eclipse/deeplearning4j

A general forum for machine learning and deep learining questions, focused on DL4J/ND4J

DarkWingMcQuack

@DarkWingMcQuack

@izzat1996 would you go higher or lower?

could this be a vanishing gradient problem?

Eduardo Gonzalez

@wmeddie

I’d double check the data.

DarkWingMcQuack

@DarkWingMcQuack

@wmeddie it does happen always at the same data interval. if i restart the training it happens at another timestamp

      .gradientNormalization(GradientNormalization.ClipElementWiseAbsoluteValue)
      .gradientNormalizationThreshold(1.0)

made it a bit better

izzat1996

@izzat1996

i will go a bit higher because before this when i try to train my model the same problems happens to my model

DarkWingMcQuack

@DarkWingMcQuack

@izzat1996 thanks i will try that

i just discovered this

i guess that this is not normal, too. This is my LSTM layer

DarkWingMcQuack

@DarkWingMcQuack

Does noone has an idea what is going on here? i tried different batch sizes and the problem remains. I checked the data but i was not able to find any NaN or other uncommon things. Maybe some outliners in the data, but could that lead to such a behaviour?

Eduardo Gonzalez

@wmeddie

Yeah.

One before I was doing e-mail processing, a few of the e-mails had huge sections of base64 encoded files in-line. Caused huge spikes in the loss.

DarkWingMcQuack

@DarkWingMcQuack

so i guess i need to do some outlier removal :/

Giri

@Giribushan

cudacudacudacudacuda

Adam Gibson

@agibsonccc

@nicoladaoud hard to tell what it is, you likely have a bad set of deps. Maybe use nd4j-native-platform instead?

ADAMYA850

@ADAMYA850

I AM ADAMYA

I AM LERNING PYTHON

CAN ANYBODY OF YOU HELP mE

cawthorne

@cawthorne

Hi.

When I try to install nd4j on an arm64 server that doesn't have glibc 2.7 I get the error:

Caused by: java.lang.UnsatisfiedLinkError: /home/*/.javacpp/cache/benchmarks.jar/org/nd4j/nativeblas/linux-arm64/libjnind4jcpu.so: /lib64/libm.so.6: versionGLIBC_2.27' not found (required by /home/*/.javacpp/cache/benchmarks.jar/org/nd4j/nativeblas/linux-arm64/libnd4jcpu.so`

However it does work when installed on a server with that glibc.

Is there anyway to support other glibc environments? Or is everyone who doesn't have glibc 2.7 considered unsupported?

Thanks,
Greg

Eilons

@Eilons

Hi All.
I have a mac and use maven.
In my pom I added:

<dependency>
<groupId>org.deeplearning4j</groupId>
<artifactId>deeplearning4j-core</artifactId>
<version>1.0.0-beta7</version>
</dependency>
<dependency>
<groupId>org.nd4j</groupId>
<artifactId>nd4j-native</artifactId>
<version>1.0.0-beta7</version>
</dependency>
<dependency>
<groupId>org.deeplearning4j</groupId>
<artifactId>deeplearning4j-nlp</artifactId>
<version>1.0.0-beta7</version>
</dependency>
<dependency>
<groupId>org.deeplearning4j</groupId>
<artifactId>deeplearning4j-modelimport</artifactId>
<version>1.0.0-beta7</version>
</dependency>

I try to call the Word2Vec class but I get that the class is not identify (attached).
Can anyone help?

Note that when I use my windows computer, everything works fine

Adam Gibson

@agibsonccc

Use nd4j-native-platform and ensure intelij/eclipse updates its cache properly

Eilons

@Eilons

Wow! perfect! Thanks!

cawthorne

@cawthorne

bump for my post :)

Eduardo Gonzalez

@wmeddie

Are you using a really new OS?

cawthorne

@cawthorne

ubuntu 16

I have 2.23 installed and it needs 2.7

what would be the best way to resolve this do you think?

Eduardo Gonzalez

@wmeddie

Ah, the opposite problem.

for ubuntu… not sure.

cawthorne

@cawthorne

What are the general steps for an os you are familiar with?

Eduardo Gonzalez

@wmeddie

In Red-hot/Centos there’s devtoolset that you can possibly install.

And you run it after you’ve activated a recent enough devtoolset.

cawthorne

@cawthorne

Ah cool. Can it also do Downgrading if needed?

Eduardo Gonzalez

@wmeddie

otherwise, you can create a chroot with a newer version.

You have to “activate it” every time you want to use it.

cawthorne

@cawthorne

Ah okay good to know thanks

Eduardo Gonzalez

@wmeddie

same goes for the chroot.

cawthorne

@cawthorne

Ah I thought you were talking about chroot. Okay thanks.

@nimishatandon

Am looking for some guidance s that my models are reproducible

When using dl4j beta-7. When using beta-2 and training with the exact same test and train data I was getting consistent models , but it's not the case with beta-7 .

Eduardo Gonzalez

@wmeddie

If you can share something reproducible in a bug report we will look into it. Properly seeding all of the libraries we integrate with is non-trivial.

@nimishatandon

It's a vanilla LSTM model that I am training . Would you still like me to share the configuration and the stopping criteria of my models? I read on your website that setting the seed in the configuration object would help make the models reproducible. If there are other such recommendations I could follow that would be really helpful.

Eduardo Gonzalez

@wmeddie

If it was reproducible in beta2 and no longer the case (and nothing about that mentioned in the release notes) then its a bug.

cawthorne

@cawthorne

Hey so I noticed you guys don't have some of the blas routines implemented in nd4j, and from building from source I found these can be implemented trivially. For example strmv triggers a java.lang.UnsupportedOperationException error but if I just change 2 lines in the code I get it to work as it has a openblas version available.

Would you guys accept a PR from me filling in all the Blas and LAPACK routines? Or is this limited coverage somehow intentional?

openblas has very good coverage

Eduardo Gonzalez

@wmeddie

I can’t see why it’d be intentional.. but I’m not 100% sure.

Where communities thrive

People

Repo info

Activity