README.md

EmotionRecServer

Ktor server for providing emotion recognition. This is done through a trained TensorFlow model that is either located locally or uploaded to Google Cloud Platform ML Engine.

I wrote an article covering the specific TensorFlow model I trained.

Modes of inference

• Local Inference
• GCP Inference [ML Engine]

To be able to choose which inference mode the server will use you will have to set a property in the application configuration found in /api_ktor/src/main/resources/application.conf.

ktor {
    ...
    application {
        ...
        gcp = false
    }
}

Setting gcp to false will configure the server to use the local inference otherwise it will use the model hosted in GCP.

Local Inference

LocalInference will use the Java TensorFlow Api to load the trained model.

To be able to use local inference you first have to point the Server to the location of your model.

TODO: Currently this is done through a static variable but this should be moved to a config file.

/api_ktor/src/main/kotlin/com/emotionrec/api/Server.kt

val LOCAL_INF_MODEL = "./src/main/resources/1"
val LOCAL_INF_TAG = "serve"

The first is the location of the model relative to the project and the other is the tag used while saving the model through the SavedModelApi.

GCP ML Engine Inference

In order to use the the predictions coming from the model hosted on GCP you will have to upload a saved model to your GCP account. You will need to have two things.

• Path to your model
• Credential File

Path to your model

"projects/ml-happy-rec/models/happy_rec_model/versions/v2:predict"

This is a hardcoded path found in the RetrofitNetwork.kt and should be changed to point to your specific model.

Credentials

TODO: Value should be set through config

/api_ktor/src/main/kotlin/com/emotionrec/api/Server.kt

val GOOGLE_CRED_FILE = "happy_rec_cred.json"

This is currently handled through the use of the Google credential file given through GCP.

Running the Server

To run the api_ktor application: ./gradlew api_ktor:run which will use the default settings (port) defined in the application.conf file.

ktor {
    deployment {
        port = 8378
        environment = development
        watch = [ emotionrec ]
    }

    application {
        id = emotionrec
        modules = [com.emotionrec.api.ServerKt.main]
        gcp = false
    }
}

Simple Api

GET /ping

• Used for sanity checks and returns "pong"

POST /prediction

• Accepts [PostPredictionData].

• Expects the [PostPredictionData.image_array]:

  • to be an array of size 2304
  • String array separated by a delimiter [PostPredictionData.delimiter] (default: [DEFAULT_DELIMITER])

• Responds with [PredictionError] or [PredictionResponse]

POST /predictionImage

• Accepts mutlipart file image upload
• Responds with [PredictionError] or [PredictionResponse]

PredictionResponse

{
    "sortedPredictions": [
        {
            "probability": 0.99999285,
            "emotion": "ANGRY"
        },
        {
            "probability": 0.0000035176417,
            "emotion": "SAD"
        },
        {
            "probability": 0.0000018190486,
            "emotion": "FEAR"
        },
        {
            "probability": 0.0000018007337,
            "emotion": "NEUTRAL"
        },
        {
            "probability": 1.873281e-8,
            "emotion": "HAPPY"
        },
        {
            "probability": 3.4072745e-11,
            "emotion": "DISGUST"
        },
        {
            "probability": 2.9763858e-12,
            "emotion": "SURPRISE"
        }
    ],
    "guessedPrediction": {
        "probability": 0.99999285,
        "emotion": "ANGRY"
    }
}