Github user asfgit closed the pull request at:
https://github.com/apache/opennlp/pull/136
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is ena
Yes, open an issue for the name samples, that should be fixed.
Jörn
On Mon, Mar 6, 2017 at 2:17 PM, Damiano Porta
wrote:
> I have to redesign it, reading the wiki you gave me i have noticed that i
> should not create two partitions (one for trainiing and one for testing).
> It avoids overfittin
I have to redesign it, reading the wiki you gave me i have noticed that i
should not create two partitions (one for trainiing and one for testing).
It avoids overfitting, so i will pass all the data!
Thanks Jorn!
P.S. Did you read my previous email about the bug in namesamples? Should i
open an is
Oh I see. Thanks!
Basically i have 30k sentences i apply the labels with a script and then i
pass 0-15k to train the model (to build the .bin) and 15k-30k to evaluate
it.
I am trying to build the model with 300 iterations again.
2017-03-06 13:31 GMT+01:00 Joern Kottmann :
> You should understan
You should understand how it works, have a look at this wikipedia article,
the picture on the right side explains it quite nicely.
https://en.wikipedia.org/wiki/Cross-validation_(statistics)
The idea is to split the data into n partitions and then use n-1 for
training and 1 for testing, this is re
Unfortunately not, 100 iterations ~ 30 minutes 300 iterations > 2 days and
it is still running... i will block it
i still do not understand what number should i set as *folds*. Ok i will
set a number > 1 but, should i have to pay more attention to this
parameter? if i set 8 or 10 does it matter an
test.evaluate(samples, 1), here the second parameter is the number of
folds, usually you use 10 or a number larger than 1.
The amount of times you need for training with perceptron is linear to the
iterations, if you use 300 instead of 100 it should take three times as
long.
Jörn
On Mon, Mar 6,
Jorn,
I am training and testing the model via api. If it is not a training
problem. How is that possible that the evaluation is taking 2 days (and
still running) to evaluate the model? As i told you with 100 iterations i
can get the model and the test in ~30 minutes.
I only have a doubt about eval
That is correct, we would be happy to merge a PR to change that.
Jörn
On Mon, Mar 6, 2017 at 10:49 AM, Damiano Porta
wrote:
> Jorn, i think it is really important. For the moment we should allow more
> threads for perceptron training. If remember correctly it is only allowed
> for MAXENT classi
Hello,
the model is only available after the training finished, hard to guess what
you are doing.
Do you use the command line? Which command?
Jörn
On Mon, Mar 6, 2017 at 10:29 AM, Damiano Porta
wrote:
> Hello Jorn,
> I tried with 300 iterations and it takes forever, reducing that number to
>
Jorn, i think it is really important. For the moment we should allow more
threads for perceptron training. If remember correctly it is only allowed
for MAXENT classifier, right ?
2017-03-06 10:17 GMT+01:00 Joern Kottmann :
> Hello,
>
> no, we don't support CUDA. At some point we probably add supp
Hello,
no, we don't support CUDA. At some point we probably add support for one of
the deep learning packages and those usually use CUDA.
Jörn
On Sat, Mar 4, 2017 at 5:17 PM, Damiano Porta
wrote:
> Hello everybody,
>
> does OpenNLP support CUDA parallel computing?
>
> Damiano
>
Hello Jorn,
I tried with 300 iterations and it takes forever, reducing that number to
100 i can finally get the model in half an hour.
The problem with 300 iterations is that i can see the model (.bin) in half
an hour too but the computations are still running. So i do not really
understand what i
Hello,
this looks like output from the cross validator.
Jörn
On Sun, Mar 5, 2017 at 11:34 AM, Damiano Porta
wrote:
> Hello,
>
> I am training a NER model with perceptron classifier (using OpenNLP 1.7.0)
>
> the output of the training is:
>
> Indexing events using cutoff of 0
>
> Computing even
14 matches
Mail list logo
