Hi Fred,
I just followed your suggestion and hard coded the changes in my Theano 
package, and ran multiple experiments with the same settings. What I 
observed is that, after applying this patch, the non-determinism reduces to 
only 2 cases but does not completely disappear.  In other words, before 
applying the changes, each experiment would end up with a different cost, 
while now there are only 2 points that each of the experiments end up. So, 
the behavior is more deterministic, but not 100%. 
Thanks to Ozan Çağlayan <https://github.com/ozancaglayan>, I found that to 
solve the issue completely (at least for my case), I need to have a recent 
version of Theano in which the following changes are applied (in 
theano/scan_module/scan_op.py):
scan/scan_op: Convert known_grads to OrderedDict 
<https://github.com/Theano/Theano/commit/8769382ff661aab15dda474a4c74456037f73cc6>
One can also manually change  theano/scan_module/scan_op.py according to 
what is described in the above link.

I still have not performed any real experiment (with large data sets and 
large number of iterations) using this modification; but it sounds 
promising. At least in 18 runs (on my toy example) I got exactly the same 
cost after fixed number updates, while before they would differ.
So, while my heavier experiments are running, I would like to start working 
on introducing the *deterministic* flag to theano, in order to avoid hard 
coding the changes, and also have the option to run different experiments 
with different determinism behavior.
May I ask you to point me to the portion of Theano code in which I can 
introduce this flag?

Thanks,
Amin


On Monday, February 1, 2016 at 3:59:43 PM UTC+1, nouiz wrote:
>
> Go in the file theano/sandbox/cuda/opt.py. Search for 
> GpuAdvancedIncSubtensor1_dev20 and make sure that it is 
> GpuAdvancedIncSubtensor1 that is used instead. We wanted to make a Theano 
> flag for this, do you want to make it?
>
> On Sun, Jan 31, 2016 at 11:33 AM, Zhenyang Li <[email protected] 
> <javascript:>> wrote:
>
>> Hi Fred,
>>
>> Yes, please, I want to make the result more consistent across different 
>> machines.
>>
>> Thank you,
>> Zhenyang
>>
>> On Thursday, January 28, 2016 at 8:34:14 PM UTC+1, nouiz wrote:
>>>
>>> About cudnn, you can use Theano flag to have it use deterministic 
>>> algorithms.
>>>
>>> Theano have a few places where we use the atomic add operation on the 
>>> GPU. This can cause in ordered addition. As this is done on floats this can 
>>> lead to d different result. We do this in the grad of advanced subtensor. 
>>> We have an older version that is deterministic but that is slower. There is 
>>> no flag to use it, but of you want to try out, I can tell you which change 
>>> is needed on Theano.
>>>
>>> Fred
>>> Le 27 janv. 2016 04:52, "Zhenyang Li" <[email protected]> a écrit :
>>>
>>>> Hi Pascal,
>>>>
>>>> Thank you very much, in the end I solved it by removing cudnn lib, then 
>>>> it's consistent on a same machine again.
>>>>
>>>> Another problem I have now is that, when I run a same RNN (standard 
>>>> LSTM) model, on same type of GPUs (Titan X) on two machines (basically two 
>>>> nodes on a cluster, so almost same platform).
>>>> Setting up proper gradient clipping, like what Keras do 
>>>> <https://github.com/fchollet/keras/blob/master/keras/optimizers.py#L48>, 
>>>> I got the exactly same results on the two machines, but without gradient 
>>>> clipping, I also observed that similar situation above, i.e.
>>>> quite similar mini-batch cost in the beginning, but the difference 
>>>> became larger and larger, is it expected?
>>>>
>>>> Best,
>>>> Zhenyang
>>>>
>>>>
>>>> On Tuesday, January 26, 2016 at 1:18:44 AM UTC+1, Pascal Lamblin wrote:
>>>>>
>>>>> This is possible, depending on what your model is. 
>>>>> More information at https://github.com/Theano/Theano/issues/3029 
>>>>>
>>>>> On Sun, Jan 24, 2016, Zhenyang Li wrote: 
>>>>> > Hi folks, 
>>>>> > 
>>>>> > I ran my theano code on a same GPU multiple times and found that for 
>>>>> > different runs, I got different results (i mean mini-batch cost 
>>>>> here), 
>>>>> > it's always the same for the beginning ~15 (param updating) rounds, 
>>>>> then 
>>>>> > got 10e-5 difference and became larger and larger, in the end, I got 
>>>>> very 
>>>>> > different results on a evaluation set. 
>>>>> > 
>>>>> > However, I also tried the same code on CPU multiple times, and I got 
>>>>> > consistently same results. 
>>>>> > 
>>>>> > What would be the issue, since I could not reproduce same results if 
>>>>> > running on GPU? And my theano GPU config is: 
>>>>> > 
>>>>> > floatX = float32 
>>>>> > 
>>>>> > device = gpu0 
>>>>> > 
>>>>> > mode = FAST_RUN 
>>>>> > 
>>>>> > optimizer = fast_run 
>>>>> > 
>>>>> > warn_float64 = warn 
>>>>> > 
>>>>> > Any help will be appreciated! 
>>>>> > 
>>>>> > 
>>>>> > Best, 
>>>>> > Zhenyang 
>>>>> > 
>>>>> > -- 
>>>>> > 
>>>>> > --- 
>>>>> > You received this message because you are subscribed to the Google 
>>>>> Groups "theano-users" group. 
>>>>> > To unsubscribe from this group and stop receiving emails from it, 
>>>>> send an email to [email protected]. 
>>>>> > For more options, visit https://groups.google.com/d/optout. 
>>>>>
>>>>>
>>>>> -- 
>>>>> Pascal 
>>>>>
>>>> -- 
>>>>
>>>> --- 
>>>> You received this message because you are subscribed to the Google 
>>>> Groups "theano-users" group.
>>>> To unsubscribe from this group and stop receiving emails from it, send 
>>>> an email to [email protected].
>>>> For more options, visit https://groups.google.com/d/optout.
>>>>
>>> -- 
>>
>> --- 
>> You received this message because you are subscribed to the Google Groups 
>> "theano-users" group.
>> To unsubscribe from this group and stop receiving emails from it, send an 
>> email to [email protected] <javascript:>.
>> For more options, visit https://groups.google.com/d/optout.
>>
>
>

-- 

--- 
You received this message because you are subscribed to the Google Groups 
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to