Theano 0.6

classic Classic list List threaded Threaded
17 messages Options
Reply | Threaded
Open this post in threaded view
|

Theano 0.6

Frédéric Bastien
Hi,

We released Theano 0.6 yesterday. I saw in that enpkg know about those
version of Theano:

    0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
    subscription), 0.5.0rc1 (no subscription)


It don't even know about Theano 0.5 that was release in February 2012.

It would be great if you add version 0.6 as available. To my
knowledge, Theano work well with EPD/Canopy on windows, mac and linux.

To my understanding Theano is included in the community section, so
without support from you. So that should not add extra work from you.
If I'm mistaken, please tell me. I would like to make Theano easily
available and having only old version shown as being available isn't
good.

If you need modification in Theano for any reason, do not hesitate to tell me.

thanks

Frédéric Bastien
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

David Cournapeau
Hi Frédéric,

I agree having ancient theano is not good.

I was considering moving it up into a supported package, but I feared that it would not be so useful without GPU support. I (unfortunately) never had time to play with theano, and don't know how useful the package is without CUDA (and I am a bit worried about support pressure when people will ask 'GPU is not used, fix it').

thanks,
David


On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]> wrote:
Hi,

We released Theano 0.6 yesterday. I saw in that enpkg know about those
version of Theano:

    0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
    subscription), 0.5.0rc1 (no subscription)


It don't even know about Theano 0.5 that was release in February 2012.

It would be great if you add version 0.6 as available. To my
knowledge, Theano work well with EPD/Canopy on windows, mac and linux.

To my understanding Theano is included in the community section, so
without support from you. So that should not add extra work from you.
If I'm mistaken, please tell me. I would like to make Theano easily
available and having only old version shown as being available isn't
good.

If you need modification in Theano for any reason, do not hesitate to tell me.

thanks

Frédéric Bastien
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev


_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
Hi,

Theano is very useful without GPU. We have some paper that show that
the CPU only back-end was faster then numpy in many cases and faster
then numexpr too.

Here is the most recent paper with benchmark but compare only to other
highly optimized packages on machine learning algo:

http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark

Here is the first paper with comparison again numpy/scipy and numexpr:

http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf look
at page 3 and 4.


This is just to show that Theano without the GPU is very useful. In
fact we use it frequently here as we don't have enough GPUs and have
access to a few CPU clusters.


I saw a few other people from Enthought play/contribute with/to Theano
like Robert Kern. If you want you can talk about this with them.

I don't know the detail of your supported status. Could you support it
on the CPU only? Or just refer to Theano installation for the GPU? On
linux and Mac, it isn't very complicated. But I agree that on Windows,
it is more complicated as the user need to install the microsoft
compiler in addition to nvcc. Officially, we don't support GPU on
Windows, but we help answer questions about this on the mailing when
people have problems, there is some instruction provided by user and
updated by others users. I have heard many people on the mailing tell
that they use that. In fact, when we make code change that MSVC don't
like(it is a more picky compiler then g++), users tell us about this
and frequently provide us with the fixes.

Frédéric


On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau <[hidden email]> wrote:

> Hi Frédéric,
>
> I agree having ancient theano is not good.
>
> I was considering moving it up into a supported package, but I feared that
> it would not be so useful without GPU support. I (unfortunately) never had
> time to play with theano, and don't know how useful the package is without
> CUDA (and I am a bit worried about support pressure when people will ask
> 'GPU is not used, fix it').
>
> thanks,
> David
>
>
> On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> Hi,
>>
>> We released Theano 0.6 yesterday. I saw in that enpkg know about those
>> version of Theano:
>>
>>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
>>     subscription), 0.5.0rc1 (no subscription)
>>
>>
>> It don't even know about Theano 0.5 that was release in February 2012.
>>
>> It would be great if you add version 0.6 as available. To my
>> knowledge, Theano work well with EPD/Canopy on windows, mac and linux.
>>
>> To my understanding Theano is included in the community section, so
>> without support from you. So that should not add extra work from you.
>> If I'm mistaken, please tell me. I would like to make Theano easily
>> available and having only old version shown as being available isn't
>> good.
>>
>> If you need modification in Theano for any reason, do not hesitate to tell
>> me.
>>
>> thanks
>>
>> Frédéric Bastien
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

David Cournapeau
Hi Frédéric,


On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]> wrote:
Hi,

Theano is very useful without GPU. We have some paper that show that
the CPU only back-end was faster then numpy in many cases and faster
then numexpr too.

Here is the most recent paper with benchmark but compare only to other
highly optimized packages on machine learning algo:

http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark

Here is the first paper with comparison again numpy/scipy and numexpr:

http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf look
at page 3 and 4.


This is just to show that Theano without the GPU is very useful. In
fact we use it frequently here as we don't have enough GPUs and have
access to a few CPU clusters.


I saw a few other people from Enthought play/contribute with/to Theano
like Robert Kern. If you want you can talk about this with them.

I don't know the detail of your supported status. Could you support it
on the CPU only? Or just refer to Theano installation for the GPU? On
linux and Mac, it isn't very complicated. But I agree that on Windows,
it is more complicated as the user need to install the microsoft
compiler in addition to nvcc. Officially, we don't support GPU on
Windows, but we help answer questions about this on the mailing when
people have problems, there is some instruction provided by user and
updated by others users. I have heard many people on the mailing tell
that they use that. In fact, when we make code change that MSVC don't
like(it is a more picky compiler then g++), users tell us about this
and frequently provide us with the fixes.

Thanks for the detailed explanation.

You convinced me it would be useful to add support for theano. It should be available within a few days, unless some unforeseen complications arise,

thanks,
David

Frédéric


On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau <[hidden email]> wrote:
> Hi Frédéric,
>
> I agree having ancient theano is not good.
>
> I was considering moving it up into a supported package, but I feared that
> it would not be so useful without GPU support. I (unfortunately) never had
> time to play with theano, and don't know how useful the package is without
> CUDA (and I am a bit worried about support pressure when people will ask
> 'GPU is not used, fix it').
>
> thanks,
> David
>
>
> On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> Hi,
>>
>> We released Theano 0.6 yesterday. I saw in that enpkg know about those
>> version of Theano:
>>
>>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
>>     subscription), 0.5.0rc1 (no subscription)
>>
>>
>> It don't even know about Theano 0.5 that was release in February 2012.
>>
>> It would be great if you add version 0.6 as available. To my
>> knowledge, Theano work well with EPD/Canopy on windows, mac and linux.
>>
>> To my understanding Theano is included in the community section, so
>> without support from you. So that should not add extra work from you.
>> If I'm mistaken, please tell me. I would like to make Theano easily
>> available and having only old version shown as being available isn't
>> good.
>>
>> If you need modification in Theano for any reason, do not hesitate to tell
>> me.
>>
>> thanks
>>
>> Frédéric Bastien
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev


_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
Thanks for the good news!

As always if you have any questions/comments don't hesitate to contact me.

Frédéric

On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau <[hidden email]> wrote:

> Hi Frédéric,
>
>
> On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> Hi,
>>
>> Theano is very useful without GPU. We have some paper that show that
>> the CPU only back-end was faster then numpy in many cases and faster
>> then numexpr too.
>>
>> Here is the most recent paper with benchmark but compare only to other
>> highly optimized packages on machine learning algo:
>>
>> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>>
>> Here is the first paper with comparison again numpy/scipy and numexpr:
>>
>> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf look
>> at page 3 and 4.
>>
>>
>> This is just to show that Theano without the GPU is very useful. In
>> fact we use it frequently here as we don't have enough GPUs and have
>> access to a few CPU clusters.
>>
>>
>> I saw a few other people from Enthought play/contribute with/to Theano
>> like Robert Kern. If you want you can talk about this with them.
>>
>> I don't know the detail of your supported status. Could you support it
>> on the CPU only? Or just refer to Theano installation for the GPU? On
>> linux and Mac, it isn't very complicated. But I agree that on Windows,
>> it is more complicated as the user need to install the microsoft
>> compiler in addition to nvcc. Officially, we don't support GPU on
>> Windows, but we help answer questions about this on the mailing when
>> people have problems, there is some instruction provided by user and
>> updated by others users. I have heard many people on the mailing tell
>> that they use that. In fact, when we make code change that MSVC don't
>> like(it is a more picky compiler then g++), users tell us about this
>> and frequently provide us with the fixes.
>
>
> Thanks for the detailed explanation.
>
> You convinced me it would be useful to add support for theano. It should be
> available within a few days, unless some unforeseen complications arise,
>
> thanks,
> David
>>
>>
>> Frédéric
>>
>>
>> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau <[hidden email]>
>> wrote:
>> > Hi Frédéric,
>> >
>> > I agree having ancient theano is not good.
>> >
>> > I was considering moving it up into a supported package, but I feared
>> > that
>> > it would not be so useful without GPU support. I (unfortunately) never
>> > had
>> > time to play with theano, and don't know how useful the package is
>> > without
>> > CUDA (and I am a bit worried about support pressure when people will ask
>> > 'GPU is not used, fix it').
>> >
>> > thanks,
>> > David
>> >
>> >
>> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]>
>> > wrote:
>> >>
>> >> Hi,
>> >>
>> >> We released Theano 0.6 yesterday. I saw in that enpkg know about those
>> >> version of Theano:
>> >>
>> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
>> >>     subscription), 0.5.0rc1 (no subscription)
>> >>
>> >>
>> >> It don't even know about Theano 0.5 that was release in February 2012.
>> >>
>> >> It would be great if you add version 0.6 as available. To my
>> >> knowledge, Theano work well with EPD/Canopy on windows, mac and linux.
>> >>
>> >> To my understanding Theano is included in the community section, so
>> >> without support from you. So that should not add extra work from you.
>> >> If I'm mistaken, please tell me. I would like to make Theano easily
>> >> available and having only old version shown as being available isn't
>> >> good.
>> >>
>> >> If you need modification in Theano for any reason, do not hesitate to
>> >> tell
>> >> me.
>> >>
>> >> thanks
>> >>
>> >> Frédéric Bastien
>> >> _______________________________________________
>> >> Enthought-Dev mailing list
>> >> [hidden email]
>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> >
>> >
>> > _______________________________________________
>> > Enthought-Dev mailing list
>> > [hidden email]
>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

David Cournapeau



On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]> wrote:
Thanks for the good news!

As always if you have any questions/comments don't hesitate to contact me.

So I have a preliminary egg for Theano, but the test suite of ~ 20 minutes is prohibitive. Is there an option for testing only a subset of Theano ?

For packaging, I care mostly about ensuring we have not introduced bugs when packaging things incorrectly, and I like running those everytime we update both forward and reverse dependencies of a given package.

thanks,
David

Frédéric

On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau <[hidden email]> wrote:
> Hi Frédéric,
>
>
> On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> Hi,
>>
>> Theano is very useful without GPU. We have some paper that show that
>> the CPU only back-end was faster then numpy in many cases and faster
>> then numexpr too.
>>
>> Here is the most recent paper with benchmark but compare only to other
>> highly optimized packages on machine learning algo:
>>
>> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>>
>> Here is the first paper with comparison again numpy/scipy and numexpr:
>>
>> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf look
>> at page 3 and 4.
>>
>>
>> This is just to show that Theano without the GPU is very useful. In
>> fact we use it frequently here as we don't have enough GPUs and have
>> access to a few CPU clusters.
>>
>>
>> I saw a few other people from Enthought play/contribute with/to Theano
>> like Robert Kern. If you want you can talk about this with them.
>>
>> I don't know the detail of your supported status. Could you support it
>> on the CPU only? Or just refer to Theano installation for the GPU? On
>> linux and Mac, it isn't very complicated. But I agree that on Windows,
>> it is more complicated as the user need to install the microsoft
>> compiler in addition to nvcc. Officially, we don't support GPU on
>> Windows, but we help answer questions about this on the mailing when
>> people have problems, there is some instruction provided by user and
>> updated by others users. I have heard many people on the mailing tell
>> that they use that. In fact, when we make code change that MSVC don't
>> like(it is a more picky compiler then g++), users tell us about this
>> and frequently provide us with the fixes.
>
>
> Thanks for the detailed explanation.
>
> You convinced me it would be useful to add support for theano. It should be
> available within a few days, unless some unforeseen complications arise,
>
> thanks,
> David
>>
>>
>> Frédéric
>>
>>
>> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau <[hidden email]>
>> wrote:
>> > Hi Frédéric,
>> >
>> > I agree having ancient theano is not good.
>> >
>> > I was considering moving it up into a supported package, but I feared
>> > that
>> > it would not be so useful without GPU support. I (unfortunately) never
>> > had
>> > time to play with theano, and don't know how useful the package is
>> > without
>> > CUDA (and I am a bit worried about support pressure when people will ask
>> > 'GPU is not used, fix it').
>> >
>> > thanks,
>> > David
>> >
>> >
>> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]>
>> > wrote:
>> >>
>> >> Hi,
>> >>
>> >> We released Theano 0.6 yesterday. I saw in that enpkg know about those
>> >> version of Theano:
>> >>
>> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
>> >>     subscription), 0.5.0rc1 (no subscription)
>> >>
>> >>
>> >> It don't even know about Theano 0.5 that was release in February 2012.
>> >>
>> >> It would be great if you add version 0.6 as available. To my
>> >> knowledge, Theano work well with EPD/Canopy on windows, mac and linux.
>> >>
>> >> To my understanding Theano is included in the community section, so
>> >> without support from you. So that should not add extra work from you.
>> >> If I'm mistaken, please tell me. I would like to make Theano easily
>> >> available and having only old version shown as being available isn't
>> >> good.
>> >>
>> >> If you need modification in Theano for any reason, do not hesitate to
>> >> tell
>> >> me.
>> >>
>> >> thanks
>> >>
>> >> Frédéric Bastien
>> >> _______________________________________________
>> >> Enthought-Dev mailing list
>> >> [hidden email]
>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> >
>> >
>> > _______________________________________________
>> > Enthought-Dev mailing list
>> > [hidden email]
>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev


_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
Currently there isn't an easy way to test just a subset of Theano. The
only way is to specify directory/files/test name as a parameter to
nosetests.

This have been asked a few times to have a partial but faster test
suite. Personally, I would prefer keeping the default like now and
having the user specify to run just a part of the test suite. Will
this fit your system?

I don't know the different way to mark tests as slow. Do you have a
recommended way to do this? I'll look at the way NumPy do this.

My daughter is sick and I'm in vacation for the next 2 weeks, so I
don't know when I can have this ready.

Fred


On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]> wrote:

>
>
>
> On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> Thanks for the good news!
>>
>> As always if you have any questions/comments don't hesitate to contact me.
>
>
> So I have a preliminary egg for Theano, but the test suite of ~ 20 minutes
> is prohibitive. Is there an option for testing only a subset of Theano ?
>
> For packaging, I care mostly about ensuring we have not introduced bugs when
> packaging things incorrectly, and I like running those everytime we update
> both forward and reverse dependencies of a given package.
>
> thanks,
> David
>>
>>
>> Frédéric
>>
>> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau <[hidden email]>
>> wrote:
>> > Hi Frédéric,
>> >
>> >
>> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>> > wrote:
>> >>
>> >> Hi,
>> >>
>> >> Theano is very useful without GPU. We have some paper that show that
>> >> the CPU only back-end was faster then numpy in many cases and faster
>> >> then numexpr too.
>> >>
>> >> Here is the most recent paper with benchmark but compare only to other
>> >> highly optimized packages on machine learning algo:
>> >>
>> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>> >>
>> >> Here is the first paper with comparison again numpy/scipy and numexpr:
>> >>
>> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf look
>> >> at page 3 and 4.
>> >>
>> >>
>> >> This is just to show that Theano without the GPU is very useful. In
>> >> fact we use it frequently here as we don't have enough GPUs and have
>> >> access to a few CPU clusters.
>> >>
>> >>
>> >> I saw a few other people from Enthought play/contribute with/to Theano
>> >> like Robert Kern. If you want you can talk about this with them.
>> >>
>> >> I don't know the detail of your supported status. Could you support it
>> >> on the CPU only? Or just refer to Theano installation for the GPU? On
>> >> linux and Mac, it isn't very complicated. But I agree that on Windows,
>> >> it is more complicated as the user need to install the microsoft
>> >> compiler in addition to nvcc. Officially, we don't support GPU on
>> >> Windows, but we help answer questions about this on the mailing when
>> >> people have problems, there is some instruction provided by user and
>> >> updated by others users. I have heard many people on the mailing tell
>> >> that they use that. In fact, when we make code change that MSVC don't
>> >> like(it is a more picky compiler then g++), users tell us about this
>> >> and frequently provide us with the fixes.
>> >
>> >
>> > Thanks for the detailed explanation.
>> >
>> > You convinced me it would be useful to add support for theano. It should
>> > be
>> > available within a few days, unless some unforeseen complications arise,
>> >
>> > thanks,
>> > David
>> >>
>> >>
>> >> Frédéric
>> >>
>> >>
>> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau <[hidden email]>
>> >> wrote:
>> >> > Hi Frédéric,
>> >> >
>> >> > I agree having ancient theano is not good.
>> >> >
>> >> > I was considering moving it up into a supported package, but I feared
>> >> > that
>> >> > it would not be so useful without GPU support. I (unfortunately)
>> >> > never
>> >> > had
>> >> > time to play with theano, and don't know how useful the package is
>> >> > without
>> >> > CUDA (and I am a bit worried about support pressure when people will
>> >> > ask
>> >> > 'GPU is not used, fix it').
>> >> >
>> >> > thanks,
>> >> > David
>> >> >
>> >> >
>> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]>
>> >> > wrote:
>> >> >>
>> >> >> Hi,
>> >> >>
>> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know about
>> >> >> those
>> >> >> version of Theano:
>> >> >>
>> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
>> >> >>     subscription), 0.5.0rc1 (no subscription)
>> >> >>
>> >> >>
>> >> >> It don't even know about Theano 0.5 that was release in February
>> >> >> 2012.
>> >> >>
>> >> >> It would be great if you add version 0.6 as available. To my
>> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac and
>> >> >> linux.
>> >> >>
>> >> >> To my understanding Theano is included in the community section, so
>> >> >> without support from you. So that should not add extra work from
>> >> >> you.
>> >> >> If I'm mistaken, please tell me. I would like to make Theano easily
>> >> >> available and having only old version shown as being available isn't
>> >> >> good.
>> >> >>
>> >> >> If you need modification in Theano for any reason, do not hesitate
>> >> >> to
>> >> >> tell
>> >> >> me.
>> >> >>
>> >> >> thanks
>> >> >>
>> >> >> Frédéric Bastien
>> >> >> _______________________________________________
>> >> >> Enthought-Dev mailing list
>> >> >> [hidden email]
>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> >
>> >> >
>> >> > _______________________________________________
>> >> > Enthought-Dev mailing list
>> >> > [hidden email]
>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> _______________________________________________
>> >> Enthought-Dev mailing list
>> >> [hidden email]
>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> >
>> >
>> > _______________________________________________
>> > Enthought-Dev mailing list
>> > [hidden email]
>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

David Cournapeau



On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]> wrote:
Currently there isn't an easy way to test just a subset of Theano. The
only way is to specify directory/files/test name as a parameter to
nosetests.

