@agibsonccc When do you think I will be able to try the software with the flatbuffers patch?
Thanks
Hi Alex:
I am on the last step trying to troubleshoot an issue with sonatype. Nothing is blocking me code wise. Should be soon.
@adonnini some snapshots are published. Not all of them yet though. The lower level c++ libraries still need to be built for some platforms but we’re about there!
@agibsonccc great! Thanks for the update
@agibsonccc Just checking in
@agibsonccc I hope you will have a good holiday weekend. Please let me know when I can take the next step. Thanks
@adonnini sorry it took a bit. The windows version had some issues I had to work out. I was assuming that’s what you were on.
This is VERY early. Use this url:
<repositories>
<repository>
<id>central-portal-snapshots</id>
<name>Central Portal Snapshots</name>
<url>https://central.sonatype.com/repository/maven-snapshots/</url>
<releases>
<enabled>false</enabled>
</releases>
<snapshots>
<enabled>true</enabled>
</snapshots>
</repository>
</repositories>
I also have to switch the namespaces. Use:
org.eclipse.deeplearning4j
for everything in place of org.nd4j,org.deeplearning4j,org.datavec.
The main reason for this was due to a HUGE migration.
@agibsonccc Thanks! I am on Debian 12. Can I still use the repository you sent me?
When I clikc on the link, the following message is displayed:
Product logo
Sonatype Nexus Repository
3.76.1-01
Repository image Repository maven-snapshots
This maven2 hosted repository is not directly browseable at this URL.
Please use the browse or HTML index views to inspect the contents of this repository.
If I click on browse it takes me to this page
clicking on or HTML index takes me to page-not-found
I must be doing something wrong sorry
@adonnini ah linux-x86_64 is also published. It should be ok. You can’t browse that one. I’m giving you a different snapshots repository.
Let me know if you have any issues.
Thanks @agibsonccc I could not find the link to the repository in your message above. Could you please resend it?
Thanks
@adonnini ah sorry the backticks didn’t save properly in my edits. Of note:
Please use:
/**
* Saves the SameDiff model, automatically determining if sharding is needed.
* Uses the “Metadata (FlatBuffers) + Appended Raw Data” format.
* Recommended method for most users.
*/
public static void save(@NonNull SameDiff sameDiff, @NonNull File baseFile, boolean saveUpdaterState, Map<String, String> metadata) throws IOException {
saveAutoShard(sameDiff, baseFile, saveUpdaterState, metadata);
}
Something like:
SameDiffSerializer.save(model,file,true,Collections.emptyMap())
SameDiff sd = SameDiffSerializer.load(modelFile,true);
note this is NOT backwards compatible. This is a completely new mechanism for saving and loading models. It uses a sharding mechanism to split the model by variable. It may not be optimized yet. Let me know what issues you run int o.
@agibsonccc Thanks. I will do as you suggest.
What’s the purpose of metadata
? Can it be null?
the basefile
parameter refers to the absolute path to file which will contain the model, right?
So, just making sure, I add this
<repositories>
<repository>
<id>central-portal-snapshots</id>
<name>Central Portal Snapshots</name>
<url>https://central.sonatype.com/repository/maven-snapshots/</url>
<releases>
<enabled>false</enabled>
</releases>
<snapshots>
<enabled>true</enabled>
</snapshots>
</repository>
</repositories>
to my pom file, and then use:
org.eclipse.deeplearning4j
for everything in place of org.nd4j,org.deeplearning4j,org.datavec.
in my import statements. Correct?
Sorry. I am a bit confused not being a regular maven user.
Thanks
@adonnini I updated the build again. I fixed some minor bugs with the model serialization and it should run much better now. You mostly have it. Pass in an empty map rather than null.
What I would recommend is copy your pom as a backup and just do a find and replace on the groupIds. Replace: org.nd4j,org.datavec,org.deeplearning4j with org.eclipse.deeplearning4j
and the version to 1.0.0-SNAPSHOT.
You should be able to just change a dl4j.version or similar property if you set your project up correctly. If you hardcoded 1.0.0-M2.1 just do a similar find and replace.
After that it was designed so the actual imports like org.nd4j.something are still the same beyond that.
The main thing that changed was the maven configuration NOT the java side imports.
I didn’t originally want to change the namespace immediately but was forced to due to the migration from the old sonatype infrastructure to the new.
I was going to do that anyways but wanted to give people a transition release with the new code so they don’t have to deal with more refactoring than they have to.
@agibsonccc Thanks! I am completing a phase of a project. As soon as I am done. I’ll implement the changes to use the new snapshot. Looking forward to doing this ASAP. I’ll keep you updated. Thanks
I don’t have
anywhere.
I do have in groupIds
org.deeplearning4j
,
nd4j-native
and
org.datavec
(in artifact)
Should I replace
org.deeplearning4j
with
org.eclipse.deeplearning4j
?
I assume I should leave all the import statements unchanged, right?
Thanks