MXNet-Scala is a Scala binding for the MXNet Deep Learning framework. It currently supports Training and Inference
| Module | Package Generated | Dependencies | Description |
|---|---|---|---|
|
|
| |
2. core | mxnet-core_2.11 |
| |
3. macros | mxnet-macros_2.11 |
| |
| 4. init | mxnet-init_2.11 | ||
5. init-native |
| mxnet-init_2.11 | |
6. native |
| mxnet-core_2.11 | |
| 7. infer | mxnet-infer_2.11 |
| |
| 8. spark | mxnet-spark_2.11 |
| |
| 9. examples | mxnet-examples_2.11 |
|
Currently 3 Packages from the above generated are published to Maven. `mxnet_full_2.11-{platform}`.
The following flags are used to build the MXNet Backend(derived from mxnet pip compiler flags)
| FLAGS | MXNet-Scala OSX | MXNet-Scala Linux CPU | MXNet-Scala Linux GPU |
|---|---|---|---|
| USE_BLAS | 0 | ||
| USE_LAPACK | 1 | 1 | 1 |
| USE_OPENCV | 1 | 1 (opencv 3.4.1) | 1 |
| USE_S3 | 1 | 1 | 1 |
| USE_DIST_KVSTORE | 0 | 1 | 1 |
| USE_OPENMP | 1 | 1 | |
| USE_CUDA | 1 | ||
USE_CUDNN | 0 | 1 | |
| USE_NCCL | 0 | 1 | |
| USE_MKLDNN | 0 | 0 | |
| Java version | $ java -version | ubuntu@ip:~$ java -version |
Setup GPG Keys https://www.apache.org/dev/openpgp.html#key-gen-install-latest-gnupg
git clone https://github.com/$YOURFORK/incubator-mxnet ~/mxnet-$RELEASEVERSION cd ~/mxnet-$RELEASEVERSION git remote add upstream https://github.com/apache/incubator-mxnet git fetch upstream v1.2.0 && git checkout -b $RELEASEVERISON --track upstream/v1.2.0 git submodule update --init --recursive |
#try opencv3 next time you release brew install lapack openblas opencv@2 |
#This will build the MXNet backend with the tested configuration for publishing MXNet-Scala to Maven. cd ~/mxnet-$RELEASEVERSION scala-package/dev/compile-mxnet-backend.sh osx-x86_64-cpu |
Now setup your maven create a master password and paste the encrypted the more info Password Encryption
touch ~/.m2/settings-security.xml && mvn --encrypt-master-password |
Paste the encrypted master-password you get from above in `~/.m2/settings-security.xml`
<settingsSecurity> <master>ENCRYPTEDMASTERPASSWORD</master> </settingsSecurity> |
Encrypt your password for `apache` login
cp /usr/share/maven/conf/settings.xml ~/.m2/ && mvn --encrypt-password |
Paste the below with your apache username and encrypted apache password updated under `<servers>`tag, this will used to publish to apache snapshot and staging in `~/.m2/settings.xml`.
<!-- To publish a snapshot of some part of Maven -->
<server>
<id>apache.snapshots.https</id>
<username>APACHEUSERNAME</username>
<password>APACHEPASSWORDENCRYPTED</password>
</server>
<!-- To stage a release of some part of Maven -->
<server>
<id>apache.releases.https</id>
<username>APACHEUSERNAME</username>
<password>APACHEPASSWORDENCRYPTED</password>
</server> |
Follow Maven Publish Link https://www.apache.org/dev/publishing-maven-artifacts.html to setup repo settings, etc., before you proceed
#if you don't set this it will fail with the error "gpg: signing failed: Inappropriate ioctl for device" export GPG_TTY=$(tty) #build scalapkg make scalapkg #make sure everything is working as expected make scalarelease-dryrun # Releases to Apache Snapshot make scaladeploy. This will upload to the snapshot repo # Verify that the package has been uploaded in https://repository.apache.org/content/repositories/snapshots/ # Preparing the release will create the new tag, automatically checking in on your behalf.(We have disabled CheckIn) make scalarelease-prepare # this will upload to the staging repository and make it for testing make scalarelease-perform |
You have to close the repo for it to become available on the Staging Repo, find details here
Once you have tested the staged package, you can promote the package and it will become available on maven central, find details here
To build for linux-cpu platform, we use the AWS Deep Learning Base AMI that comes built with all the necessary
libraries such as opencv/cudnn etc,.
First you need to export the gpg keys you created on Mac and import to the linux host
Install Maven, gnupg, gnupg2 & gnupg-agent
sudo apt-get install -y maven gnupg gnupg2 gnupg-agent && #soft link gnupg2 to gnupg, Update the below when you find a better solution sudo mv /usr/bin/gpg /usr/bin/gpg.1 && sudo ln -s /usr/bin/gpg2 /usr/bin/gpg |
Export and ship gpg keys you created on the Mac
gpg --export-secret-keys -a xxx@apache.org > ~/Downloads/keys.asc scp ~/Downloads/keys.asc ubuntu@host:~/keys.asc |
Import and verify they look fine
gpg --import ~/keys.asc #note the key id gpg --list-keys |
Update the gnupg-agent cachettl, it takes a long time to build on linux and linux-gpu instances and will time out during release if you don't baby sit the build
mkdir -p ~/.gnupg && touch ~/.gnupg/gpg-agent.conf #set to 4 hours echo "default-cache-ttl 14400" > ~/.gnupg/gpg-agent.conf && echo "max-cache-ttl 14400" >> ~/.gnupg/gpg-agent.conf #key id you go when you ran gpg --listkeys echo "default-key KEYIDFROMABOVE" >> ~/.gnupg/gpg-agent.conf |
Start gpg-agent to use the above
eval $(gpg-agent --daemon) |
Follow steps to clone the repo and checkout the correct release branch
cd ~/mxnet-$RELEASEVERSION ./scala-package/dev/compile-mxnet-backend.sh linux-x86_64-gpu ./ |
Follow steps from OSX
Build on GPU
Follow all steps for Linux-CPU
Update the nccl location approriately in ~/incubator-mxnet/scala-package/dev/compile-backend.sh
```echo "ADD_LDFLAGS += -L/lib/nccl/cuda-9.0/lib```
cd ~/mxnet-$RELEASEVERSION ./scala-package/dev/compile-mxnet-backend.sh linux-x86_64-gpu ./ |
Update your pom file to include the repository info like below:
<repository>
<id>snapshot</id>
<url>https://repository.apache.org/content/repositories/staging</url>
</repository>
</repositories>
<dependencies>
<dependency>
<groupId>org.apache.mxnet</groupId>
<artifactId>mxnet-full_2.11-osx-x86_64-cpu</artifactId>
<version>1.2.0</version>
</dependency>
</dependencies |
git push --force --delete ./ refs/heads/mxnet-parent_2.11-1.2.0 git tag -d mxnet-parent_2.11-1.2.0 |
sudo apt-get install gnupg2 sudo mv /usr/bin/gpg /usr/bin/gpg.1 sudo ln -s /usr/bin/gpg2 /usr/bin/gpg |
```export GPG_TTY=$(tty)```