This have been asked a few times to have a partial but faster test
suite. Personally, I would prefer keeping the default like now and
having the user specify to run just a part of the test suite. Will
this fit your system?

Sure.
 

I don't know the different way to mark tests as slow. Do you have a
recommended way to do this? I'll look at the way NumPy do this.

nose makes that pretty easy: a slow decorator that add the slow attribue to the passed callable, then you can call nose -a '!slow'.

That's how pandas does it, not sure about numpy.


My daughter is sick and I'm in vacation for the next 2 weeks, so I
don't know when I can have this ready.

Well, what about me doing it if the above strategy works for you :)

David

Fred


On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]> wrote:
>
>
>
> On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> Thanks for the good news!
>>
>> As always if you have any questions/comments don't hesitate to contact me.
>
>
> So I have a preliminary egg for Theano, but the test suite of ~ 20 minutes
> is prohibitive. Is there an option for testing only a subset of Theano ?
>
> For packaging, I care mostly about ensuring we have not introduced bugs when
> packaging things incorrectly, and I like running those everytime we update
> both forward and reverse dependencies of a given package.
>
> thanks,
> David
>>
>>
>> Frédéric
>>
>> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau <[hidden email]>
>> wrote:
>> > Hi Frédéric,
>> >
>> >
>> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>> > wrote:
>> >>
>> >> Hi,
>> >>
>> >> Theano is very useful without GPU. We have some paper that show that
>> >> the CPU only back-end was faster then numpy in many cases and faster
>> >> then numexpr too.
>> >>
>> >> Here is the most recent paper with benchmark but compare only to other
>> >> highly optimized packages on machine learning algo:
>> >>
>> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>> >>
>> >> Here is the first paper with comparison again numpy/scipy and numexpr:
>> >>
>> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf look
>> >> at page 3 and 4.
>> >>
>> >>
>> >> This is just to show that Theano without the GPU is very useful. In
>> >> fact we use it frequently here as we don't have enough GPUs and have
>> >> access to a few CPU clusters.
>> >>
>> >>
>> >> I saw a few other people from Enthought play/contribute with/to Theano
>> >> like Robert Kern. If you want you can talk about this with them.
>> >>
>> >> I don't know the detail of your supported status. Could you support it
>> >> on the CPU only? Or just refer to Theano installation for the GPU? On
>> >> linux and Mac, it isn't very complicated. But I agree that on Windows,
>> >> it is more complicated as the user need to install the microsoft
>> >> compiler in addition to nvcc. Officially, we don't support GPU on
>> >> Windows, but we help answer questions about this on the mailing when
>> >> people have problems, there is some instruction provided by user and
>> >> updated by others users. I have heard many people on the mailing tell
>> >> that they use that. In fact, when we make code change that MSVC don't
>> >> like(it is a more picky compiler then g++), users tell us about this
>> >> and frequently provide us with the fixes.
>> >
>> >
>> > Thanks for the detailed explanation.
>> >
>> > You convinced me it would be useful to add support for theano. It should
>> > be
>> > available within a few days, unless some unforeseen complications arise,
>> >
>> > thanks,
>> > David
>> >>
>> >>
>> >> Frédéric
>> >>
>> >>
>> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau <[hidden email]>
>> >> wrote:
>> >> > Hi Frédéric,
>> >> >
>> >> > I agree having ancient theano is not good.
>> >> >
>> >> > I was considering moving it up into a supported package, but I feared
>> >> > that
>> >> > it would not be so useful without GPU support. I (unfortunately)
>> >> > never
>> >> > had
>> >> > time to play with theano, and don't know how useful the package is
>> >> > without
>> >> > CUDA (and I am a bit worried about support pressure when people will
>> >> > ask
>> >> > 'GPU is not used, fix it').
>> >> >
>> >> > thanks,
>> >> > David
>> >> >
>> >> >
>> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]>
>> >> > wrote:
>> >> >>
>> >> >> Hi,
>> >> >>
>> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know about
>> >> >> those
>> >> >> version of Theano:
>> >> >>
>> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
>> >> >>     subscription), 0.5.0rc1 (no subscription)
>> >> >>
>> >> >>
>> >> >> It don't even know about Theano 0.5 that was release in February
>> >> >> 2012.
>> >> >>
>> >> >> It would be great if you add version 0.6 as available. To my
>> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac and
>> >> >> linux.
>> >> >>
>> >> >> To my understanding Theano is included in the community section, so
>> >> >> without support from you. So that should not add extra work from
>> >> >> you.
>> >> >> If I'm mistaken, please tell me. I would like to make Theano easily
>> >> >> available and having only old version shown as being available isn't
>> >> >> good.
>> >> >>
>> >> >> If you need modification in Theano for any reason, do not hesitate
>> >> >> to
>> >> >> tell
>> >> >> me.
>> >> >>
>> >> >> thanks
>> >> >>
>> >> >> Frédéric Bastien
>> >> >> _______________________________________________
>> >> >> Enthought-Dev mailing list
>> >> >> [hidden email]
>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> >
>> >> >
>> >> > _______________________________________________
>> >> > Enthought-Dev mailing list
>> >> > [hidden email]
>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> _______________________________________________
>> >> Enthought-Dev mailing list
>> >> [hidden email]
>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> >
>> >
>> > _______________________________________________
>> > Enthought-Dev mailing list
>> > [hidden email]
>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev


_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Arnd Baecker
On Wed, 18 Dec 2013, David Cournapeau wrote:

> On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]> wrote:
>       Currently there isn't an easy way to test just a subset of
>       Theano. The
>       only way is to specify directory/files/test name as a parameter
>       to
>       nosetests.
>
>       This have been asked a few times to have a partial but faster
>       test
>       suite. Personally, I would prefer keeping the default like now
>       and
>       having the user specify to run just a part of the test suite.
>       Will
>       this fit your system?
>
>
> Sure.
>  
>
>       I don't know the different way to mark tests as slow. Do you
>       have a
>       recommended way to do this? I'll look at the way NumPy do this.
>
>
> nose makes that pretty easy: a slow decorator that add the slow attribue to
> the passed callable, then you can call nose -a '!slow'.
>
> That's how pandas does it, not sure about numpy.
Maybe https://pypi.python.org/pypi/nose-timelimit
could be used? (I haven't tried it myself yet...)

Best, Arnd
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
In reply to this post by David Cournapeau
On Wed, Dec 18, 2013 at 9:43 AM, David Cournapeau <[hidden email]> wrote:

>
>
>
> On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> Currently there isn't an easy way to test just a subset of Theano. The
>> only way is to specify directory/files/test name as a parameter to
>> nosetests.
>>
>> This have been asked a few times to have a partial but faster test
>> suite. Personally, I would prefer keeping the default like now and
>> having the user specify to run just a part of the test suite. Will
>> this fit your system?
>
>
> Sure.
>
>>
>>
>> I don't know the different way to mark tests as slow. Do you have a
>> recommended way to do this? I'll look at the way NumPy do this.
>
>
> nose makes that pretty easy: a slow decorator that add the slow attribue to
> the passed callable, then you can call nose -a '!slow'.
>
> That's how pandas does it, not sure about numpy.
>
>>
>> My daughter is sick and I'm in vacation for the next 2 weeks, so I
>> don't know when I can have this ready.
>
>
> Well, what about me doing it if the above strategy works for you :)

Thanks. I will be happy that someone else do that. I'm good with the
strategy you told.

The command

theano-nose --time-profile

would output 3 files timeprof_*. They will contain the timing of each
test. That can help you find the long running timming.

In travis-ci config file, we split Theano tests into 4 parts. Two of
those part are theano/scan_module/ and the file
theano/tensor/tests/test_basic.py as they are long running. In the
test_basic.py file, there is many elemwise test for each scalar.
Skipping many of them would remove much coverage.

Also, there is 2 type of timming, with a theano cache of compiled c
code "full" or "empty". This make a big difference in tests run time.
I suppose you want the timming with the empty cache? For your final
timming you can empty the cache with this command:

theano-cache clear

to be sure the timming is as you would like. Much fast would you like it to be?

thanks

Fred

>
> David
>>
>>
>> Fred
>>
>>
>> On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]>
>> wrote:
>> >
>> >
>> >
>> > On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]>
>> > wrote:
>> >>
>> >> Thanks for the good news!
>> >>
>> >> As always if you have any questions/comments don't hesitate to contact
>> >> me.
>> >
>> >
>> > So I have a preliminary egg for Theano, but the test suite of ~ 20
>> > minutes
>> > is prohibitive. Is there an option for testing only a subset of Theano ?
>> >
>> > For packaging, I care mostly about ensuring we have not introduced bugs
>> > when
>> > packaging things incorrectly, and I like running those everytime we
>> > update
>> > both forward and reverse dependencies of a given package.
>> >
>> > thanks,
>> > David
>> >>
>> >>
>> >> Frédéric
>> >>
>> >> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau <[hidden email]>
>> >> wrote:
>> >> > Hi Frédéric,
>> >> >
>> >> >
>> >> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>> >> > wrote:
>> >> >>
>> >> >> Hi,
>> >> >>
>> >> >> Theano is very useful without GPU. We have some paper that show that
>> >> >> the CPU only back-end was faster then numpy in many cases and faster
>> >> >> then numexpr too.
>> >> >>
>> >> >> Here is the most recent paper with benchmark but compare only to
>> >> >> other
>> >> >> highly optimized packages on machine learning algo:
>> >> >>
>> >> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>> >> >>
>> >> >> Here is the first paper with comparison again numpy/scipy and
>> >> >> numexpr:
>> >> >>
>> >> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf
>> >> >> look
>> >> >> at page 3 and 4.
>> >> >>
>> >> >>
>> >> >> This is just to show that Theano without the GPU is very useful. In
>> >> >> fact we use it frequently here as we don't have enough GPUs and have
>> >> >> access to a few CPU clusters.
>> >> >>
>> >> >>
>> >> >> I saw a few other people from Enthought play/contribute with/to
>> >> >> Theano
>> >> >> like Robert Kern. If you want you can talk about this with them.
>> >> >>
>> >> >> I don't know the detail of your supported status. Could you support
>> >> >> it
>> >> >> on the CPU only? Or just refer to Theano installation for the GPU?
>> >> >> On
>> >> >> linux and Mac, it isn't very complicated. But I agree that on
>> >> >> Windows,
>> >> >> it is more complicated as the user need to install the microsoft
>> >> >> compiler in addition to nvcc. Officially, we don't support GPU on
>> >> >> Windows, but we help answer questions about this on the mailing when
>> >> >> people have problems, there is some instruction provided by user and
>> >> >> updated by others users. I have heard many people on the mailing
>> >> >> tell
>> >> >> that they use that. In fact, when we make code change that MSVC
>> >> >> don't
>> >> >> like(it is a more picky compiler then g++), users tell us about this
>> >> >> and frequently provide us with the fixes.
>> >> >
>> >> >
>> >> > Thanks for the detailed explanation.
>> >> >
>> >> > You convinced me it would be useful to add support for theano. It
>> >> > should
>> >> > be
>> >> > available within a few days, unless some unforeseen complications
>> >> > arise,
>> >> >
>> >> > thanks,
>> >> > David
>> >> >>
>> >> >>
>> >> >> Frédéric
>> >> >>
>> >> >>
>> >> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau
>> >> >> <[hidden email]>
>> >> >> wrote:
>> >> >> > Hi Frédéric,
>> >> >> >
>> >> >> > I agree having ancient theano is not good.
>> >> >> >
>> >> >> > I was considering moving it up into a supported package, but I
>> >> >> > feared
>> >> >> > that
>> >> >> > it would not be so useful without GPU support. I (unfortunately)
>> >> >> > never
>> >> >> > had
>> >> >> > time to play with theano, and don't know how useful the package is
>> >> >> > without
>> >> >> > CUDA (and I am a bit worried about support pressure when people
>> >> >> > will
>> >> >> > ask
>> >> >> > 'GPU is not used, fix it').
>> >> >> >
>> >> >> > thanks,
>> >> >> > David
>> >> >> >
>> >> >> >
>> >> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]>
>> >> >> > wrote:
>> >> >> >>
>> >> >> >> Hi,
>> >> >> >>
>> >> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know about
>> >> >> >> those
>> >> >> >> version of Theano:
>> >> >> >>
>> >> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
>> >> >> >>     subscription), 0.5.0rc1 (no subscription)
>> >> >> >>
>> >> >> >>
>> >> >> >> It don't even know about Theano 0.5 that was release in February
>> >> >> >> 2012.
>> >> >> >>
>> >> >> >> It would be great if you add version 0.6 as available. To my
>> >> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac and
>> >> >> >> linux.
>> >> >> >>
>> >> >> >> To my understanding Theano is included in the community section,
>> >> >> >> so
>> >> >> >> without support from you. So that should not add extra work from
>> >> >> >> you.
>> >> >> >> If I'm mistaken, please tell me. I would like to make Theano
>> >> >> >> easily
>> >> >> >> available and having only old version shown as being available
>> >> >> >> isn't
>> >> >> >> good.
>> >> >> >>
>> >> >> >> If you need modification in Theano for any reason, do not
>> >> >> >> hesitate
>> >> >> >> to
>> >> >> >> tell
>> >> >> >> me.
>> >> >> >>
>> >> >> >> thanks
>> >> >> >>
>> >> >> >> Frédéric Bastien
>> >> >> >> _______________________________________________
>> >> >> >> Enthought-Dev mailing list
>> >> >> >> [hidden email]
>> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >> >
>> >> >> >
>> >> >> >
>> >> >> > _______________________________________________
>> >> >> > Enthought-Dev mailing list
>> >> >> > [hidden email]
>> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >> >
>> >> >> _______________________________________________
>> >> >> Enthought-Dev mailing list
>> >> >> [hidden email]
>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> >
>> >> >
>> >> > _______________________________________________
>> >> > Enthought-Dev mailing list
>> >> > [hidden email]
>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> _______________________________________________
>> >> Enthought-Dev mailing list
>> >> [hidden email]
>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> >
>> >
>> > _______________________________________________
>> > Enthought-Dev mailing list
>> > [hidden email]
>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

David Cournapeau



On Wed, Dec 18, 2013 at 7:48 PM, Frédéric Bastien <[hidden email]> wrote:
On Wed, Dec 18, 2013 at 9:43 AM, David Cournapeau <[hidden email]> wrote:
>
>
>
> On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> Currently there isn't an easy way to test just a subset of Theano. The
>> only way is to specify directory/files/test name as a parameter to
>> nosetests.
>>
>> This have been asked a few times to have a partial but faster test
>> suite. Personally, I would prefer keeping the default like now and
>> having the user specify to run just a part of the test suite. Will
>> this fit your system?
>
>
> Sure.
>
>>
>>
>> I don't know the different way to mark tests as slow. Do you have a
>> recommended way to do this? I'll look at the way NumPy do this.
>
>
> nose makes that pretty easy: a slow decorator that add the slow attribue to
> the passed callable, then you can call nose -a '!slow'.
>
> That's how pandas does it, not sure about numpy.
>
>>
>> My daughter is sick and I'm in vacation for the next 2 weeks, so I
>> don't know when I can have this ready.
>
>
> Well, what about me doing it if the above strategy works for you :)

Thanks. I will be happy that someone else do that. I'm good with the
strategy you told.

