Description

Run image classification engine on Apache PredictionIO using a Tensorflow model.

This template

• follows the workflow in official Tensorflow tutorial for image classification.
• uses a pre-trained model from the inception challenge. The purpose of the engine is to be able to deploy a Tensorflow model and do inference via HTTP.
• for experimental purposes, builds Tensorflow from source using bazel build, instead of calling the Tensorflow Java API. For more details, see src/main/cpp/jna.cc and Future explorations section below.

Installation

Prerequisites

• PredictionIO
• Bazel
• Swig

Instructions

./build1.sh

tensorflow/configure

./build2.sh

Workflow

Downloading the model

Run data/download.sh to download the pre-trained imagenet model.

Configuring PredictionIO

In conf/pio-env.sh,

• PIO_STORAGE_REPOSITORIES_MODELDATA_SOURCE=LOCALFS
• PIO_STORAGE_SOURCES_LOCALFS_TYPE=localfs
• Set PIO_STORAGE_SOURCES_LOCALFS_PATH to the directory where you want to store the model metadata information. By default, this is $HOME/.pio_store/models.

Deploying the engine

• Create a new PredictionIO app, and modify engine.json accordingly.
• After installing PredictionIO, run pio build in the engine directory. This produces fat jars in target/scala-2.x/.
• Run pio train. While we're using a pre-trained model, this command writes the new engine metadata to the PredictionIO metadata storage, and is required for deploy.
• Run pio deploy. By default, this deploys the engine using port 8000.

Testing the engine

There are two ways of serving data to the engine.

1. Put the target image in data/images. You can change this path in engine.json. Then use image param as the filename of the target image such as curl -H "Content-Type: application/json" -d '{ "image":"cropped_panda.jpg" }' http://localhost:8000/queries.json.

2. Use data param to send a UTF-8 encoded string of the target image data.

If all goes well, the engine will return a JSON result such as {"labelId":169,"categories":"giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca","confidence":0.8910738229751587}

Future explorations

Enabling native training

In the Tensorflow community, there is active work in progress to support gradient operations via the Tensorflow C API. Using the C API, we can include such operations in our native java binding that we create. This will open possibilities to train a deep learning model natively, without using python scripts to create a protobuf file, or converting data between numpy arrays and java arrays which causes overhead.