Ok, let's have both, since both will be useful with different datasets.

On Mon, Jun 1, 2015 at 12:21 AM, Maheshakya Wijewardena <[email protected]
> wrote:

> I should agree with Upul about adding both if possible. Mini-batch adds
> the question of determining the right size for batch size, but finding the
> right batch size may greatly improve our results as well as time for
> convergence. But still, it can depend heavily on the dataset.
>
> Have you tried with different datasets? Different in terms of size as well
> as other statistical properties of features(such as standard deviation,
> skewness, etc.)?
>
> On Sun, May 31, 2015 at 10:28 PM, Nirmal Fernando <[email protected]> wrote:
>
>> yes.. but from the simple test I did, I felt L-BFGS is faster. Will
>> confirm anyway.
>>
>> On Sun, May 31, 2015 at 10:13 PM, Upul Bandara <[email protected]> wrote:
>>
>>> Actually, I'm thinking in terms of training time, even for large data
>>> sets prediction accuracy of L-BFGS will outperform SGD. But its training
>>> time would be considerably bigger than the training time of SGD.
>>> On the other hand, SGD model gives a decent prediction accuracy in
>>> relatively short period of training time.
>>>
>>>
>>> On Sun, May 31, 2015 at 9:52 PM, Nirmal Fernando <[email protected]>
>>> wrote:
>>>
>>>> Thanks Upul. So, are you thinking along the lines of performance? Sure,
>>>> I'll run a test.
>>>>
>>>> On Sun, May 31, 2015 at 9:50 PM, Upul Bandara <[email protected]> wrote:
>>>>
>>>>> If it is possible, I would like to have both.
>>>>>
>>>>> L-BFGS converges faster than SGD. But it goes through the entire data
>>>>> set before moving from one iteration to the next.
>>>>> Whereas, SGD uses a minit-batch of the training data set for
>>>>> calculating and updating its gradient.
>>>>> Hence, for large data sets SGD is more practical than L-BFGS.
>>>>>
>>>>> I think we can test this scenario by running these two algorithms
>>>>> against a large data set (~ 1GB)
>>>>>
>>>>> Thanks,
>>>>> Upul
>>>>>
>>>>> On Sun, May 31, 2015 at 8:02 PM, Nirmal Fernando <[email protected]>
>>>>> wrote:
>>>>>
>>>>>> One other benefit of switching is, this API supports multi-class
>>>>>> classification too. I've tested this API with Iris dataset.
>>>>>>
>>>>>> On Sun, May 31, 2015 at 7:33 PM, Nirmal Fernando <[email protected]>
>>>>>> wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> Currently in ML, we use mini-batch gradient descent algorithm when
>>>>>>> running logistic regression. But Spark-mllib recommends L-BFGS over
>>>>>>> mini-batch gradient descent for faster convergence [1].
>>>>>>>
>>>>>>> I tested both the implementation with the same dataset and gained an
>>>>>>> improved accuracy in L-BFGS (80% vs 67% for SGD).
>>>>>>>
>>>>>>> Shall we switch?
>>>>>>>
>>>>>>> [1]
>>>>>>> https://spark.apache.org/docs/latest/mllib-linear-methods.html#logistic-regression
>>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>>
>>>>>>> Thanks & regards,
>>>>>>> Nirmal
>>>>>>>
>>>>>>> Associate Technical Lead - Data Technologies Team, WSO2 Inc.
>>>>>>> Mobile: +94715779733
>>>>>>> Blog: http://nirmalfdo.blogspot.com/
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>>
>>>>>> Thanks & regards,
>>>>>> Nirmal
>>>>>>
>>>>>> Associate Technical Lead - Data Technologies Team, WSO2 Inc.
>>>>>> Mobile: +94715779733
>>>>>> Blog: http://nirmalfdo.blogspot.com/
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> Upul Bandara,
>>>>> Associate Technical Lead, WSO2, Inc.,
>>>>> Mob: +94 715 468 345.
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>>
>>>> Thanks & regards,
>>>> Nirmal
>>>>
>>>> Associate Technical Lead - Data Technologies Team, WSO2 Inc.
>>>> Mobile: +94715779733
>>>> Blog: http://nirmalfdo.blogspot.com/
>>>>
>>>>
>>>>
>>>
>>>
>>> --
>>> Upul Bandara,
>>> Associate Technical Lead, WSO2, Inc.,
>>> Mob: +94 715 468 345.
>>>
>>
>>
>>
>> --
>>
>> Thanks & regards,
>> Nirmal
>>
>> Associate Technical Lead - Data Technologies Team, WSO2 Inc.
>> Mobile: +94715779733
>> Blog: http://nirmalfdo.blogspot.com/
>>
>>
>>
>
>
> --
> Pruthuvi Maheshakya Wijewardena
> Software Engineer
> WSO2 Lanka (Pvt) Ltd
> Email: [email protected]
> Mobile: +94711228855
>
>
>


-- 

Thanks & regards,
Nirmal

Associate Technical Lead - Data Technologies Team, WSO2 Inc.
Mobile: +94715779733
Blog: http://nirmalfdo.blogspot.com/
_______________________________________________
Dev mailing list
[email protected]
http://wso2.org/cgi-bin/mailman/listinfo/dev

Reply via email to