The command

theano-nose --time-profile

would output 3 files timeprof_*. They will contain the timing of each
test. That can help you find the long running timming.

In travis-ci config file, we split Theano tests into 4 parts.

Ah, that's very useful information.

I see errors like the following:

======================================================================
ERROR: test_grad (theano.tensor.tests.test_basic.T_max_and_argmax)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py", line 2564, in test_grad
    [data])
  File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py", line 2540, in safe_verify_grad
    utt.verify_grad(func, data, eps=eps)
  File "/home/vagrant/src/tmp/Theano/theano/tests/unittest_tools.py", line 90, in verify_grad
    T.verify_grad(op, pt, n_tests, rng, *args, **kwargs)
  File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1470, in verify_grad
    eps, out_type)
  File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1233, in __init__
    f_x = f(*[p.copy() for p in apt])
  File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py", line 588, in __call__
    self.fn.thunks[self.fn.position_of_error])
  File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py", line 579, in __call__
    outputs = self.fn()
TypeError: ("expected type_num 9 (NPY_INT64) got 5\nApply node that caused the error: Elemwise{mul,no_inplace}(random_projection, argmax)\nInputs shapes: [(2,), (2,)]\nInputs strides: [(8,), (4,)]\nInputs types: [TensorType(float64, vector), TensorType(int64, vector)]\nUse the Theano flag 'exception_verbosity=high' for a debugprint of this apply node.", '\nThe error happened with the following inputs:', [array([[ 0.40087424,  0.68837329,  0.35302867],
       [ 0.45571611,  0.90291591, -0.97459361]])], '\nThe value of eps is:', 1e-07, '\nThe out_type is:', None)


Which I can't make much sense of I am afraid. This is on rh5 32 bits.

David
 
Two of
those part are theano/scan_module/ and the file
theano/tensor/tests/test_basic.py as they are long running. In the
test_basic.py file, there is many elemwise test for each scalar.
Skipping many of them would remove much coverage.

Also, there is 2 type of timming, with a theano cache of compiled c
code "full" or "empty". This make a big difference in tests run time.
I suppose you want the timming with the empty cache? For your final
timming you can empty the cache with this command:

theano-cache clear

to be sure the timming is as you would like. Much fast would you like it to be?

thanks

Fred

>
> David
>>
>>
>> Fred
>>
>>
>> On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]>
>> wrote:
>> >
>> >
>> >
>> > On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]>
>> > wrote:
>> >>
>> >> Thanks for the good news!
>> >>
>> >> As always if you have any questions/comments don't hesitate to contact
>> >> me.
>> >
>> >
>> > So I have a preliminary egg for Theano, but the test suite of ~ 20
>> > minutes
>> > is prohibitive. Is there an option for testing only a subset of Theano ?
>> >
>> > For packaging, I care mostly about ensuring we have not introduced bugs
>> > when
>> > packaging things incorrectly, and I like running those everytime we
>> > update
>> > both forward and reverse dependencies of a given package.
>> >
>> > thanks,
>> > David
>> >>
>> >>
>> >> Frédéric
>> >>
>> >> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau <[hidden email]>
>> >> wrote:
>> >> > Hi Frédéric,
>> >> >
>> >> >
>> >> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>> >> > wrote:
>> >> >>
>> >> >> Hi,
>> >> >>
>> >> >> Theano is very useful without GPU. We have some paper that show that
>> >> >> the CPU only back-end was faster then numpy in many cases and faster
>> >> >> then numexpr too.
>> >> >>
>> >> >> Here is the most recent paper with benchmark but compare only to
>> >> >> other
>> >> >> highly optimized packages on machine learning algo:
>> >> >>
>> >> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>> >> >>
>> >> >> Here is the first paper with comparison again numpy/scipy and
>> >> >> numexpr:
>> >> >>
>> >> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf
>> >> >> look
>> >> >> at page 3 and 4.
>> >> >>
>> >> >>
>> >> >> This is just to show that Theano without the GPU is very useful. In
>> >> >> fact we use it frequently here as we don't have enough GPUs and have
>> >> >> access to a few CPU clusters.
>> >> >>
>> >> >>
>> >> >> I saw a few other people from Enthought play/contribute with/to
>> >> >> Theano
>> >> >> like Robert Kern. If you want you can talk about this with them.
>> >> >>
>> >> >> I don't know the detail of your supported status. Could you support
>> >> >> it
>> >> >> on the CPU only? Or just refer to Theano installation for the GPU?
>> >> >> On
>> >> >> linux and Mac, it isn't very complicated. But I agree that on
>> >> >> Windows,
>> >> >> it is more complicated as the user need to install the microsoft
>> >> >> compiler in addition to nvcc. Officially, we don't support GPU on
>> >> >> Windows, but we help answer questions about this on the mailing when
>> >> >> people have problems, there is some instruction provided by user and
>> >> >> updated by others users. I have heard many people on the mailing
>> >> >> tell
>> >> >> that they use that. In fact, when we make code change that MSVC
>> >> >> don't
>> >> >> like(it is a more picky compiler then g++), users tell us about this
>> >> >> and frequently provide us with the fixes.
>> >> >
>> >> >
>> >> > Thanks for the detailed explanation.
>> >> >
>> >> > You convinced me it would be useful to add support for theano. It
>> >> > should
>> >> > be
>> >> > available within a few days, unless some unforeseen complications
>> >> > arise,
>> >> >
>> >> > thanks,
>> >> > David
>> >> >>
>> >> >>
>> >> >> Frédéric
>> >> >>
>> >> >>
>> >> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau
>> >> >> <[hidden email]>
>> >> >> wrote:
>> >> >> > Hi Frédéric,
>> >> >> >
>> >> >> > I agree having ancient theano is not good.
>> >> >> >
>> >> >> > I was considering moving it up into a supported package, but I
>> >> >> > feared
>> >> >> > that
>> >> >> > it would not be so useful without GPU support. I (unfortunately)
>> >> >> > never
>> >> >> > had
>> >> >> > time to play with theano, and don't know how useful the package is
>> >> >> > without
>> >> >> > CUDA (and I am a bit worried about support pressure when people
>> >> >> > will
>> >> >> > ask
>> >> >> > 'GPU is not used, fix it').
>> >> >> >
>> >> >> > thanks,
>> >> >> > David
>> >> >> >
>> >> >> >
>> >> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien <[hidden email]>
>> >> >> > wrote:
>> >> >> >>
>> >> >> >> Hi,
>> >> >> >>
>> >> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know about
>> >> >> >> those
>> >> >> >> version of Theano:
>> >> >> >>
>> >> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1 (no
>> >> >> >>     subscription), 0.5.0rc1 (no subscription)
>> >> >> >>
>> >> >> >>
>> >> >> >> It don't even know about Theano 0.5 that was release in February
>> >> >> >> 2012.
>> >> >> >>
>> >> >> >> It would be great if you add version 0.6 as available. To my
>> >> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac and
>> >> >> >> linux.
>> >> >> >>
>> >> >> >> To my understanding Theano is included in the community section,
>> >> >> >> so
>> >> >> >> without support from you. So that should not add extra work from
>> >> >> >> you.
>> >> >> >> If I'm mistaken, please tell me. I would like to make Theano
>> >> >> >> easily
>> >> >> >> available and having only old version shown as being available
>> >> >> >> isn't
>> >> >> >> good.
>> >> >> >>
>> >> >> >> If you need modification in Theano for any reason, do not
>> >> >> >> hesitate
>> >> >> >> to
>> >> >> >> tell
>> >> >> >> me.
>> >> >> >>
>> >> >> >> thanks
>> >> >> >>
>> >> >> >> Frédéric Bastien
>> >> >> >> _______________________________________________
>> >> >> >> Enthought-Dev mailing list
>> >> >> >> [hidden email]
>> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >> >
>> >> >> >
>> >> >> >
>> >> >> > _______________________________________________
>> >> >> > Enthought-Dev mailing list
>> >> >> > [hidden email]
>> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >> >
>> >> >> _______________________________________________
>> >> >> Enthought-Dev mailing list
>> >> >> [hidden email]
>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> >
>> >> >
>> >> > _______________________________________________
>> >> > Enthought-Dev mailing list
>> >> > [hidden email]
>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> _______________________________________________
>> >> Enthought-Dev mailing list
>> >> [hidden email]
>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> >
>> >
>> > _______________________________________________
>> > Enthought-Dev mailing list
>> > [hidden email]
>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev


_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
I suppose this is with the development version of Theano? I added c
code to MaxAndArgmax and forgot to upcast to int64 the output of
PyArray_ArgMax(). This affect only 32 bit python version. Theano
always use int64 for the shape, but not NumPy. I only test on 32 bit
python once in a while or before release.

Thanks for letting me know. A simple work around until I fix this in
the development version (I'll try to make the PR today), is t change
the method name of class MaxAndArgmax c_code() to _c_code() to disable
its c code.



Fred

On Fri, Dec 20, 2013 at 2:27 PM, David Cournapeau <[hidden email]> wrote:

>
>
>
> On Wed, Dec 18, 2013 at 7:48 PM, Frédéric Bastien <[hidden email]> wrote:
>>
>> On Wed, Dec 18, 2013 at 9:43 AM, David Cournapeau <[hidden email]>
>> wrote:
>> >
>> >
>> >
>> > On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]>
>> > wrote:
>> >>
>> >> Currently there isn't an easy way to test just a subset of Theano. The
>> >> only way is to specify directory/files/test name as a parameter to
>> >> nosetests.
>> >>
>> >> This have been asked a few times to have a partial but faster test
>> >> suite. Personally, I would prefer keeping the default like now and
>> >> having the user specify to run just a part of the test suite. Will
>> >> this fit your system?
>> >
>> >
>> > Sure.
>> >
>> >>
>> >>
>> >> I don't know the different way to mark tests as slow. Do you have a
>> >> recommended way to do this? I'll look at the way NumPy do this.
>> >
>> >
>> > nose makes that pretty easy: a slow decorator that add the slow attribue
>> > to
>> > the passed callable, then you can call nose -a '!slow'.
>> >
>> > That's how pandas does it, not sure about numpy.
>> >
>> >>
>> >> My daughter is sick and I'm in vacation for the next 2 weeks, so I
>> >> don't know when I can have this ready.
>> >
>> >
>> > Well, what about me doing it if the above strategy works for you :)
>>
>> Thanks. I will be happy that someone else do that. I'm good with the
>> strategy you told.
>>
>> The command
>>
>> theano-nose --time-profile
>>
>> would output 3 files timeprof_*. They will contain the timing of each
>> test. That can help you find the long running timming.
>>
>> In travis-ci config file, we split Theano tests into 4 parts.
>
>
> Ah, that's very useful information.
>
> I see errors like the following:
>
> ======================================================================
> ERROR: test_grad (theano.tensor.tests.test_basic.T_max_and_argmax)
> ----------------------------------------------------------------------
> Traceback (most recent call last):
>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
> line 2564, in test_grad
>     [data])
>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
> line 2540, in safe_verify_grad
>     utt.verify_grad(func, data, eps=eps)
>   File "/home/vagrant/src/tmp/Theano/theano/tests/unittest_tools.py", line
> 90, in verify_grad
>     T.verify_grad(op, pt, n_tests, rng, *args, **kwargs)
>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1470, in
> verify_grad
>     eps, out_type)
>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1233, in
> __init__
>     f_x = f(*[p.copy() for p in apt])
>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
> line 588, in __call__
>     self.fn.thunks[self.fn.position_of_error])
>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
> line 579, in __call__
>     outputs = self.fn()
> TypeError: ("expected type_num 9 (NPY_INT64) got 5\nApply node that caused
> the error: Elemwise{mul,no_inplace}(random_projection, argmax)\nInputs
> shapes: [(2,), (2,)]\nInputs strides: [(8,), (4,)]\nInputs types:
> [TensorType(float64, vector), TensorType(int64, vector)]\nUse the Theano
> flag 'exception_verbosity=high' for a debugprint of this apply node.",
> '\nThe error happened with the following inputs:', [array([[ 0.40087424,
> 0.68837329,  0.35302867],
>        [ 0.45571611,  0.90291591, -0.97459361]])], '\nThe value of eps is:',
> 1e-07, '\nThe out_type is:', None)
>
>
> Which I can't make much sense of I am afraid. This is on rh5 32 bits.
>
> David
>
>>
>> Two of
>> those part are theano/scan_module/ and the file
>> theano/tensor/tests/test_basic.py as they are long running. In the
>> test_basic.py file, there is many elemwise test for each scalar.
>> Skipping many of them would remove much coverage.
>>
>> Also, there is 2 type of timming, with a theano cache of compiled c
>> code "full" or "empty". This make a big difference in tests run time.
>> I suppose you want the timming with the empty cache? For your final
>> timming you can empty the cache with this command:
>>
>> theano-cache clear
>>
>> to be sure the timming is as you would like. Much fast would you like it
>> to be?
>>
>> thanks
>>
>> Fred
>>
>> >
>> > David
>> >>
>> >>
>> >> Fred
>> >>
>> >>
>> >> On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]>
>> >> wrote:
>> >> >
>> >> >
>> >> >
>> >> > On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]>
>> >> > wrote:
>> >> >>
>> >> >> Thanks for the good news!
>> >> >>
>> >> >> As always if you have any questions/comments don't hesitate to
>> >> >> contact
>> >> >> me.
>> >> >
>> >> >
>> >> > So I have a preliminary egg for Theano, but the test suite of ~ 20
>> >> > minutes
>> >> > is prohibitive. Is there an option for testing only a subset of
>> >> > Theano ?
>> >> >
>> >> > For packaging, I care mostly about ensuring we have not introduced
>> >> > bugs
>> >> > when
>> >> > packaging things incorrectly, and I like running those everytime we
>> >> > update
>> >> > both forward and reverse dependencies of a given package.
>> >> >
>> >> > thanks,
>> >> > David
>> >> >>
>> >> >>
>> >> >> Frédéric
>> >> >>
>> >> >> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau
>> >> >> <[hidden email]>
>> >> >> wrote:
>> >> >> > Hi Frédéric,
>> >> >> >
>> >> >> >
>> >> >> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>> >> >> > wrote:
>> >> >> >>
>> >> >> >> Hi,
>> >> >> >>
>> >> >> >> Theano is very useful without GPU. We have some paper that show
>> >> >> >> that
>> >> >> >> the CPU only back-end was faster then numpy in many cases and
>> >> >> >> faster
>> >> >> >> then numexpr too.
>> >> >> >>
>> >> >> >> Here is the most recent paper with benchmark but compare only to
>> >> >> >> other
>> >> >> >> highly optimized packages on machine learning algo:
>> >> >> >>
>> >> >> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>> >> >> >>
>> >> >> >> Here is the first paper with comparison again numpy/scipy and
>> >> >> >> numexpr:
>> >> >> >>
>> >> >> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf
>> >> >> >> look
>> >> >> >> at page 3 and 4.
>> >> >> >>
>> >> >> >>
>> >> >> >> This is just to show that Theano without the GPU is very useful.
>> >> >> >> In
>> >> >> >> fact we use it frequently here as we don't have enough GPUs and
>> >> >> >> have
>> >> >> >> access to a few CPU clusters.
>> >> >> >>
>> >> >> >>
>> >> >> >> I saw a few other people from Enthought play/contribute with/to
>> >> >> >> Theano
>> >> >> >> like Robert Kern. If you want you can talk about this with them.
>> >> >> >>
>> >> >> >> I don't know the detail of your supported status. Could you
>> >> >> >> support
>> >> >> >> it
>> >> >> >> on the CPU only? Or just refer to Theano installation for the
>> >> >> >> GPU?
>> >> >> >> On
>> >> >> >> linux and Mac, it isn't very complicated. But I agree that on
>> >> >> >> Windows,
>> >> >> >> it is more complicated as the user need to install the microsoft
>> >> >> >> compiler in addition to nvcc. Officially, we don't support GPU on
>> >> >> >> Windows, but we help answer questions about this on the mailing
>> >> >> >> when
>> >> >> >> people have problems, there is some instruction provided by user
>> >> >> >> and
>> >> >> >> updated by others users. I have heard many people on the mailing
>> >> >> >> tell
>> >> >> >> that they use that. In fact, when we make code change that MSVC
>> >> >> >> don't
>> >> >> >> like(it is a more picky compiler then g++), users tell us about
>> >> >> >> this
>> >> >> >> and frequently provide us with the fixes.
>> >> >> >
>> >> >> >
>> >> >> > Thanks for the detailed explanation.
>> >> >> >
>> >> >> > You convinced me it would be useful to add support for theano. It
>> >> >> > should
>> >> >> > be
>> >> >> > available within a few days, unless some unforeseen complications
>> >> >> > arise,
>> >> >> >
>> >> >> > thanks,
>> >> >> > David
>> >> >> >>
>> >> >> >>
>> >> >> >> Frédéric
>> >> >> >>
>> >> >> >>
>> >> >> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau
>> >> >> >> <[hidden email]>
>> >> >> >> wrote:
>> >> >> >> > Hi Frédéric,
>> >> >> >> >
>> >> >> >> > I agree having ancient theano is not good.
>> >> >> >> >
>> >> >> >> > I was considering moving it up into a supported package, but I
>> >> >> >> > feared
>> >> >> >> > that
>> >> >> >> > it would not be so useful without GPU support. I
>> >> >> >> > (unfortunately)
>> >> >> >> > never
>> >> >> >> > had
>> >> >> >> > time to play with theano, and don't know how useful the package
>> >> >> >> > is
>> >> >> >> > without
>> >> >> >> > CUDA (and I am a bit worried about support pressure when people
>> >> >> >> > will
>> >> >> >> > ask
>> >> >> >> > 'GPU is not used, fix it').
>> >> >> >> >
>> >> >> >> > thanks,
>> >> >> >> > David
>> >> >> >> >
>> >> >> >> >
>> >> >> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien
>> >> >> >> > <[hidden email]>
>> >> >> >> > wrote:
>> >> >> >> >>
>> >> >> >> >> Hi,
>> >> >> >> >>
>> >> >> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know
>> >> >> >> >> about
>> >> >> >> >> those
>> >> >> >> >> version of Theano:
>> >> >> >> >>
>> >> >> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1
>> >> >> >> >> (no
>> >> >> >> >>     subscription), 0.5.0rc1 (no subscription)
>> >> >> >> >>
>> >> >> >> >>
>> >> >> >> >> It don't even know about Theano 0.5 that was release in
>> >> >> >> >> February
>> >> >> >> >> 2012.
>> >> >> >> >>
>> >> >> >> >> It would be great if you add version 0.6 as available. To my
>> >> >> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac
>> >> >> >> >> and
>> >> >> >> >> linux.
>> >> >> >> >>
>> >> >> >> >> To my understanding Theano is included in the community
>> >> >> >> >> section,
>> >> >> >> >> so
>> >> >> >> >> without support from you. So that should not add extra work
>> >> >> >> >> from
>> >> >> >> >> you.
>> >> >> >> >> If I'm mistaken, please tell me. I would like to make Theano
>> >> >> >> >> easily
>> >> >> >> >> available and having only old version shown as being available
>> >> >> >> >> isn't
>> >> >> >> >> good.
>> >> >> >> >>
>> >> >> >> >> If you need modification in Theano for any reason, do not
>> >> >> >> >> hesitate
>> >> >> >> >> to
>> >> >> >> >> tell
>> >> >> >> >> me.
>> >> >> >> >>
>> >> >> >> >> thanks
>> >> >> >> >>
>> >> >> >> >> Frédéric Bastien
>> >> >> >> >> _______________________________________________
>> >> >> >> >> Enthought-Dev mailing list
>> >> >> >> >> [hidden email]
>> >> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >> >> >
>> >> >> >> >
>> >> >> >> >
>> >> >> >> > _______________________________________________
>> >> >> >> > Enthought-Dev mailing list
>> >> >> >> > [hidden email]
>> >> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >> >> >
>> >> >> >> _______________________________________________
>> >> >> >> Enthought-Dev mailing list
>> >> >> >> [hidden email]
>> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >> >
>> >> >> >
>> >> >> >
>> >> >> > _______________________________________________
>> >> >> > Enthought-Dev mailing list
>> >> >> > [hidden email]
>> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >> >
>> >> >> _______________________________________________
>> >> >> Enthought-Dev mailing list
>> >> >> [hidden email]
>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> >
>> >> >
>> >> > _______________________________________________
>> >> > Enthought-Dev mailing list
>> >> > [hidden email]
>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >> >
>> >> _______________________________________________
>> >> Enthought-Dev mailing list
>> >> [hidden email]
>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> >
>> >
>> > _______________________________________________
>> > Enthought-Dev mailing list
>> > [hidden email]
>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>> >
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
>
>
> _______________________________________________
> Enthought-Dev mailing list
> [hidden email]
> https://mail.enthought.com/mailman/listinfo/enthought-dev
>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
Here is the PR with a fix:

