Hi,
Thank you now I really understand what's happening. I started to take  
a look at the data files in hpxml and it seems that they only contain  
the inputs, not the optimal policies. So I assume that you find the  
optimal training policies on training data in the  
binary_regression_model.hpp file. Is that correct? Since I want to  
train my data on python, I think I will need to generate a file of the  
optimal policies for the examples (output.dat). And then I could train  
in python using (input.dat and output.dat) what do you think?

Gabriel


Zahra Khatami <[email protected]> a écrit :

> Hi Gabriel,
>
> About your cat example, you understand machine learning concepts correctly
> and you are in a right direction!
> About your question, when collecting training data, we know which policy is
> the optimum one for the that hpx loop for example. But how we know that?
> Actually we run that test loop twice, first using sequential policy and in
> the second run using parallel policy, and by comparing their execution time
> we can find out that which policy worked better for that specific loop. So
> we collected training data like this!
> For testing, we use those training data to predict the policy of completely
> new hpx loop. Imagine that after executing our learning model, it’s policy
> was chosen to be parallel for that loop. To be sure that if our model was
> right, we run that new loop again by setting it’s policy  to be sequential.
> That’s how we determine the accuracy of our model. If you look at the
> comparison results in our paper, you will find out that we compare the
> results of our learning model with the one that their policies ( our chunk
> sizes or prefetching distances) was set manually with different candidates.
>
> Thanks,
> Zahra
>
> On Tue, Feb 27, 2018 at 7:39 PM Gabriel Laberge <[email protected]>
> wrote:
>
>> Hi,
>> I think the last time I asked you this question I wasn't very clear so
>>   will clarify. This could be due to the fact that i'm new to machine
>> learning.
>>
>>  From my understanding, if you want to train a binary logistic
>> regression, you need a matrix Xdata which represent training features
>> and you need Ydata which is a vector representing the expected output
>> (0 or 1) for each of the training examples. For example, if I want to
>> train a regression to identify cat pictures, I need Xdata which
>> represents a set of 'm' cat pictures and Ydata which is a vector of 0
>> and 1 if the corresponding picture is a cat.
>>
>> I assume the same logic applies with a logistic regression used to
>> find the execution policy (seq or par)
>> Xdata represents the inputs features for the examples and Ydata
>> represent the expected output (seq or par) for each of the training
>> examples.
>>
>>  From our previous discussion, It seemed like you told me that the
>> data was only constituted of input features (Xdata) which would mean
>> that we don't know the optimal policy on our data set. If this is the
>> case, I simply don't understand how you can train a regression.
>>
>> My analogy would be that you can't train a regression to recognize cat
>> pictures if you don't tell it which one are carts and which aren't.
>>
>> Could you tell me if I'm wrong?
>>
>> If I'm right then I wonder how do you find the optimal policy for each
>> of the training examples.
>>
>> Thank you very much.
>> Gabriel.
>>
>>
>> --
> Best Regards, Zahra Khatami | PhD Student Center for Computation &
> Technology (CCT) School of Electrical Engineering & Computer Science
> Louisiana State University 2027 Digital Media Center (DMC) Baton Rouge, LA
> 70803



_______________________________________________
hpx-users mailing list
[email protected]
https://mail.cct.lsu.edu/mailman/listinfo/hpx-users

Reply via email to