zaleslaw

val trainDataset = OnHeapDataset.create(File(datasetPath, "train"), labelGenerator, preprocessing)
val valDataset = OnHeapDataset.create(File(datasetPath, "val"), labelGenerator, preprocessing)

zaleslaw

val preprocessing = preprocess {
 transformImage {
   centerCrop {
     size = 214
   }
   pad {
     top = 10
     bottom = 10
     left = 10
     right = 10

zaleslaw

model.use {
   it.compile(
       optimizer = Adam(clipGradient = ClipGradientByValue(0.1f)),
       loss = Losses.SOFT_MAX_CROSS_ENTROPY_WITH_LOGITS,
       metric = Metrics.ACCURACY
   )

   it.logSummary()

   it.fit(

zaleslaw

val earlyStopping = EarlyStopping(
   monitor = EpochTrainingEvent::valLossValue,
   minDelta = 0.0,
   patience = 2,
   verbose = true,
   mode = EarlyStoppingMode.AUTO,
   baseline = 0.1,
   restoreBestWeights = false
)
val terminateOnNaN = TerminateOnNaN()

zaleslaw

model.use {
   it.compile(
       optimizer = Adam(),
       loss = Losses.SOFT_MAX_CROSS_ENTROPY_WITH_LOGITS,
       metric = Metrics.ACCURACY
   )

   it.loadWeightsForFrozenLayers(hdfFile)

   it.fit(

zaleslaw

val topModel = Sequential.of(
   GlobalAvgPool2D(
       name = "top_avg_pool",
   ),
   Dense(
       name = "top_dense",
       kernelInitializer = GlorotUniform(),
       biasInitializer = GlorotUniform(),
       outputSize = 200,
       activation = Activations.Relu

zaleslaw

val modelHub = TFModelHub(cacheDirectory = File("cache/pretrainedModels"))

val modelType = TFModels.CV.ResNet50(noTop = true, inputShape = intArrayOf(IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))

val noTopModel = modelHub.loadModel(modelType)

zaleslaw

model.use { poseDetectionModel ->
       val imageFile = …
       val detectedPoses = poseDetectionModel.detectPoses(imageFile = imageFile, confidence = 0.0f)

       detectedPoses.multiplePoses.forEach { detectedPose ->
           println("Found ${detectedPose.first.classLabel} with probability ${detectedPose.first.probability}")
           detectedPose.second.poseLandmarks.forEach {
               println("Found ${it.poseLandmarkLabel} with probability ${it.probability}")
           }

zaleslaw

    val modelHub = ONNXModelHub(cacheDirectory = File("cache/pretrainedModels"))
    val model = ONNXModels.PoseDetection.MoveNetMultiPoseLighting.pretrainedModel(modelHub)

zaleslaw

model.use { poseDetectionModel ->
       val imageFile = …
       val detectedPose = poseDetectionModel.detectPose(imageFile = imageFile)

       detectedPose.poseLandmarks.forEach {
           println("Found ${it.poseLandmarkLabel} with probability ${it.probability}")
       }

       detectedPose.edges.forEach {
           println("The ${it.poseEdgeLabel} starts at ${it.start.poseLandmarkLabel} and ends with ${it.end.poseLandmarkLabel}")