https://github.com/Theano/Theano/pull/1682

I don't know when it will get merged.

I'm rerunning our tests on a 32 bit VM just in case something else happened.

thanks

Fred

On Fri, Dec 20, 2013 at 3:34 PM, Frédéric Bastien <[hidden email]> wrote:

> I suppose this is with the development version of Theano? I added c
> code to MaxAndArgmax and forgot to upcast to int64 the output of
> PyArray_ArgMax(). This affect only 32 bit python version. Theano
> always use int64 for the shape, but not NumPy. I only test on 32 bit
> python once in a while or before release.
>
> Thanks for letting me know. A simple work around until I fix this in
> the development version (I'll try to make the PR today), is t change
> the method name of class MaxAndArgmax c_code() to _c_code() to disable
> its c code.
>
>
>
> Fred
>
> On Fri, Dec 20, 2013 at 2:27 PM, David Cournapeau <[hidden email]> wrote:
>>
>>
>>
>> On Wed, Dec 18, 2013 at 7:48 PM, Frédéric Bastien <[hidden email]> wrote:
>>>
>>> On Wed, Dec 18, 2013 at 9:43 AM, David Cournapeau <[hidden email]>
>>> wrote:
>>> >
>>> >
>>> >
>>> > On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]>
>>> > wrote:
>>> >>
>>> >> Currently there isn't an easy way to test just a subset of Theano. The
>>> >> only way is to specify directory/files/test name as a parameter to
>>> >> nosetests.
>>> >>
>>> >> This have been asked a few times to have a partial but faster test
>>> >> suite. Personally, I would prefer keeping the default like now and
>>> >> having the user specify to run just a part of the test suite. Will
>>> >> this fit your system?
>>> >
>>> >
>>> > Sure.
>>> >
>>> >>
>>> >>
>>> >> I don't know the different way to mark tests as slow. Do you have a
>>> >> recommended way to do this? I'll look at the way NumPy do this.
>>> >
>>> >
>>> > nose makes that pretty easy: a slow decorator that add the slow attribue
>>> > to
>>> > the passed callable, then you can call nose -a '!slow'.
>>> >
>>> > That's how pandas does it, not sure about numpy.
>>> >
>>> >>
>>> >> My daughter is sick and I'm in vacation for the next 2 weeks, so I
>>> >> don't know when I can have this ready.
>>> >
>>> >
>>> > Well, what about me doing it if the above strategy works for you :)
>>>
>>> Thanks. I will be happy that someone else do that. I'm good with the
>>> strategy you told.
>>>
>>> The command
>>>
>>> theano-nose --time-profile
>>>
>>> would output 3 files timeprof_*. They will contain the timing of each
>>> test. That can help you find the long running timming.
>>>
>>> In travis-ci config file, we split Theano tests into 4 parts.
>>
>>
>> Ah, that's very useful information.
>>
>> I see errors like the following:
>>
>> ======================================================================
>> ERROR: test_grad (theano.tensor.tests.test_basic.T_max_and_argmax)
>> ----------------------------------------------------------------------
>> Traceback (most recent call last):
>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>> line 2564, in test_grad
>>     [data])
>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>> line 2540, in safe_verify_grad
>>     utt.verify_grad(func, data, eps=eps)
>>   File "/home/vagrant/src/tmp/Theano/theano/tests/unittest_tools.py", line
>> 90, in verify_grad
>>     T.verify_grad(op, pt, n_tests, rng, *args, **kwargs)
>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1470, in
>> verify_grad
>>     eps, out_type)
>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1233, in
>> __init__
>>     f_x = f(*[p.copy() for p in apt])
>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>> line 588, in __call__
>>     self.fn.thunks[self.fn.position_of_error])
>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>> line 579, in __call__
>>     outputs = self.fn()
>> TypeError: ("expected type_num 9 (NPY_INT64) got 5\nApply node that caused
>> the error: Elemwise{mul,no_inplace}(random_projection, argmax)\nInputs
>> shapes: [(2,), (2,)]\nInputs strides: [(8,), (4,)]\nInputs types:
>> [TensorType(float64, vector), TensorType(int64, vector)]\nUse the Theano
>> flag 'exception_verbosity=high' for a debugprint of this apply node.",
>> '\nThe error happened with the following inputs:', [array([[ 0.40087424,
>> 0.68837329,  0.35302867],
>>        [ 0.45571611,  0.90291591, -0.97459361]])], '\nThe value of eps is:',
>> 1e-07, '\nThe out_type is:', None)
>>
>>
>> Which I can't make much sense of I am afraid. This is on rh5 32 bits.
>>
>> David
>>
>>>
>>> Two of
>>> those part are theano/scan_module/ and the file
>>> theano/tensor/tests/test_basic.py as they are long running. In the
>>> test_basic.py file, there is many elemwise test for each scalar.
>>> Skipping many of them would remove much coverage.
>>>
>>> Also, there is 2 type of timming, with a theano cache of compiled c
>>> code "full" or "empty". This make a big difference in tests run time.
>>> I suppose you want the timming with the empty cache? For your final
>>> timming you can empty the cache with this command:
>>>
>>> theano-cache clear
>>>
>>> to be sure the timming is as you would like. Much fast would you like it
>>> to be?
>>>
>>> thanks
>>>
>>> Fred
>>>
>>> >
>>> > David
>>> >>
>>> >>
>>> >> Fred
>>> >>
>>> >>
>>> >> On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]>
>>> >> wrote:
>>> >> >
>>> >> >
>>> >> >
>>> >> > On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]>
>>> >> > wrote:
>>> >> >>
>>> >> >> Thanks for the good news!
>>> >> >>
>>> >> >> As always if you have any questions/comments don't hesitate to
>>> >> >> contact
>>> >> >> me.
>>> >> >
>>> >> >
>>> >> > So I have a preliminary egg for Theano, but the test suite of ~ 20
>>> >> > minutes
>>> >> > is prohibitive. Is there an option for testing only a subset of
>>> >> > Theano ?
>>> >> >
>>> >> > For packaging, I care mostly about ensuring we have not introduced
>>> >> > bugs
>>> >> > when
>>> >> > packaging things incorrectly, and I like running those everytime we
>>> >> > update
>>> >> > both forward and reverse dependencies of a given package.
>>> >> >
>>> >> > thanks,
>>> >> > David
>>> >> >>
>>> >> >>
>>> >> >> Frédéric
>>> >> >>
>>> >> >> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau
>>> >> >> <[hidden email]>
>>> >> >> wrote:
>>> >> >> > Hi Frédéric,
>>> >> >> >
>>> >> >> >
>>> >> >> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>>> >> >> > wrote:
>>> >> >> >>
>>> >> >> >> Hi,
>>> >> >> >>
>>> >> >> >> Theano is very useful without GPU. We have some paper that show
>>> >> >> >> that
>>> >> >> >> the CPU only back-end was faster then numpy in many cases and
>>> >> >> >> faster
>>> >> >> >> then numexpr too.
>>> >> >> >>
>>> >> >> >> Here is the most recent paper with benchmark but compare only to
>>> >> >> >> other
>>> >> >> >> highly optimized packages on machine learning algo:
>>> >> >> >>
>>> >> >> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>>> >> >> >>
>>> >> >> >> Here is the first paper with comparison again numpy/scipy and
>>> >> >> >> numexpr:
>>> >> >> >>
>>> >> >> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf
>>> >> >> >> look
>>> >> >> >> at page 3 and 4.
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> This is just to show that Theano without the GPU is very useful.
>>> >> >> >> In
>>> >> >> >> fact we use it frequently here as we don't have enough GPUs and
>>> >> >> >> have
>>> >> >> >> access to a few CPU clusters.
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> I saw a few other people from Enthought play/contribute with/to
>>> >> >> >> Theano
>>> >> >> >> like Robert Kern. If you want you can talk about this with them.
>>> >> >> >>
>>> >> >> >> I don't know the detail of your supported status. Could you
>>> >> >> >> support
>>> >> >> >> it
>>> >> >> >> on the CPU only? Or just refer to Theano installation for the
>>> >> >> >> GPU?
>>> >> >> >> On
>>> >> >> >> linux and Mac, it isn't very complicated. But I agree that on
>>> >> >> >> Windows,
>>> >> >> >> it is more complicated as the user need to install the microsoft
>>> >> >> >> compiler in addition to nvcc. Officially, we don't support GPU on
>>> >> >> >> Windows, but we help answer questions about this on the mailing
>>> >> >> >> when
>>> >> >> >> people have problems, there is some instruction provided by user
>>> >> >> >> and
>>> >> >> >> updated by others users. I have heard many people on the mailing
>>> >> >> >> tell
>>> >> >> >> that they use that. In fact, when we make code change that MSVC
>>> >> >> >> don't
>>> >> >> >> like(it is a more picky compiler then g++), users tell us about
>>> >> >> >> this
>>> >> >> >> and frequently provide us with the fixes.
>>> >> >> >
>>> >> >> >
>>> >> >> > Thanks for the detailed explanation.
>>> >> >> >
>>> >> >> > You convinced me it would be useful to add support for theano. It
>>> >> >> > should
>>> >> >> > be
>>> >> >> > available within a few days, unless some unforeseen complications
>>> >> >> > arise,
>>> >> >> >
>>> >> >> > thanks,
>>> >> >> > David
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> Frédéric
>>> >> >> >>
>>> >> >> >>
>>> >> >> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau
>>> >> >> >> <[hidden email]>
>>> >> >> >> wrote:
>>> >> >> >> > Hi Frédéric,
>>> >> >> >> >
>>> >> >> >> > I agree having ancient theano is not good.
>>> >> >> >> >
>>> >> >> >> > I was considering moving it up into a supported package, but I
>>> >> >> >> > feared
>>> >> >> >> > that
>>> >> >> >> > it would not be so useful without GPU support. I
>>> >> >> >> > (unfortunately)
>>> >> >> >> > never
>>> >> >> >> > had
>>> >> >> >> > time to play with theano, and don't know how useful the package
>>> >> >> >> > is
>>> >> >> >> > without
>>> >> >> >> > CUDA (and I am a bit worried about support pressure when people
>>> >> >> >> > will
>>> >> >> >> > ask
>>> >> >> >> > 'GPU is not used, fix it').
>>> >> >> >> >
>>> >> >> >> > thanks,
>>> >> >> >> > David
>>> >> >> >> >
>>> >> >> >> >
>>> >> >> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien
>>> >> >> >> > <[hidden email]>
>>> >> >> >> > wrote:
>>> >> >> >> >>
>>> >> >> >> >> Hi,
>>> >> >> >> >>
>>> >> >> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know
>>> >> >> >> >> about
>>> >> >> >> >> those
>>> >> >> >> >> version of Theano:
>>> >> >> >> >>
>>> >> >> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1
>>> >> >> >> >> (no
>>> >> >> >> >>     subscription), 0.5.0rc1 (no subscription)
>>> >> >> >> >>
>>> >> >> >> >>
>>> >> >> >> >> It don't even know about Theano 0.5 that was release in
>>> >> >> >> >> February
>>> >> >> >> >> 2012.
>>> >> >> >> >>
>>> >> >> >> >> It would be great if you add version 0.6 as available. To my
>>> >> >> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac
>>> >> >> >> >> and
>>> >> >> >> >> linux.
>>> >> >> >> >>
>>> >> >> >> >> To my understanding Theano is included in the community
>>> >> >> >> >> section,
>>> >> >> >> >> so
>>> >> >> >> >> without support from you. So that should not add extra work
>>> >> >> >> >> from
>>> >> >> >> >> you.
>>> >> >> >> >> If I'm mistaken, please tell me. I would like to make Theano
>>> >> >> >> >> easily
>>> >> >> >> >> available and having only old version shown as being available
>>> >> >> >> >> isn't
>>> >> >> >> >> good.
>>> >> >> >> >>
>>> >> >> >> >> If you need modification in Theano for any reason, do not
>>> >> >> >> >> hesitate
>>> >> >> >> >> to
>>> >> >> >> >> tell
>>> >> >> >> >> me.
>>> >> >> >> >>
>>> >> >> >> >> thanks
>>> >> >> >> >>
>>> >> >> >> >> Frédéric Bastien
>>> >> >> >> >> _______________________________________________
>>> >> >> >> >> Enthought-Dev mailing list
>>> >> >> >> >> [hidden email]
>>> >> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>> >> >> >> >
>>> >> >> >> >
>>> >> >> >> >
>>> >> >> >> > _______________________________________________
>>> >> >> >> > Enthought-Dev mailing list
>>> >> >> >> > [hidden email]
>>> >> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>> >> >> >> >
>>> >> >> >> _______________________________________________
>>> >> >> >> Enthought-Dev mailing list
>>> >> >> >> [hidden email]
>>> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>> >> >> >
>>> >> >> >
>>> >> >> >
>>> >> >> > _______________________________________________
>>> >> >> > Enthought-Dev mailing list
>>> >> >> > [hidden email]
>>> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>> >> >> >
>>> >> >> _______________________________________________
>>> >> >> Enthought-Dev mailing list
>>> >> >> [hidden email]
>>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>> >> >
>>> >> >
>>> >> >
>>> >> > _______________________________________________
>>> >> > Enthought-Dev mailing list
>>> >> > [hidden email]
>>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>> >> >
>>> >> _______________________________________________
>>> >> Enthought-Dev mailing list
>>> >> [hidden email]
>>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>> >
>>> >
>>> >
>>> > _______________________________________________
>>> > Enthought-Dev mailing list
>>> > [hidden email]
>>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>> >
>>> _______________________________________________
>>> Enthought-Dev mailing list
>>> [hidden email]
>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>
>>
>>
>> _______________________________________________
>> Enthought-Dev mailing list
>> [hidden email]
>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
The PR is fixed.

Fred

On Fri, Dec 20, 2013 at 4:06 PM, Frédéric Bastien <[hidden email]> wrote:

> Here is the PR with a fix:
>
> https://github.com/Theano/Theano/pull/1682
>
> I don't know when it will get merged.
>
> I'm rerunning our tests on a 32 bit VM just in case something else happened.
>
> thanks
>
> Fred
>
> On Fri, Dec 20, 2013 at 3:34 PM, Frédéric Bastien <[hidden email]> wrote:
>> I suppose this is with the development version of Theano? I added c
>> code to MaxAndArgmax and forgot to upcast to int64 the output of
>> PyArray_ArgMax(). This affect only 32 bit python version. Theano
>> always use int64 for the shape, but not NumPy. I only test on 32 bit
>> python once in a while or before release.
>>
>> Thanks for letting me know. A simple work around until I fix this in
>> the development version (I'll try to make the PR today), is t change
>> the method name of class MaxAndArgmax c_code() to _c_code() to disable
>> its c code.
>>
>>
>>
>> Fred
>>
>> On Fri, Dec 20, 2013 at 2:27 PM, David Cournapeau <[hidden email]> wrote:
>>>
>>>
>>>
>>> On Wed, Dec 18, 2013 at 7:48 PM, Frédéric Bastien <[hidden email]> wrote:
>>>>
>>>> On Wed, Dec 18, 2013 at 9:43 AM, David Cournapeau <[hidden email]>
>>>> wrote:
>>>> >
>>>> >
>>>> >
>>>> > On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]>
>>>> > wrote:
>>>> >>
>>>> >> Currently there isn't an easy way to test just a subset of Theano. The
>>>> >> only way is to specify directory/files/test name as a parameter to
>>>> >> nosetests.
>>>> >>
>>>> >> This have been asked a few times to have a partial but faster test
>>>> >> suite. Personally, I would prefer keeping the default like now and
>>>> >> having the user specify to run just a part of the test suite. Will
>>>> >> this fit your system?
>>>> >
>>>> >
>>>> > Sure.
>>>> >
>>>> >>
>>>> >>
>>>> >> I don't know the different way to mark tests as slow. Do you have a
>>>> >> recommended way to do this? I'll look at the way NumPy do this.
>>>> >
>>>> >
>>>> > nose makes that pretty easy: a slow decorator that add the slow attribue
>>>> > to
>>>> > the passed callable, then you can call nose -a '!slow'.
>>>> >
>>>> > That's how pandas does it, not sure about numpy.
>>>> >
>>>> >>
>>>> >> My daughter is sick and I'm in vacation for the next 2 weeks, so I
>>>> >> don't know when I can have this ready.
>>>> >
>>>> >
>>>> > Well, what about me doing it if the above strategy works for you :)
>>>>
>>>> Thanks. I will be happy that someone else do that. I'm good with the
>>>> strategy you told.
>>>>
>>>> The command
>>>>
>>>> theano-nose --time-profile
>>>>
>>>> would output 3 files timeprof_*. They will contain the timing of each
>>>> test. That can help you find the long running timming.
>>>>
>>>> In travis-ci config file, we split Theano tests into 4 parts.
>>>
>>>
>>> Ah, that's very useful information.
>>>
>>> I see errors like the following:
>>>
>>> ======================================================================
>>> ERROR: test_grad (theano.tensor.tests.test_basic.T_max_and_argmax)
>>> ----------------------------------------------------------------------
>>> Traceback (most recent call last):
>>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>>> line 2564, in test_grad
>>>     [data])
>>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>>> line 2540, in safe_verify_grad
>>>     utt.verify_grad(func, data, eps=eps)
>>>   File "/home/vagrant/src/tmp/Theano/theano/tests/unittest_tools.py", line
>>> 90, in verify_grad
>>>     T.verify_grad(op, pt, n_tests, rng, *args, **kwargs)
>>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1470, in
>>> verify_grad
>>>     eps, out_type)
>>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1233, in
>>> __init__
>>>     f_x = f(*[p.copy() for p in apt])
>>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>>> line 588, in __call__
>>>     self.fn.thunks[self.fn.position_of_error])
>>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>>> line 579, in __call__
>>>     outputs = self.fn()
>>> TypeError: ("expected type_num 9 (NPY_INT64) got 5\nApply node that caused
>>> the error: Elemwise{mul,no_inplace}(random_projection, argmax)\nInputs
>>> shapes: [(2,), (2,)]\nInputs strides: [(8,), (4,)]\nInputs types:
>>> [TensorType(float64, vector), TensorType(int64, vector)]\nUse the Theano
>>> flag 'exception_verbosity=high' for a debugprint of this apply node.",
>>> '\nThe error happened with the following inputs:', [array([[ 0.40087424,
>>> 0.68837329,  0.35302867],
>>>        [ 0.45571611,  0.90291591, -0.97459361]])], '\nThe value of eps is:',
>>> 1e-07, '\nThe out_type is:', None)
>>>
>>>
>>> Which I can't make much sense of I am afraid. This is on rh5 32 bits.
>>>
>>> David
>>>
>>>>
>>>> Two of
>>>> those part are theano/scan_module/ and the file
>>>> theano/tensor/tests/test_basic.py as they are long running. In the
>>>> test_basic.py file, there is many elemwise test for each scalar.
>>>> Skipping many of them would remove much coverage.
>>>>
>>>> Also, there is 2 type of timming, with a theano cache of compiled c
>>>> code "full" or "empty". This make a big difference in tests run time.
>>>> I suppose you want the timming with the empty cache? For your final
>>>> timming you can empty the cache with this command:
>>>>
>>>> theano-cache clear
>>>>
>>>> to be sure the timming is as you would like. Much fast would you like it
>>>> to be?
>>>>
>>>> thanks
>>>>
>>>> Fred
>>>>
>>>> >
>>>> > David
>>>> >>
>>>> >>
>>>> >> Fred
>>>> >>
>>>> >>
>>>> >> On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]>
>>>> >> wrote:
>>>> >> >
>>>> >> >
>>>> >> >
>>>> >> > On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]>
>>>> >> > wrote:
>>>> >> >>
>>>> >> >> Thanks for the good news!
>>>> >> >>
>>>> >> >> As always if you have any questions/comments don't hesitate to
>>>> >> >> contact
>>>> >> >> me.
>>>> >> >
>>>> >> >
>>>> >> > So I have a preliminary egg for Theano, but the test suite of ~ 20
>>>> >> > minutes
>>>> >> > is prohibitive. Is there an option for testing only a subset of
>>>> >> > Theano ?
>>>> >> >
>>>> >> > For packaging, I care mostly about ensuring we have not introduced
>>>> >> > bugs
>>>> >> > when
>>>> >> > packaging things incorrectly, and I like running those everytime we
>>>> >> > update
>>>> >> > both forward and reverse dependencies of a given package.
>>>> >> >
>>>> >> > thanks,
>>>> >> > David
>>>> >> >>
>>>> >> >>
>>>> >> >> Frédéric
>>>> >> >>
>>>> >> >> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau
>>>> >> >> <[hidden email]>
>>>> >> >> wrote:
>>>> >> >> > Hi Frédéric,
>>>> >> >> >
>>>> >> >> >
>>>> >> >> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>>>> >> >> > wrote:
>>>> >> >> >>
>>>> >> >> >> Hi,
>>>> >> >> >>
>>>> >> >> >> Theano is very useful without GPU. We have some paper that show
>>>> >> >> >> that
>>>> >> >> >> the CPU only back-end was faster then numpy in many cases and
>>>> >> >> >> faster
>>>> >> >> >> then numexpr too.
>>>> >> >> >>
>>>> >> >> >> Here is the most recent paper with benchmark but compare only to
>>>> >> >> >> other
>>>> >> >> >> highly optimized packages on machine learning algo:
>>>> >> >> >>
>>>> >> >> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>>>> >> >> >>
>>>> >> >> >> Here is the first paper with comparison again numpy/scipy and
>>>> >> >> >> numexpr:
>>>> >> >> >>
>>>> >> >> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf
>>>> >> >> >> look
>>>> >> >> >> at page 3 and 4.
>>>> >> >> >>
>>>> >> >> >>
>>>> >> >> >> This is just to show that Theano without the GPU is very useful.
>>>> >> >> >> In
>>>> >> >> >> fact we use it frequently here as we don't have enough GPUs and
>>>> >> >> >> have
>>>> >> >> >> access to a few CPU clusters.
>>>> >> >> >>
>>>> >> >> >>
>>>> >> >> >> I saw a few other people from Enthought play/contribute with/to
>>>> >> >> >> Theano
>>>> >> >> >> like Robert Kern. If you want you can talk about this with them.
>>>> >> >> >>
>>>> >> >> >> I don't know the detail of your supported status. Could you
>>>> >> >> >> support
>>>> >> >> >> it
>>>> >> >> >> on the CPU only? Or just refer to Theano installation for the
>>>> >> >> >> GPU?
>>>> >> >> >> On
>>>> >> >> >> linux and Mac, it isn't very complicated. But I agree that on
>>>> >> >> >> Windows,
>>>> >> >> >> it is more complicated as the user need to install the microsoft
>>>> >> >> >> compiler in addition to nvcc. Officially, we don't support GPU on
>>>> >> >> >> Windows, but we help answer questions about this on the mailing
>>>> >> >> >> when
>>>> >> >> >> people have problems, there is some instruction provided by user
>>>> >> >> >> and
>>>> >> >> >> updated by others users. I have heard many people on the mailing
>>>> >> >> >> tell
>>>> >> >> >> that they use that. In fact, when we make code change that MSVC
>>>> >> >> >> don't
>>>> >> >> >> like(it is a more picky compiler then g++), users tell us about
>>>> >> >> >> this
>>>> >> >> >> and frequently provide us with the fixes.
>>>> >> >> >
>>>> >> >> >
>>>> >> >> > Thanks for the detailed explanation.
>>>> >> >> >
>>>> >> >> > You convinced me it would be useful to add support for theano. It
>>>> >> >> > should
>>>> >> >> > be
>>>> >> >> > available within a few days, unless some unforeseen complications
>>>> >> >> > arise,
>>>> >> >> >
>>>> >> >> > thanks,
>>>> >> >> > David
>>>> >> >> >>
>>>> >> >> >>
>>>> >> >> >> Frédéric
>>>> >> >> >>
>>>> >> >> >>
>>>> >> >> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau
>>>> >> >> >> <[hidden email]>
>>>> >> >> >> wrote:
>>>> >> >> >> > Hi Frédéric,
>>>> >> >> >> >
>>>> >> >> >> > I agree having ancient theano is not good.
>>>> >> >> >> >
>>>> >> >> >> > I was considering moving it up into a supported package, but I
>>>> >> >> >> > feared
>>>> >> >> >> > that
>>>> >> >> >> > it would not be so useful without GPU support. I
>>>> >> >> >> > (unfortunately)
>>>> >> >> >> > never
>>>> >> >> >> > had
>>>> >> >> >> > time to play with theano, and don't know how useful the package
>>>> >> >> >> > is
>>>> >> >> >> > without
>>>> >> >> >> > CUDA (and I am a bit worried about support pressure when people
>>>> >> >> >> > will
>>>> >> >> >> > ask
>>>> >> >> >> > 'GPU is not used, fix it').
>>>> >> >> >> >
>>>> >> >> >> > thanks,
>>>> >> >> >> > David
>>>> >> >> >> >
>>>> >> >> >> >
>>>> >> >> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien
>>>> >> >> >> > <[hidden email]>
>>>> >> >> >> > wrote:
>>>> >> >> >> >>
>>>> >> >> >> >> Hi,
>>>> >> >> >> >>
>>>> >> >> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know
>>>> >> >> >> >> about
>>>> >> >> >> >> those
>>>> >> >> >> >> version of Theano:
>>>> >> >> >> >>
>>>> >> >> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1
>>>> >> >> >> >> (no
>>>> >> >> >> >>     subscription), 0.5.0rc1 (no subscription)
>>>> >> >> >> >>
>>>> >> >> >> >>
>>>> >> >> >> >> It don't even know about Theano 0.5 that was release in
>>>> >> >> >> >> February
>>>> >> >> >> >> 2012.
>>>> >> >> >> >>
>>>> >> >> >> >> It would be great if you add version 0.6 as available. To my
>>>> >> >> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac
>>>> >> >> >> >> and
>>>> >> >> >> >> linux.
>>>> >> >> >> >>
>>>> >> >> >> >> To my understanding Theano is included in the community
>>>> >> >> >> >> section,
>>>> >> >> >> >> so
>>>> >> >> >> >> without support from you. So that should not add extra work
>>>> >> >> >> >> from
>>>> >> >> >> >> you.
>>>> >> >> >> >> If I'm mistaken, please tell me. I would like to make Theano
>>>> >> >> >> >> easily
>>>> >> >> >> >> available and having only old version shown as being available
>>>> >> >> >> >> isn't
>>>> >> >> >> >> good.
>>>> >> >> >> >>
>>>> >> >> >> >> If you need modification in Theano for any reason, do not
>>>> >> >> >> >> hesitate
>>>> >> >> >> >> to
>>>> >> >> >> >> tell
>>>> >> >> >> >> me.
>>>> >> >> >> >>
>>>> >> >> >> >> thanks
>>>> >> >> >> >>
>>>> >> >> >> >> Frédéric Bastien
>>>> >> >> >> >> _______________________________________________
>>>> >> >> >> >> Enthought-Dev mailing list
>>>> >> >> >> >> [hidden email]
>>>> >> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>> >> >> >> >
>>>> >> >> >> >
>>>> >> >> >> >
>>>> >> >> >> > _______________________________________________
>>>> >> >> >> > Enthought-Dev mailing list
>>>> >> >> >> > [hidden email]
>>>> >> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>> >> >> >> >
>>>> >> >> >> _______________________________________________
>>>> >> >> >> Enthought-Dev mailing list
>>>> >> >> >> [hidden email]
>>>> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>> >> >> >
>>>> >> >> >
>>>> >> >> >
>>>> >> >> > _______________________________________________
>>>> >> >> > Enthought-Dev mailing list
>>>> >> >> > [hidden email]
>>>> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>> >> >> >
>>>> >> >> _______________________________________________
>>>> >> >> Enthought-Dev mailing list
>>>> >> >> [hidden email]
>>>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>> >> >
>>>> >> >
>>>> >> >
>>>> >> > _______________________________________________
>>>> >> > Enthought-Dev mailing list
>>>> >> > [hidden email]
>>>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>> >> >
>>>> >> _______________________________________________
>>>> >> Enthought-Dev mailing list
>>>> >> [hidden email]
>>>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>> >
>>>> >
>>>> >
>>>> > _______________________________________________
>>>> > Enthought-Dev mailing list
>>>> > [hidden email]
>>>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>> >
>>>> _______________________________________________
>>>> Enthought-Dev mailing list
>>>> [hidden email]
>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>
>>>
>>>
>>> _______________________________________________
>>> Enthought-Dev mailing list
>>> [hidden email]
>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
I mean, the PR is merged:)

Fred

On Fri, Dec 20, 2013 at 8:26 PM, Frédéric Bastien <[hidden email]> wrote:

> The PR is fixed.
>
> Fred
>
> On Fri, Dec 20, 2013 at 4:06 PM, Frédéric Bastien <[hidden email]> wrote:
>> Here is the PR with a fix:
>>
>> https://github.com/Theano/Theano/pull/1682
>>
>> I don't know when it will get merged.
>>
>> I'm rerunning our tests on a 32 bit VM just in case something else happened.
>>
>> thanks
>>
>> Fred
>>
>> On Fri, Dec 20, 2013 at 3:34 PM, Frédéric Bastien <[hidden email]> wrote:
>>> I suppose this is with the development version of Theano? I added c
>>> code to MaxAndArgmax and forgot to upcast to int64 the output of
>>> PyArray_ArgMax(). This affect only 32 bit python version. Theano
>>> always use int64 for the shape, but not NumPy. I only test on 32 bit
>>> python once in a while or before release.
>>>
>>> Thanks for letting me know. A simple work around until I fix this in
>>> the development version (I'll try to make the PR today), is t change
>>> the method name of class MaxAndArgmax c_code() to _c_code() to disable
>>> its c code.
>>>
>>>
>>>
>>> Fred
>>>
>>> On Fri, Dec 20, 2013 at 2:27 PM, David Cournapeau <[hidden email]> wrote:
>>>>
>>>>
>>>>
>>>> On Wed, Dec 18, 2013 at 7:48 PM, Frédéric Bastien <[hidden email]> wrote:
>>>>>
>>>>> On Wed, Dec 18, 2013 at 9:43 AM, David Cournapeau <[hidden email]>
>>>>> wrote:
>>>>> >
>>>>> >
>>>>> >
>>>>> > On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]>
>>>>> > wrote:
>>>>> >>
>>>>> >> Currently there isn't an easy way to test just a subset of Theano. The
>>>>> >> only way is to specify directory/files/test name as a parameter to
>>>>> >> nosetests.
>>>>> >>
>>>>> >> This have been asked a few times to have a partial but faster test
>>>>> >> suite. Personally, I would prefer keeping the default like now and
>>>>> >> having the user specify to run just a part of the test suite. Will
>>>>> >> this fit your system?
>>>>> >
>>>>> >
>>>>> > Sure.
>>>>> >
>>>>> >>
>>>>> >>
>>>>> >> I don't know the different way to mark tests as slow. Do you have a
>>>>> >> recommended way to do this? I'll look at the way NumPy do this.
>>>>> >
>>>>> >
>>>>> > nose makes that pretty easy: a slow decorator that add the slow attribue
>>>>> > to
>>>>> > the passed callable, then you can call nose -a '!slow'.
>>>>> >
>>>>> > That's how pandas does it, not sure about numpy.
>>>>> >
>>>>> >>
>>>>> >> My daughter is sick and I'm in vacation for the next 2 weeks, so I
>>>>> >> don't know when I can have this ready.
>>>>> >
>>>>> >
>>>>> > Well, what about me doing it if the above strategy works for you :)
>>>>>
>>>>> Thanks. I will be happy that someone else do that. I'm good with the
>>>>> strategy you told.
>>>>>
>>>>> The command
>>>>>
>>>>> theano-nose --time-profile
>>>>>
>>>>> would output 3 files timeprof_*. They will contain the timing of each
>>>>> test. That can help you find the long running timming.
>>>>>
>>>>> In travis-ci config file, we split Theano tests into 4 parts.
>>>>
>>>>
>>>> Ah, that's very useful information.
>>>>
>>>> I see errors like the following:
>>>>
>>>> ======================================================================
>>>> ERROR: test_grad (theano.tensor.tests.test_basic.T_max_and_argmax)
>>>> ----------------------------------------------------------------------
>>>> Traceback (most recent call last):
>>>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>>>> line 2564, in test_grad
>>>>     [data])
>>>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>>>> line 2540, in safe_verify_grad
>>>>     utt.verify_grad(func, data, eps=eps)
>>>>   File "/home/vagrant/src/tmp/Theano/theano/tests/unittest_tools.py", line
>>>> 90, in verify_grad
>>>>     T.verify_grad(op, pt, n_tests, rng, *args, **kwargs)
>>>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1470, in
>>>> verify_grad
>>>>     eps, out_type)
>>>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1233, in
>>>> __init__
>>>>     f_x = f(*[p.copy() for p in apt])
>>>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>>>> line 588, in __call__
>>>>     self.fn.thunks[self.fn.position_of_error])
>>>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>>>> line 579, in __call__
>>>>     outputs = self.fn()
>>>> TypeError: ("expected type_num 9 (NPY_INT64) got 5\nApply node that caused
>>>> the error: Elemwise{mul,no_inplace}(random_projection, argmax)\nInputs
>>>> shapes: [(2,), (2,)]\nInputs strides: [(8,), (4,)]\nInputs types:
>>>> [TensorType(float64, vector), TensorType(int64, vector)]\nUse the Theano
>>>> flag 'exception_verbosity=high' for a debugprint of this apply node.",
>>>> '\nThe error happened with the following inputs:', [array([[ 0.40087424,
>>>> 0.68837329,  0.35302867],
>>>>        [ 0.45571611,  0.90291591, -0.97459361]])], '\nThe value of eps is:',
>>>> 1e-07, '\nThe out_type is:', None)
>>>>
>>>>
>>>> Which I can't make much sense of I am afraid. This is on rh5 32 bits.
>>>>
>>>> David
>>>>
>>>>>
>>>>> Two of
>>>>> those part are theano/scan_module/ and the file
>>>>> theano/tensor/tests/test_basic.py as they are long running. In the
>>>>> test_basic.py file, there is many elemwise test for each scalar.
>>>>> Skipping many of them would remove much coverage.
>>>>>
>>>>> Also, there is 2 type of timming, with a theano cache of compiled c
>>>>> code "full" or "empty". This make a big difference in tests run time.
>>>>> I suppose you want the timming with the empty cache? For your final
>>>>> timming you can empty the cache with this command:
>>>>>
>>>>> theano-cache clear
>>>>>
>>>>> to be sure the timming is as you would like. Much fast would you like it
>>>>> to be?
>>>>>
>>>>> thanks
>>>>>
>>>>> Fred
>>>>>
>>>>> >
>>>>> > David
>>>>> >>
>>>>> >>
>>>>> >> Fred
>>>>> >>
>>>>> >>
>>>>> >> On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]>
>>>>> >> wrote:
>>>>> >> >
>>>>> >> >
>>>>> >> >
>>>>> >> > On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]>
>>>>> >> > wrote:
>>>>> >> >>
>>>>> >> >> Thanks for the good news!
>>>>> >> >>
>>>>> >> >> As always if you have any questions/comments don't hesitate to
>>>>> >> >> contact
>>>>> >> >> me.
>>>>> >> >
>>>>> >> >
>>>>> >> > So I have a preliminary egg for Theano, but the test suite of ~ 20
>>>>> >> > minutes
>>>>> >> > is prohibitive. Is there an option for testing only a subset of
>>>>> >> > Theano ?
>>>>> >> >
>>>>> >> > For packaging, I care mostly about ensuring we have not introduced
>>>>> >> > bugs
>>>>> >> > when
>>>>> >> > packaging things incorrectly, and I like running those everytime we
>>>>> >> > update
>>>>> >> > both forward and reverse dependencies of a given package.
>>>>> >> >
>>>>> >> > thanks,
>>>>> >> > David
>>>>> >> >>
>>>>> >> >>
>>>>> >> >> Frédéric
>>>>> >> >>
>>>>> >> >> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau
>>>>> >> >> <[hidden email]>
>>>>> >> >> wrote:
>>>>> >> >> > Hi Frédéric,
>>>>> >> >> >
>>>>> >> >> >
>>>>> >> >> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>>>>> >> >> > wrote:
>>>>> >> >> >>
>>>>> >> >> >> Hi,
>>>>> >> >> >>
>>>>> >> >> >> Theano is very useful without GPU. We have some paper that show
>>>>> >> >> >> that
>>>>> >> >> >> the CPU only back-end was faster then numpy in many cases and
>>>>> >> >> >> faster
>>>>> >> >> >> then numexpr too.
>>>>> >> >> >>
>>>>> >> >> >> Here is the most recent paper with benchmark but compare only to
>>>>> >> >> >> other
>>>>> >> >> >> highly optimized packages on machine learning algo:
>>>>> >> >> >>
>>>>> >> >> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>>>>> >> >> >>
>>>>> >> >> >> Here is the first paper with comparison again numpy/scipy and
>>>>> >> >> >> numexpr:
>>>>> >> >> >>
>>>>> >> >> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf
>>>>> >> >> >> look
>>>>> >> >> >> at page 3 and 4.
>>>>> >> >> >>
>>>>> >> >> >>
>>>>> >> >> >> This is just to show that Theano without the GPU is very useful.
>>>>> >> >> >> In
>>>>> >> >> >> fact we use it frequently here as we don't have enough GPUs and
>>>>> >> >> >> have
>>>>> >> >> >> access to a few CPU clusters.
>>>>> >> >> >>
>>>>> >> >> >>
>>>>> >> >> >> I saw a few other people from Enthought play/contribute with/to
>>>>> >> >> >> Theano
>>>>> >> >> >> like Robert Kern. If you want you can talk about this with them.
>>>>> >> >> >>
>>>>> >> >> >> I don't know the detail of your supported status. Could you
>>>>> >> >> >> support
>>>>> >> >> >> it
>>>>> >> >> >> on the CPU only? Or just refer to Theano installation for the
>>>>> >> >> >> GPU?
>>>>> >> >> >> On
>>>>> >> >> >> linux and Mac, it isn't very complicated. But I agree that on
>>>>> >> >> >> Windows,
>>>>> >> >> >> it is more complicated as the user need to install the microsoft
>>>>> >> >> >> compiler in addition to nvcc. Officially, we don't support GPU on
>>>>> >> >> >> Windows, but we help answer questions about this on the mailing
>>>>> >> >> >> when
>>>>> >> >> >> people have problems, there is some instruction provided by user
>>>>> >> >> >> and
>>>>> >> >> >> updated by others users. I have heard many people on the mailing
>>>>> >> >> >> tell
>>>>> >> >> >> that they use that. In fact, when we make code change that MSVC
>>>>> >> >> >> don't
>>>>> >> >> >> like(it is a more picky compiler then g++), users tell us about
>>>>> >> >> >> this
>>>>> >> >> >> and frequently provide us with the fixes.
>>>>> >> >> >
>>>>> >> >> >
>>>>> >> >> > Thanks for the detailed explanation.
>>>>> >> >> >
>>>>> >> >> > You convinced me it would be useful to add support for theano. It
>>>>> >> >> > should
>>>>> >> >> > be
>>>>> >> >> > available within a few days, unless some unforeseen complications
>>>>> >> >> > arise,
>>>>> >> >> >
>>>>> >> >> > thanks,
>>>>> >> >> > David
>>>>> >> >> >>
>>>>> >> >> >>
>>>>> >> >> >> Frédéric
>>>>> >> >> >>
>>>>> >> >> >>
>>>>> >> >> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau
>>>>> >> >> >> <[hidden email]>
>>>>> >> >> >> wrote:
>>>>> >> >> >> > Hi Frédéric,
>>>>> >> >> >> >
>>>>> >> >> >> > I agree having ancient theano is not good.
>>>>> >> >> >> >
>>>>> >> >> >> > I was considering moving it up into a supported package, but I
>>>>> >> >> >> > feared
>>>>> >> >> >> > that
>>>>> >> >> >> > it would not be so useful without GPU support. I
>>>>> >> >> >> > (unfortunately)
>>>>> >> >> >> > never
>>>>> >> >> >> > had
>>>>> >> >> >> > time to play with theano, and don't know how useful the package
>>>>> >> >> >> > is
>>>>> >> >> >> > without
>>>>> >> >> >> > CUDA (and I am a bit worried about support pressure when people
>>>>> >> >> >> > will
>>>>> >> >> >> > ask
>>>>> >> >> >> > 'GPU is not used, fix it').
>>>>> >> >> >> >
>>>>> >> >> >> > thanks,
>>>>> >> >> >> > David
>>>>> >> >> >> >
>>>>> >> >> >> >
>>>>> >> >> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien
>>>>> >> >> >> > <[hidden email]>
>>>>> >> >> >> > wrote:
>>>>> >> >> >> >>
>>>>> >> >> >> >> Hi,
>>>>> >> >> >> >>
>>>>> >> >> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know
>>>>> >> >> >> >> about
>>>>> >> >> >> >> those
>>>>> >> >> >> >> version of Theano:
>>>>> >> >> >> >>
>>>>> >> >> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1
>>>>> >> >> >> >> (no
>>>>> >> >> >> >>     subscription), 0.5.0rc1 (no subscription)
>>>>> >> >> >> >>
>>>>> >> >> >> >>
>>>>> >> >> >> >> It don't even know about Theano 0.5 that was release in
>>>>> >> >> >> >> February
>>>>> >> >> >> >> 2012.
>>>>> >> >> >> >>
>>>>> >> >> >> >> It would be great if you add version 0.6 as available. To my
>>>>> >> >> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac
>>>>> >> >> >> >> and
>>>>> >> >> >> >> linux.
>>>>> >> >> >> >>
>>>>> >> >> >> >> To my understanding Theano is included in the community
>>>>> >> >> >> >> section,
>>>>> >> >> >> >> so
>>>>> >> >> >> >> without support from you. So that should not add extra work
>>>>> >> >> >> >> from
>>>>> >> >> >> >> you.
>>>>> >> >> >> >> If I'm mistaken, please tell me. I would like to make Theano
>>>>> >> >> >> >> easily
>>>>> >> >> >> >> available and having only old version shown as being available
>>>>> >> >> >> >> isn't
>>>>> >> >> >> >> good.
>>>>> >> >> >> >>
>>>>> >> >> >> >> If you need modification in Theano for any reason, do not
>>>>> >> >> >> >> hesitate
>>>>> >> >> >> >> to
>>>>> >> >> >> >> tell
>>>>> >> >> >> >> me.
>>>>> >> >> >> >>
>>>>> >> >> >> >> thanks
>>>>> >> >> >> >>
>>>>> >> >> >> >> Frédéric Bastien
>>>>> >> >> >> >> _______________________________________________
>>>>> >> >> >> >> Enthought-Dev mailing list
>>>>> >> >> >> >> [hidden email]
>>>>> >> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>> >> >> >> >
>>>>> >> >> >> >
>>>>> >> >> >> >
>>>>> >> >> >> > _______________________________________________
>>>>> >> >> >> > Enthought-Dev mailing list
>>>>> >> >> >> > [hidden email]
>>>>> >> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>> >> >> >> >
>>>>> >> >> >> _______________________________________________
>>>>> >> >> >> Enthought-Dev mailing list
>>>>> >> >> >> [hidden email]
>>>>> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>> >> >> >
>>>>> >> >> >
>>>>> >> >> >
>>>>> >> >> > _______________________________________________
>>>>> >> >> > Enthought-Dev mailing list
>>>>> >> >> > [hidden email]
>>>>> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>> >> >> >
>>>>> >> >> _______________________________________________
>>>>> >> >> Enthought-Dev mailing list
>>>>> >> >> [hidden email]
>>>>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>> >> >
>>>>> >> >
>>>>> >> >
>>>>> >> > _______________________________________________
>>>>> >> > Enthought-Dev mailing list
>>>>> >> > [hidden email]
>>>>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>> >> >
>>>>> >> _______________________________________________
>>>>> >> Enthought-Dev mailing list
>>>>> >> [hidden email]
>>>>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>> >
>>>>> >
>>>>> >
>>>>> > _______________________________________________
>>>>> > Enthought-Dev mailing list
>>>>> > [hidden email]
>>>>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>> >
>>>>> _______________________________________________
>>>>> Enthought-Dev mailing list
>>>>> [hidden email]
>>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> Enthought-Dev mailing list
>>>> [hidden email]
>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
Hi David,

Did you had the time to work on this or should I put this on my todo?

thanks

Fred

On Fri, Dec 20, 2013 at 8:31 PM, Frédéric Bastien <[hidden email]> wrote:

> I mean, the PR is merged:)
>
> Fred
>
> On Fri, Dec 20, 2013 at 8:26 PM, Frédéric Bastien <[hidden email]> wrote:
>> The PR is fixed.
>>
>> Fred
>>
>> On Fri, Dec 20, 2013 at 4:06 PM, Frédéric Bastien <[hidden email]> wrote:
>>> Here is the PR with a fix:
>>>
>>> https://github.com/Theano/Theano/pull/1682
>>>
>>> I don't know when it will get merged.
>>>
>>> I'm rerunning our tests on a 32 bit VM just in case something else happened.
>>>
>>> thanks
>>>
>>> Fred
>>>
>>> On Fri, Dec 20, 2013 at 3:34 PM, Frédéric Bastien <[hidden email]> wrote:
>>>> I suppose this is with the development version of Theano? I added c
>>>> code to MaxAndArgmax and forgot to upcast to int64 the output of
>>>> PyArray_ArgMax(). This affect only 32 bit python version. Theano
>>>> always use int64 for the shape, but not NumPy. I only test on 32 bit
>>>> python once in a while or before release.
>>>>
>>>> Thanks for letting me know. A simple work around until I fix this in
>>>> the development version (I'll try to make the PR today), is t change
>>>> the method name of class MaxAndArgmax c_code() to _c_code() to disable
>>>> its c code.
>>>>
>>>>
>>>>
>>>> Fred
>>>>
>>>> On Fri, Dec 20, 2013 at 2:27 PM, David Cournapeau <[hidden email]> wrote:
>>>>>
>>>>>
>>>>>
>>>>> On Wed, Dec 18, 2013 at 7:48 PM, Frédéric Bastien <[hidden email]> wrote:
>>>>>>
>>>>>> On Wed, Dec 18, 2013 at 9:43 AM, David Cournapeau <[hidden email]>
>>>>>> wrote:
>>>>>> >
>>>>>> >
>>>>>> >
>>>>>> > On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]>
>>>>>> > wrote:
>>>>>> >>
>>>>>> >> Currently there isn't an easy way to test just a subset of Theano. The
>>>>>> >> only way is to specify directory/files/test name as a parameter to
>>>>>> >> nosetests.
>>>>>> >>
>>>>>> >> This have been asked a few times to have a partial but faster test
>>>>>> >> suite. Personally, I would prefer keeping the default like now and
>>>>>> >> having the user specify to run just a part of the test suite. Will
>>>>>> >> this fit your system?
>>>>>> >
>>>>>> >
>>>>>> > Sure.
>>>>>> >
>>>>>> >>
>>>>>> >>
>>>>>> >> I don't know the different way to mark tests as slow. Do you have a
>>>>>> >> recommended way to do this? I'll look at the way NumPy do this.
>>>>>> >
>>>>>> >
>>>>>> > nose makes that pretty easy: a slow decorator that add the slow attribue
>>>>>> > to
>>>>>> > the passed callable, then you can call nose -a '!slow'.
>>>>>> >
>>>>>> > That's how pandas does it, not sure about numpy.
>>>>>> >
>>>>>> >>
>>>>>> >> My daughter is sick and I'm in vacation for the next 2 weeks, so I
>>>>>> >> don't know when I can have this ready.
>>>>>> >
>>>>>> >
>>>>>> > Well, what about me doing it if the above strategy works for you :)
>>>>>>
>>>>>> Thanks. I will be happy that someone else do that. I'm good with the
>>>>>> strategy you told.
>>>>>>
>>>>>> The command
>>>>>>
>>>>>> theano-nose --time-profile
>>>>>>
>>>>>> would output 3 files timeprof_*. They will contain the timing of each
>>>>>> test. That can help you find the long running timming.
>>>>>>
>>>>>> In travis-ci config file, we split Theano tests into 4 parts.
>>>>>
>>>>>
>>>>> Ah, that's very useful information.
>>>>>
>>>>> I see errors like the following:
>>>>>
>>>>> ======================================================================
>>>>> ERROR: test_grad (theano.tensor.tests.test_basic.T_max_and_argmax)
>>>>> ----------------------------------------------------------------------
>>>>> Traceback (most recent call last):
>>>>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>>>>> line 2564, in test_grad
>>>>>     [data])
>>>>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>>>>> line 2540, in safe_verify_grad
>>>>>     utt.verify_grad(func, data, eps=eps)
>>>>>   File "/home/vagrant/src/tmp/Theano/theano/tests/unittest_tools.py", line
>>>>> 90, in verify_grad
>>>>>     T.verify_grad(op, pt, n_tests, rng, *args, **kwargs)
>>>>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1470, in
>>>>> verify_grad
>>>>>     eps, out_type)
>>>>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1233, in
>>>>> __init__
>>>>>     f_x = f(*[p.copy() for p in apt])
>>>>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>>>>> line 588, in __call__
>>>>>     self.fn.thunks[self.fn.position_of_error])
>>>>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>>>>> line 579, in __call__
>>>>>     outputs = self.fn()
>>>>> TypeError: ("expected type_num 9 (NPY_INT64) got 5\nApply node that caused
>>>>> the error: Elemwise{mul,no_inplace}(random_projection, argmax)\nInputs
>>>>> shapes: [(2,), (2,)]\nInputs strides: [(8,), (4,)]\nInputs types:
>>>>> [TensorType(float64, vector), TensorType(int64, vector)]\nUse the Theano
>>>>> flag 'exception_verbosity=high' for a debugprint of this apply node.",
>>>>> '\nThe error happened with the following inputs:', [array([[ 0.40087424,
>>>>> 0.68837329,  0.35302867],
>>>>>        [ 0.45571611,  0.90291591, -0.97459361]])], '\nThe value of eps is:',
>>>>> 1e-07, '\nThe out_type is:', None)
>>>>>
>>>>>
>>>>> Which I can't make much sense of I am afraid. This is on rh5 32 bits.
>>>>>
>>>>> David
>>>>>
>>>>>>
>>>>>> Two of
>>>>>> those part are theano/scan_module/ and the file
>>>>>> theano/tensor/tests/test_basic.py as they are long running. In the
>>>>>> test_basic.py file, there is many elemwise test for each scalar.
>>>>>> Skipping many of them would remove much coverage.
>>>>>>
>>>>>> Also, there is 2 type of timming, with a theano cache of compiled c
>>>>>> code "full" or "empty". This make a big difference in tests run time.
>>>>>> I suppose you want the timming with the empty cache? For your final
>>>>>> timming you can empty the cache with this command:
>>>>>>
>>>>>> theano-cache clear
>>>>>>
>>>>>> to be sure the timming is as you would like. Much fast would you like it
>>>>>> to be?
>>>>>>
>>>>>> thanks
>>>>>>
>>>>>> Fred
>>>>>>
>>>>>> >
>>>>>> > David
>>>>>> >>
>>>>>> >>
>>>>>> >> Fred
>>>>>> >>
>>>>>> >>
>>>>>> >> On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]>
>>>>>> >> wrote:
>>>>>> >> >
>>>>>> >> >
>>>>>> >> >
>>>>>> >> > On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]>
>>>>>> >> > wrote:
>>>>>> >> >>
>>>>>> >> >> Thanks for the good news!
>>>>>> >> >>
>>>>>> >> >> As always if you have any questions/comments don't hesitate to
>>>>>> >> >> contact
>>>>>> >> >> me.
>>>>>> >> >
>>>>>> >> >
>>>>>> >> > So I have a preliminary egg for Theano, but the test suite of ~ 20
>>>>>> >> > minutes
>>>>>> >> > is prohibitive. Is there an option for testing only a subset of
>>>>>> >> > Theano ?
>>>>>> >> >
>>>>>> >> > For packaging, I care mostly about ensuring we have not introduced
>>>>>> >> > bugs
>>>>>> >> > when
>>>>>> >> > packaging things incorrectly, and I like running those everytime we
>>>>>> >> > update
>>>>>> >> > both forward and reverse dependencies of a given package.
>>>>>> >> >
>>>>>> >> > thanks,
>>>>>> >> > David
>>>>>> >> >>
>>>>>> >> >>
>>>>>> >> >> Frédéric
>>>>>> >> >>
>>>>>> >> >> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau
>>>>>> >> >> <[hidden email]>
>>>>>> >> >> wrote:
>>>>>> >> >> > Hi Frédéric,
>>>>>> >> >> >
>>>>>> >> >> >
>>>>>> >> >> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>>>>>> >> >> > wrote:
>>>>>> >> >> >>
>>>>>> >> >> >> Hi,
>>>>>> >> >> >>
>>>>>> >> >> >> Theano is very useful without GPU. We have some paper that show
>>>>>> >> >> >> that
>>>>>> >> >> >> the CPU only back-end was faster then numpy in many cases and
>>>>>> >> >> >> faster
>>>>>> >> >> >> then numexpr too.
>>>>>> >> >> >>
>>>>>> >> >> >> Here is the most recent paper with benchmark but compare only to
>>>>>> >> >> >> other
>>>>>> >> >> >> highly optimized packages on machine learning algo:
>>>>>> >> >> >>
>>>>>> >> >> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>>>>>> >> >> >>
>>>>>> >> >> >> Here is the first paper with comparison again numpy/scipy and
>>>>>> >> >> >> numexpr:
>>>>>> >> >> >>
>>>>>> >> >> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf
>>>>>> >> >> >> look
>>>>>> >> >> >> at page 3 and 4.
>>>>>> >> >> >>
>>>>>> >> >> >>
>>>>>> >> >> >> This is just to show that Theano without the GPU is very useful.
>>>>>> >> >> >> In
>>>>>> >> >> >> fact we use it frequently here as we don't have enough GPUs and
>>>>>> >> >> >> have
>>>>>> >> >> >> access to a few CPU clusters.
>>>>>> >> >> >>
>>>>>> >> >> >>
>>>>>> >> >> >> I saw a few other people from Enthought play/contribute with/to
>>>>>> >> >> >> Theano
>>>>>> >> >> >> like Robert Kern. If you want you can talk about this with them.
>>>>>> >> >> >>
>>>>>> >> >> >> I don't know the detail of your supported status. Could you
>>>>>> >> >> >> support
>>>>>> >> >> >> it
>>>>>> >> >> >> on the CPU only? Or just refer to Theano installation for the
>>>>>> >> >> >> GPU?
>>>>>> >> >> >> On
>>>>>> >> >> >> linux and Mac, it isn't very complicated. But I agree that on
>>>>>> >> >> >> Windows,
>>>>>> >> >> >> it is more complicated as the user need to install the microsoft
>>>>>> >> >> >> compiler in addition to nvcc. Officially, we don't support GPU on
>>>>>> >> >> >> Windows, but we help answer questions about this on the mailing
>>>>>> >> >> >> when
>>>>>> >> >> >> people have problems, there is some instruction provided by user
>>>>>> >> >> >> and
>>>>>> >> >> >> updated by others users. I have heard many people on the mailing
>>>>>> >> >> >> tell
>>>>>> >> >> >> that they use that. In fact, when we make code change that MSVC
>>>>>> >> >> >> don't
>>>>>> >> >> >> like(it is a more picky compiler then g++), users tell us about
>>>>>> >> >> >> this
>>>>>> >> >> >> and frequently provide us with the fixes.
>>>>>> >> >> >
>>>>>> >> >> >
>>>>>> >> >> > Thanks for the detailed explanation.
>>>>>> >> >> >
>>>>>> >> >> > You convinced me it would be useful to add support for theano. It
>>>>>> >> >> > should
>>>>>> >> >> > be
>>>>>> >> >> > available within a few days, unless some unforeseen complications
>>>>>> >> >> > arise,
>>>>>> >> >> >
>>>>>> >> >> > thanks,
>>>>>> >> >> > David
>>>>>> >> >> >>
>>>>>> >> >> >>
>>>>>> >> >> >> Frédéric
>>>>>> >> >> >>
>>>>>> >> >> >>
>>>>>> >> >> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau
>>>>>> >> >> >> <[hidden email]>
>>>>>> >> >> >> wrote:
>>>>>> >> >> >> > Hi Frédéric,
>>>>>> >> >> >> >
>>>>>> >> >> >> > I agree having ancient theano is not good.
>>>>>> >> >> >> >
>>>>>> >> >> >> > I was considering moving it up into a supported package, but I
>>>>>> >> >> >> > feared
>>>>>> >> >> >> > that
>>>>>> >> >> >> > it would not be so useful without GPU support. I
>>>>>> >> >> >> > (unfortunately)
>>>>>> >> >> >> > never
>>>>>> >> >> >> > had
>>>>>> >> >> >> > time to play with theano, and don't know how useful the package
>>>>>> >> >> >> > is
>>>>>> >> >> >> > without
>>>>>> >> >> >> > CUDA (and I am a bit worried about support pressure when people
>>>>>> >> >> >> > will
>>>>>> >> >> >> > ask
>>>>>> >> >> >> > 'GPU is not used, fix it').
>>>>>> >> >> >> >
>>>>>> >> >> >> > thanks,
>>>>>> >> >> >> > David
>>>>>> >> >> >> >
>>>>>> >> >> >> >
>>>>>> >> >> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien
>>>>>> >> >> >> > <[hidden email]>
>>>>>> >> >> >> > wrote:
>>>>>> >> >> >> >>
>>>>>> >> >> >> >> Hi,
>>>>>> >> >> >> >>
>>>>>> >> >> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know
>>>>>> >> >> >> >> about
>>>>>> >> >> >> >> those
>>>>>> >> >> >> >> version of Theano:
>>>>>> >> >> >> >>
>>>>>> >> >> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1
>>>>>> >> >> >> >> (no
>>>>>> >> >> >> >>     subscription), 0.5.0rc1 (no subscription)
>>>>>> >> >> >> >>
>>>>>> >> >> >> >>
>>>>>> >> >> >> >> It don't even know about Theano 0.5 that was release in
>>>>>> >> >> >> >> February
>>>>>> >> >> >> >> 2012.
>>>>>> >> >> >> >>
>>>>>> >> >> >> >> It would be great if you add version 0.6 as available. To my
>>>>>> >> >> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac
>>>>>> >> >> >> >> and
>>>>>> >> >> >> >> linux.
>>>>>> >> >> >> >>
>>>>>> >> >> >> >> To my understanding Theano is included in the community
>>>>>> >> >> >> >> section,
>>>>>> >> >> >> >> so
>>>>>> >> >> >> >> without support from you. So that should not add extra work
>>>>>> >> >> >> >> from
>>>>>> >> >> >> >> you.
>>>>>> >> >> >> >> If I'm mistaken, please tell me. I would like to make Theano
>>>>>> >> >> >> >> easily
>>>>>> >> >> >> >> available and having only old version shown as being available
>>>>>> >> >> >> >> isn't
>>>>>> >> >> >> >> good.
>>>>>> >> >> >> >>
>>>>>> >> >> >> >> If you need modification in Theano for any reason, do not
>>>>>> >> >> >> >> hesitate
>>>>>> >> >> >> >> to
>>>>>> >> >> >> >> tell
>>>>>> >> >> >> >> me.
>>>>>> >> >> >> >>
>>>>>> >> >> >> >> thanks
>>>>>> >> >> >> >>
>>>>>> >> >> >> >> Frédéric Bastien
>>>>>> >> >> >> >> _______________________________________________
>>>>>> >> >> >> >> Enthought-Dev mailing list
>>>>>> >> >> >> >> [hidden email]
>>>>>> >> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>> >> >> >> >
>>>>>> >> >> >> >
>>>>>> >> >> >> >
>>>>>> >> >> >> > _______________________________________________
>>>>>> >> >> >> > Enthought-Dev mailing list
>>>>>> >> >> >> > [hidden email]
>>>>>> >> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>> >> >> >> >
>>>>>> >> >> >> _______________________________________________
>>>>>> >> >> >> Enthought-Dev mailing list
>>>>>> >> >> >> [hidden email]
>>>>>> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>> >> >> >
>>>>>> >> >> >
>>>>>> >> >> >
>>>>>> >> >> > _______________________________________________
>>>>>> >> >> > Enthought-Dev mailing list
>>>>>> >> >> > [hidden email]
>>>>>> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>> >> >> >
>>>>>> >> >> _______________________________________________
>>>>>> >> >> Enthought-Dev mailing list
>>>>>> >> >> [hidden email]
>>>>>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>> >> >
>>>>>> >> >
>>>>>> >> >
>>>>>> >> > _______________________________________________
>>>>>> >> > Enthought-Dev mailing list
>>>>>> >> > [hidden email]
>>>>>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>> >> >
>>>>>> >> _______________________________________________
>>>>>> >> Enthought-Dev mailing list
>>>>>> >> [hidden email]
>>>>>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>> >
>>>>>> >
>>>>>> >
>>>>>> > _______________________________________________
>>>>>> > Enthought-Dev mailing list
>>>>>> > [hidden email]
>>>>>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>> >
>>>>>> _______________________________________________
>>>>>> Enthought-Dev mailing list
>>>>>> [hidden email]
>>>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> Enthought-Dev mailing list
>>>>> [hidden email]
>>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev
Reply | Threaded
Open this post in threaded view
|

Re: Theano 0.6

Frédéric Bastien
Hi,

As I didn't got an answer, I suppose you are overloaded. Arnaud, can
you do this? I would like that you decorate Theano tests that are slow
with a "slow" flag to allow doing

nose -a "!slow"

But don't disable all tests for an op. You can use to know the run
time of each tests.

theano-nose --time-profile

After that, can you tell use the time Theano tests take with an empty
and full theano cache?

thanks

Fred


On Mon, Jan 6, 2014 at 12:30 PM, Frédéric Bastien <[hidden email]> wrote:

