The main purpose of this demo is to illustrate how to connect TensorFlow, Infinispan and Ceph. However, it also shows how to connect various parts of the application (even written in different languages) and pass the data through whole application stack using Infinispan.
The demo uses MNIST data sample. From the user perspective, user selects a test image in a custom application and once he clicks on the image, the result (which number is on the image) is immediately shown together with original image in the web application.
High level architecture is on the figure bellow.
Data flow is as follows:
-
The test data sample is loaded into the simple MNIST data browser writtein in Qt/C++.
-
Upon the user click on the selected image, image is sent to the Infinispan server via REST.
-
Once the data arrives into Infinispan, custom cache event (containing image ID and its byte content) is sent to TensorFlow classifier - a separate application which contains remote Infinispan client and TF classifier itself.
-
The result is sent back to Infinispan.
-
Having the result, Infinispan notifies NodeJS application listening for the new results, which sends the result of the classification together with original image to the web brower of the user.
-
Behind the scene, Infinispan offloads the data into Ceph data store as the amount of data grows. Eviction policy is set to keep 5 most recent entries in Infispan, rest is moved into the Ceph as new data arrivies.
To be able to build and run the demo, you will need following components installed:
-
TensorFlow (tested with TenforFlow
0.11) -
javacpp-presets (tested with javacpp-presets
0.11.0-1.3.1-SNAPSHOT) -
Qt framework (tested with Qt
5.7) -
Infinispan server (tested with
9.0.0.Alpha4) -
NodeJS (tested with Fedora NodeJS package
nodejs-4.6.1-6) -
Ceph (tested with Fedora Ceph packages
10.2.4-2, installed byceph-deploy-1.5.32-2)
You need to download at least test MNIST data sample (t10k-images-idx3-ubyte.gz, you can download it here), put it into /tmp/mnist/data/ and unpack.
If you store it in diffrent location, just change the path in qt-sender/src/mnist/main.cpp.
The app also expects that the neural network is already trained and saved in TF checkpoint file.
For training NN and creating checkpoint file you can use script img-classifier/src/main/python/nn2h.py.
Simply run the script, it will download required data automatically.
Once the NN is trained, checkpoint file is store in /tmp/my-model/nn2h.ckpt.
You should see success rate of the NN on the standard output.
You need to build following pieces:
-
Infinispan server converter in
hr-server-converterby simplemvn clean package. Copy the resulting bits (target/hr-server-converter-0.0.1-SNAPSHOT.jar) into$ISPN_HOME/standalone/deployments. -
TF classifier app in
img-classifieragain by simplemvn clean package -
Build MNIST sender app in
qt-senderby
mkdir build && cd build /opt/qt/5.7/gcc_64/bin/qmake ../mnist-sender.pro make
-
Clone Infinispan Ceph Cachestore and build it by
mvn package -Puberjar. Copy the resulting bits (target/infinispan-cachestore-ceph-1.0-SNAPSHOT-jar-with-dependencies.jar) into$ISPN_HOME/standalone/deployments. -
Download NodeJS dependencies by runing
npm installinnodejs-consumerdir or wherever you want to run this app.
-
Copy
conf/standalone.xmlinto$ISPN_HOME/standalone/configuration. -
In
standalone.xmlfind element<store class="org.infinispan.persistence.ceph.CephStore">and setup correct IP addres of the Ceph monitor and credentials in properties of this element.
To avoid problems, start the pieces of the app in the following order:
-
Start Ceph data store.
-
Start Infinispan server.
-
Run TF classifier in
img-classifierbyjava -Djava.library.path=/opt/lib/tensorflow/r0.11 -jar target/tf-ispn-classifier-0.0.1-SNAPSHOT.jar -
Run NodeJs app
-
Run
mnist-senderinqt-sender/build -
Start the web browser and connect to
http://localhost:3000