[NLPL Task Force (A)] CNN slow-dow

[Andrei Kutuzov]

I can confirm that the problem is solved for me.

This module ('nlpl-pytorch/1.1.0/3.7') runs convolutions at more or less
the same speed as 'nlpl-in5550' (with old PyTorch 0.4), and MUCH faster
than 'nlpl-pytorch/1.0.0/3.7'.

Thanks!

On 5/10/19 10:25 AM, Stephan Oepen wrote:
> dear PyTorch users,
> 
> i just installed PyTorch 1.1.0 (now with NumPy 1.16.3) on Abel, and
> for all i can tell this upgrade resolves the exceedingly slow
> convolutions.  could i encourage you to test-drive this new module a
> little?  it is called 'nlpl-pytorch/1.1.0/3.7'.
> 
> with thanks in advance, oe
> 
> 
> On Thu, Mar 28, 2019 at 2:38 PM Andrei Kutuzov <andreku at ifi.uio.no> wrote:
>>
>> Hi,
>>
>> I tested our solution code for Obligatories 1 and 2 with Pytorch 0.4.1,
>> and it seems to work.
>> Thus, the best we can do right now I think is to downgrade PyTorch in
>> the IN5550 environment to this version.
>> This should be announced on the course web page and on Piazza, of course.
>>
>> On 3/26/19 6:36 PM, Vinit Ravishankar wrote:
>>> FWIW, I did try setting up a local environment with numpy==1.16.1 and torch==1.0.0, but it runs fine on my laptop. It’s plausible that there might be a version mismatch with some C library or other, especially because (I think) most optimisation-related code is written in C, right?
>>>
>>> – Vinit
>>>
>>>> On 26 Mar 2019, at 17:41, Andrei Kutuzov <andreku at ifi.uio.no> wrote:
>>>>
>>>> Note also that the fast nlpl-pytorch/0.4.1 environment has Numpy 1.15.1,
>>>> while both slow environments have Numpy 1.16.1.
>>>> Can be of importance as well.
>>>>
>>>> On 3/26/19 4:25 PM, Andrei Kutuzov wrote:
>>>>> I checked, and it doesn't depend on the Python version: Pytorch 1.0 runs
>>>>> equally slow on both Python 3.5 and Python 3.7
>>>>> Pytorch 0.4.1 runs on Python 3.5, and it is fast.
>>>>>
>>>>>
>>>>>
>>>>> On 3/26/19 4:21 PM, Vinit Ravishankar wrote:
>>>>>> Here’s a super tiny bit of code that just runs the same convolutional layer on the same input 10ish times.
>>>>>>
>>>>>> – Vinit
>>>>>>
>>>>>>> On 26 Mar 2019, at 15:58, Stephan Oepen <oe at ifi.uio.no> wrote:
>>>>>>>
>>>>>>> do you guys have a ready-to-run code snippet and some instructions on
>>>>>>> how to replicate the problem of unduly slow CNN performance in our
>>>>>>> newest PyTorch installation?  also, could you quickly summarize the
>>>>>>> results of benchmarking you have done so far?
>>>>>>>
>>>>>>> NLPL has three PyTorch modules:
>>>>>>>
>>>>>>> nlpl-pytorch/0.4.1
>>>>>>> nlpl-pytorch/1.0.0/3.5
>>>>>>> nlpl-pytorch/1.0.0/3.7
>>>>>>>
>>>>>>> it actually could be very nice to work towards a set of standard
>>>>>>> benchmarks, e.g. one each for a simple deep MLP, CNN, and RNN?
>>>>>>>
>>>>>>> cheers, oe
>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> Andrei
>>>> PhD Candidate at Language Technology Group (LTG)
>>>> University of Oslo
>>>
>>
>>
>> --
>> Andrei
>> PhD Candidate at Language Technology Group (LTG)
>> University of Oslo


-- 
Andrei
PhD Candidate at Language Technology Group (LTG)
University of Oslo