> Hi David,
>
> Did you had the time to work on this or should I put this on my todo?
>
> thanks
>
> Fred
>
> On Fri, Dec 20, 2013 at 8:31 PM, Frédéric Bastien <[hidden email]> wrote:
>> I mean, the PR is merged:)
>>
>> Fred
>>
>> On Fri, Dec 20, 2013 at 8:26 PM, Frédéric Bastien <[hidden email]> wrote:
>>> The PR is fixed.
>>>
>>> Fred
>>>
>>> On Fri, Dec 20, 2013 at 4:06 PM, Frédéric Bastien <[hidden email]> wrote:
>>>> Here is the PR with a fix:
>>>>
>>>> https://github.com/Theano/Theano/pull/1682
>>>>
>>>> I don't know when it will get merged.
>>>>
>>>> I'm rerunning our tests on a 32 bit VM just in case something else happened.
>>>>
>>>> thanks
>>>>
>>>> Fred
>>>>
>>>> On Fri, Dec 20, 2013 at 3:34 PM, Frédéric Bastien <[hidden email]> wrote:
>>>>> I suppose this is with the development version of Theano? I added c
>>>>> code to MaxAndArgmax and forgot to upcast to int64 the output of
>>>>> PyArray_ArgMax(). This affect only 32 bit python version. Theano
>>>>> always use int64 for the shape, but not NumPy. I only test on 32 bit
>>>>> python once in a while or before release.
>>>>>
>>>>> Thanks for letting me know. A simple work around until I fix this in
>>>>> the development version (I'll try to make the PR today), is t change
>>>>> the method name of class MaxAndArgmax c_code() to _c_code() to disable
>>>>> its c code.
>>>>>
>>>>>
>>>>>
>>>>> Fred
>>>>>
>>>>> On Fri, Dec 20, 2013 at 2:27 PM, David Cournapeau <[hidden email]> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Wed, Dec 18, 2013 at 7:48 PM, Frédéric Bastien <[hidden email]> wrote:
>>>>>>>
>>>>>>> On Wed, Dec 18, 2013 at 9:43 AM, David Cournapeau <[hidden email]>
>>>>>>> wrote:
>>>>>>> >
>>>>>>> >
>>>>>>> >
>>>>>>> > On Wed, Dec 18, 2013 at 12:31 PM, Frédéric Bastien <[hidden email]>
>>>>>>> > wrote:
>>>>>>> >>
>>>>>>> >> Currently there isn't an easy way to test just a subset of Theano. The
>>>>>>> >> only way is to specify directory/files/test name as a parameter to
>>>>>>> >> nosetests.
>>>>>>> >>
>>>>>>> >> This have been asked a few times to have a partial but faster test
>>>>>>> >> suite. Personally, I would prefer keeping the default like now and
>>>>>>> >> having the user specify to run just a part of the test suite. Will
>>>>>>> >> this fit your system?
>>>>>>> >
>>>>>>> >
>>>>>>> > Sure.
>>>>>>> >
>>>>>>> >>
>>>>>>> >>
>>>>>>> >> I don't know the different way to mark tests as slow. Do you have a
>>>>>>> >> recommended way to do this? I'll look at the way NumPy do this.
>>>>>>> >
>>>>>>> >
>>>>>>> > nose makes that pretty easy: a slow decorator that add the slow attribue
>>>>>>> > to
>>>>>>> > the passed callable, then you can call nose -a '!slow'.
>>>>>>> >
>>>>>>> > That's how pandas does it, not sure about numpy.
>>>>>>> >
>>>>>>> >>
>>>>>>> >> My daughter is sick and I'm in vacation for the next 2 weeks, so I
>>>>>>> >> don't know when I can have this ready.
>>>>>>> >
>>>>>>> >
>>>>>>> > Well, what about me doing it if the above strategy works for you :)
>>>>>>>
>>>>>>> Thanks. I will be happy that someone else do that. I'm good with the
>>>>>>> strategy you told.
>>>>>>>
>>>>>>> The command
>>>>>>>
>>>>>>> theano-nose --time-profile
>>>>>>>
>>>>>>> would output 3 files timeprof_*. They will contain the timing of each
>>>>>>> test. That can help you find the long running timming.
>>>>>>>
>>>>>>> In travis-ci config file, we split Theano tests into 4 parts.
>>>>>>
>>>>>>
>>>>>> Ah, that's very useful information.
>>>>>>
>>>>>> I see errors like the following:
>>>>>>
>>>>>> ======================================================================
>>>>>> ERROR: test_grad (theano.tensor.tests.test_basic.T_max_and_argmax)
>>>>>> ----------------------------------------------------------------------
>>>>>> Traceback (most recent call last):
>>>>>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>>>>>> line 2564, in test_grad
>>>>>>     [data])
>>>>>>   File "/home/vagrant/src/tmp/Theano/theano/tensor/tests/test_basic.py",
>>>>>> line 2540, in safe_verify_grad
>>>>>>     utt.verify_grad(func, data, eps=eps)
>>>>>>   File "/home/vagrant/src/tmp/Theano/theano/tests/unittest_tools.py", line
>>>>>> 90, in verify_grad
>>>>>>     T.verify_grad(op, pt, n_tests, rng, *args, **kwargs)
>>>>>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1470, in
>>>>>> verify_grad
>>>>>>     eps, out_type)
>>>>>>   File "/home/vagrant/src/tmp/Theano/theano/gradient.py", line 1233, in
>>>>>> __init__
>>>>>>     f_x = f(*[p.copy() for p in apt])
>>>>>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>>>>>> line 588, in __call__
>>>>>>     self.fn.thunks[self.fn.position_of_error])
>>>>>>   File "/home/vagrant/src/tmp/Theano/theano/compile/function_module.py",
>>>>>> line 579, in __call__
>>>>>>     outputs = self.fn()
>>>>>> TypeError: ("expected type_num 9 (NPY_INT64) got 5\nApply node that caused
>>>>>> the error: Elemwise{mul,no_inplace}(random_projection, argmax)\nInputs
>>>>>> shapes: [(2,), (2,)]\nInputs strides: [(8,), (4,)]\nInputs types:
>>>>>> [TensorType(float64, vector), TensorType(int64, vector)]\nUse the Theano
>>>>>> flag 'exception_verbosity=high' for a debugprint of this apply node.",
>>>>>> '\nThe error happened with the following inputs:', [array([[ 0.40087424,
>>>>>> 0.68837329,  0.35302867],
>>>>>>        [ 0.45571611,  0.90291591, -0.97459361]])], '\nThe value of eps is:',
>>>>>> 1e-07, '\nThe out_type is:', None)
>>>>>>
>>>>>>
>>>>>> Which I can't make much sense of I am afraid. This is on rh5 32 bits.
>>>>>>
>>>>>> David
>>>>>>
>>>>>>>
>>>>>>> Two of
>>>>>>> those part are theano/scan_module/ and the file
>>>>>>> theano/tensor/tests/test_basic.py as they are long running. In the
>>>>>>> test_basic.py file, there is many elemwise test for each scalar.
>>>>>>> Skipping many of them would remove much coverage.
>>>>>>>
>>>>>>> Also, there is 2 type of timming, with a theano cache of compiled c
>>>>>>> code "full" or "empty". This make a big difference in tests run time.
>>>>>>> I suppose you want the timming with the empty cache? For your final
>>>>>>> timming you can empty the cache with this command:
>>>>>>>
>>>>>>> theano-cache clear
>>>>>>>
>>>>>>> to be sure the timming is as you would like. Much fast would you like it
>>>>>>> to be?
>>>>>>>
>>>>>>> thanks
>>>>>>>
>>>>>>> Fred
>>>>>>>
>>>>>>> >
>>>>>>> > David
>>>>>>> >>
>>>>>>> >>
>>>>>>> >> Fred
>>>>>>> >>
>>>>>>> >>
>>>>>>> >> On Wed, Dec 18, 2013 at 5:42 AM, David Cournapeau <[hidden email]>
>>>>>>> >> wrote:
>>>>>>> >> >
>>>>>>> >> >
>>>>>>> >> >
>>>>>>> >> > On Thu, Dec 12, 2013 at 8:58 PM, Frédéric Bastien <[hidden email]>
>>>>>>> >> > wrote:
>>>>>>> >> >>
>>>>>>> >> >> Thanks for the good news!
>>>>>>> >> >>
>>>>>>> >> >> As always if you have any questions/comments don't hesitate to
>>>>>>> >> >> contact
>>>>>>> >> >> me.
>>>>>>> >> >
>>>>>>> >> >
>>>>>>> >> > So I have a preliminary egg for Theano, but the test suite of ~ 20
>>>>>>> >> > minutes
>>>>>>> >> > is prohibitive. Is there an option for testing only a subset of
>>>>>>> >> > Theano ?
>>>>>>> >> >
>>>>>>> >> > For packaging, I care mostly about ensuring we have not introduced
>>>>>>> >> > bugs
>>>>>>> >> > when
>>>>>>> >> > packaging things incorrectly, and I like running those everytime we
>>>>>>> >> > update
>>>>>>> >> > both forward and reverse dependencies of a given package.
>>>>>>> >> >
>>>>>>> >> > thanks,
>>>>>>> >> > David
>>>>>>> >> >>
>>>>>>> >> >>
>>>>>>> >> >> Frédéric
>>>>>>> >> >>
>>>>>>> >> >> On Thu, Dec 12, 2013 at 2:51 PM, David Cournapeau
>>>>>>> >> >> <[hidden email]>
>>>>>>> >> >> wrote:
>>>>>>> >> >> > Hi Frédéric,
>>>>>>> >> >> >
>>>>>>> >> >> >
>>>>>>> >> >> > On Fri, Dec 6, 2013 at 9:01 PM, Frédéric Bastien <[hidden email]>
>>>>>>> >> >> > wrote:
>>>>>>> >> >> >>
>>>>>>> >> >> >> Hi,
>>>>>>> >> >> >>
>>>>>>> >> >> >> Theano is very useful without GPU. We have some paper that show
>>>>>>> >> >> >> that
>>>>>>> >> >> >> the CPU only back-end was faster then numpy in many cases and
>>>>>>> >> >> >> faster
>>>>>>> >> >> >> then numexpr too.
>>>>>>> >> >> >>
>>>>>>> >> >> >> Here is the most recent paper with benchmark but compare only to
>>>>>>> >> >> >> other
>>>>>>> >> >> >> highly optimized packages on machine learning algo:
>>>>>>> >> >> >>
>>>>>>> >> >> >> http://arxiv.org/abs/1211.5590 look at page 7 and 9 for benchmark
>>>>>>> >> >> >>
>>>>>>> >> >> >> Here is the first paper with comparison again numpy/scipy and
>>>>>>> >> >> >> numexpr:
>>>>>>> >> >> >>
>>>>>>> >> >> >> http://www.iro.umontreal.ca/~lisa/pointeurs/theano_scipy2010.pdf
>>>>>>> >> >> >> look
>>>>>>> >> >> >> at page 3 and 4.
>>>>>>> >> >> >>
>>>>>>> >> >> >>
>>>>>>> >> >> >> This is just to show that Theano without the GPU is very useful.
>>>>>>> >> >> >> In
>>>>>>> >> >> >> fact we use it frequently here as we don't have enough GPUs and
>>>>>>> >> >> >> have
>>>>>>> >> >> >> access to a few CPU clusters.
>>>>>>> >> >> >>
>>>>>>> >> >> >>
>>>>>>> >> >> >> I saw a few other people from Enthought play/contribute with/to
>>>>>>> >> >> >> Theano
>>>>>>> >> >> >> like Robert Kern. If you want you can talk about this with them.
>>>>>>> >> >> >>
>>>>>>> >> >> >> I don't know the detail of your supported status. Could you
>>>>>>> >> >> >> support
>>>>>>> >> >> >> it
>>>>>>> >> >> >> on the CPU only? Or just refer to Theano installation for the
>>>>>>> >> >> >> GPU?
>>>>>>> >> >> >> On
>>>>>>> >> >> >> linux and Mac, it isn't very complicated. But I agree that on
>>>>>>> >> >> >> Windows,
>>>>>>> >> >> >> it is more complicated as the user need to install the microsoft
>>>>>>> >> >> >> compiler in addition to nvcc. Officially, we don't support GPU on
>>>>>>> >> >> >> Windows, but we help answer questions about this on the mailing
>>>>>>> >> >> >> when
>>>>>>> >> >> >> people have problems, there is some instruction provided by user
>>>>>>> >> >> >> and
>>>>>>> >> >> >> updated by others users. I have heard many people on the mailing
>>>>>>> >> >> >> tell
>>>>>>> >> >> >> that they use that. In fact, when we make code change that MSVC
>>>>>>> >> >> >> don't
>>>>>>> >> >> >> like(it is a more picky compiler then g++), users tell us about
>>>>>>> >> >> >> this
>>>>>>> >> >> >> and frequently provide us with the fixes.
>>>>>>> >> >> >
>>>>>>> >> >> >
>>>>>>> >> >> > Thanks for the detailed explanation.
>>>>>>> >> >> >
>>>>>>> >> >> > You convinced me it would be useful to add support for theano. It
>>>>>>> >> >> > should
>>>>>>> >> >> > be
>>>>>>> >> >> > available within a few days, unless some unforeseen complications
>>>>>>> >> >> > arise,
>>>>>>> >> >> >
>>>>>>> >> >> > thanks,
>>>>>>> >> >> > David
>>>>>>> >> >> >>
>>>>>>> >> >> >>
>>>>>>> >> >> >> Frédéric
>>>>>>> >> >> >>
>>>>>>> >> >> >>
>>>>>>> >> >> >> On Fri, Dec 6, 2013 at 3:32 PM, David Cournapeau
>>>>>>> >> >> >> <[hidden email]>
>>>>>>> >> >> >> wrote:
>>>>>>> >> >> >> > Hi Frédéric,
>>>>>>> >> >> >> >
>>>>>>> >> >> >> > I agree having ancient theano is not good.
>>>>>>> >> >> >> >
>>>>>>> >> >> >> > I was considering moving it up into a supported package, but I
>>>>>>> >> >> >> > feared
>>>>>>> >> >> >> > that
>>>>>>> >> >> >> > it would not be so useful without GPU support. I
>>>>>>> >> >> >> > (unfortunately)
>>>>>>> >> >> >> > never
>>>>>>> >> >> >> > had
>>>>>>> >> >> >> > time to play with theano, and don't know how useful the package
>>>>>>> >> >> >> > is
>>>>>>> >> >> >> > without
>>>>>>> >> >> >> > CUDA (and I am a bit worried about support pressure when people
>>>>>>> >> >> >> > will
>>>>>>> >> >> >> > ask
>>>>>>> >> >> >> > 'GPU is not used, fix it').
>>>>>>> >> >> >> >
>>>>>>> >> >> >> > thanks,
>>>>>>> >> >> >> > David
>>>>>>> >> >> >> >
>>>>>>> >> >> >> >
>>>>>>> >> >> >> > On Wed, Dec 4, 2013 at 3:18 PM, Frédéric Bastien
>>>>>>> >> >> >> > <[hidden email]>
>>>>>>> >> >> >> > wrote:
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >> Hi,
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >> We released Theano 0.6 yesterday. I saw in that enpkg know
>>>>>>> >> >> >> >> about
>>>>>>> >> >> >> >> those
>>>>>>> >> >> >> >> version of Theano:
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >>     0.3.1 (no subscription), 0.4.0 (no subscription), 0.4.1
>>>>>>> >> >> >> >> (no
>>>>>>> >> >> >> >>     subscription), 0.5.0rc1 (no subscription)
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >> It don't even know about Theano 0.5 that was release in
>>>>>>> >> >> >> >> February
>>>>>>> >> >> >> >> 2012.
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >> It would be great if you add version 0.6 as available. To my
>>>>>>> >> >> >> >> knowledge, Theano work well with EPD/Canopy on windows, mac
>>>>>>> >> >> >> >> and
>>>>>>> >> >> >> >> linux.
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >> To my understanding Theano is included in the community
>>>>>>> >> >> >> >> section,
>>>>>>> >> >> >> >> so
>>>>>>> >> >> >> >> without support from you. So that should not add extra work
>>>>>>> >> >> >> >> from
>>>>>>> >> >> >> >> you.
>>>>>>> >> >> >> >> If I'm mistaken, please tell me. I would like to make Theano
>>>>>>> >> >> >> >> easily
>>>>>>> >> >> >> >> available and having only old version shown as being available
>>>>>>> >> >> >> >> isn't
>>>>>>> >> >> >> >> good.
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >> If you need modification in Theano for any reason, do not
>>>>>>> >> >> >> >> hesitate
>>>>>>> >> >> >> >> to
>>>>>>> >> >> >> >> tell
>>>>>>> >> >> >> >> me.
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >> thanks
>>>>>>> >> >> >> >>
>>>>>>> >> >> >> >> Frédéric Bastien
>>>>>>> >> >> >> >> _______________________________________________
>>>>>>> >> >> >> >> Enthought-Dev mailing list
>>>>>>> >> >> >> >> [hidden email]
>>>>>>> >> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>> >> >> >> >
>>>>>>> >> >> >> >
>>>>>>> >> >> >> >
>>>>>>> >> >> >> > _______________________________________________
>>>>>>> >> >> >> > Enthought-Dev mailing list
>>>>>>> >> >> >> > [hidden email]
>>>>>>> >> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>> >> >> >> >
>>>>>>> >> >> >> _______________________________________________
>>>>>>> >> >> >> Enthought-Dev mailing list
>>>>>>> >> >> >> [hidden email]
>>>>>>> >> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>> >> >> >
>>>>>>> >> >> >
>>>>>>> >> >> >
>>>>>>> >> >> > _______________________________________________
>>>>>>> >> >> > Enthought-Dev mailing list
>>>>>>> >> >> > [hidden email]
>>>>>>> >> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>> >> >> >
>>>>>>> >> >> _______________________________________________
>>>>>>> >> >> Enthought-Dev mailing list
>>>>>>> >> >> [hidden email]
>>>>>>> >> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>> >> >
>>>>>>> >> >
>>>>>>> >> >
>>>>>>> >> > _______________________________________________
>>>>>>> >> > Enthought-Dev mailing list
>>>>>>> >> > [hidden email]
>>>>>>> >> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>> >> >
>>>>>>> >> _______________________________________________
>>>>>>> >> Enthought-Dev mailing list
>>>>>>> >> [hidden email]
>>>>>>> >> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>> >
>>>>>>> >
>>>>>>> >
>>>>>>> > _______________________________________________
>>>>>>> > Enthought-Dev mailing list
>>>>>>> > [hidden email]
>>>>>>> > https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>> >
>>>>>>> _______________________________________________
>>>>>>> Enthought-Dev mailing list
>>>>>>> [hidden email]
>>>>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> Enthought-Dev mailing list
>>>>>> [hidden email]
>>>>>> https://mail.enthought.com/mailman/listinfo/enthought-dev
>>>>>>
_______________________________________________
Enthought-Dev mailing list
[hidden email]
https://mail.enthought.com/mailman/listinfo/enthought-dev