
Hello nouiz, I would like to work on this. #1957
<5eb9b4f>
has implemented part of this (I guess). What are the modifications that are
supposed to be carried out further ?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1216 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/AALC-wyhffYtEs1WHeS4MbRjLp2fS1VLks5rTMp8gaJpZM4AbCDe>
.
|
NumPy 1.7 support a new pad instruction. We should support it. Some mode are easy to support with current ops, but some of then could need new opt or more complicated graph. In the cases were we don't need a new complicated graph, it is better to don't use a new op. But may to support all mode, making a new op "Pad" that wrap this fct is needed.
Ability to pad rank-n arrays
A pad module containing functions for padding n-dimensional arrays has been
added. The various private padding functions are exposed as options to a
public 'pad' function. Example::
Current modes are
constant,edge,linear_ramp,maximum,mean,median,minimum,reflect,symmetric,wrap, and<function>.The text was updated successfully, but these errors were encountered: