Updated inference examples

examples/src/main/kotlin/examples/inference/production/LeNetFashionMnistExportImport.kt → examples/src/main/kotlin/examples/inference/production/fashionmnist/LeNetFashionMnistExportImport.kt

@@ -32,6 +32,15 @@ private val fashionMnistLabelEncoding = mapOf(
     9 to "Ankle boot"
 )
 
+/**
+ * This examples demonstrates model and model weights export and import back.
+ *
+ * Models is exported as graph in .pb format, weights are exported in custom (txt) format.
+ *
+ * Model is trained on FashionMnist dataset.
+ *
+ * It saves all the data to the project root directory.
+ */
 fun main() {
     val (train, test) = Dataset.createTrainAndTestDatasets(
         FASHION_TRAIN_IMAGES_ARCHIVE,
@@ -57,10 +66,6 @@ fun main() {
             verbose = true
         )
 
-        val weights = it.layers[0].getWeights() // first conv2d layer
-
-        drawFilters(weights[0])
-
         val accuracy = it.evaluate(dataset = test, batchSize = TEST_BATCH_SIZE).metrics[Metrics.ACCURACY]
 
         println("Accuracy $accuracy")

examples/src/main/kotlin/examples/inference/production/LeNetFashionMnistImport.kt → examples/src/main/kotlin/examples/inference/production/fashionmnist/LeNetFashionMnistImport.kt

@@ -12,7 +12,11 @@ import java.io.File
 
 private const val PATH_TO_MODEL = "savedmodels/fashionLenet"
 
-
+/**
+ * Inference model is used here, separately from model training code to illustrate the ability to load model graph and weights to start prediction process.
+ *
+ * NOTE: The example requires the saved model in the appropriate directory (run LeNetFashionMnistExportImport.kt firstly).
+ */
 fun main() {
     val (train, test) = Dataset.createTrainAndTestDatasets(
         FASHION_TRAIN_IMAGES_ARCHIVE,

examples/src/main/kotlin/examples/inference/production/LeNetMnistExportImport.kt → examples/src/main/kotlin/examples/inference/production/mnist/LeNetMnistExportImport.kt

@@ -19,6 +19,15 @@ private const val EPOCHS = 1
 private const val TRAINING_BATCH_SIZE = 500
 private const val TEST_BATCH_SIZE = 1000
 
+/**
+ * This examples demonstrates model and model weights export and import back.
+ *
+ * Models is exported as graph in .pb format, weights are exported in custom (txt) format.
+ *
+ * Model is trained on Mnist dataset.
+ *
+ * It saves all the data to the project root directory.
+ */
 fun main() {
     val (train, test) = Dataset.createTrainAndTestDatasets(
         TRAIN_IMAGES_ARCHIVE,

examples/src/main/kotlin/examples/inference/production/LeNetMnistExportImportWithJSON.kt → examples/src/main/kotlin/examples/inference/production/mnist/LeNetMnistExportImportWithJSON.kt

@@ -22,6 +22,17 @@ private const val EPOCHS = 1
 private const val TRAINING_BATCH_SIZE = 1000
 private const val TEST_BATCH_SIZE = 1000
 
+/**
+ * This examples demonstrates model and model weights export and import back.
+ *
+ * Models is exported in Keras-style json format, weights are exported in custom (txt) format.
+ *
+ * After loading model is trained again with another optimizer with frozen Conv2D layers. Only weights in Dense layers can be updated.
+ *
+ * Model is trained on Mnist dataset.
+ *
+ * It saves all the data to the project root directory.
+ */
 fun main() {
     val (train, test) = Dataset.createTrainAndTestDatasets(
         TRAIN_IMAGES_ARCHIVE,

examples/src/main/kotlin/examples/inference/production/LeNetMnistImport.kt → examples/src/main/kotlin/examples/inference/production/mnist/LeNetMnistImport.kt

@@ -12,6 +12,11 @@ import java.io.File
 
 private const val PATH_TO_MODEL = "savedmodels/lenet5"
 
+/**
+ * Inference model is used here, separately from model training code to illustrate the ability to load model graph and weights to start prediction process.
+ *
+ * NOTE: The example requires the saved model in the appropriate directory (run LeNetMnistExportImport.kt firstly).
+ */
 fun main() {
     val (train, test) = Dataset.createTrainAndTestDatasets(
         TRAIN_IMAGES_ARCHIVE,
@@ -35,8 +40,6 @@ fun main() {
 
             if (prediction == getLabel(train, imageId))
                 accuracy += (1.0 / amountOfTestSet)
-
-            //println("Prediction: $prediction Ground Truth: ${getLabel(train, imageId)}")
         }
         println("Accuracy: $accuracy")
     }

examples/src/main/kotlin/examples/inference/production/optimizers/LeNetMnistExportImportWithJSON_Adam.kt

@@ -22,6 +22,13 @@ private const val EPOCHS = 1
 private const val TRAINING_BATCH_SIZE = 1000
 private const val TEST_BATCH_SIZE = 1000
 
+/**
+ * This examples demonstrates model, model weights and optimizer weights export and import back.
+ *
+ * Models is exported in json format, weights are exported in custom (txt) format.
+ *
+ * It saves all the data to the project root directory.
+ */
 fun main() {
     val (train, test) = Dataset.createTrainAndTestDatasets(
         TRAIN_IMAGES_ARCHIVE,

examples/src/main/kotlin/examples/inference/production/LeNetFashionMnistVisualisation.kt → examples/src/main/kotlin/examples/inference/production/visualisation/LeNetFashionMnistVisualisation.kt

@@ -28,6 +28,11 @@ private val fashionMnistLabelEncoding = mapOf(
     9 to "Ankle boot"
 )
 
+/**
+ * This examples demonstrates model activations and Conv2D filters visualisation.
+ *
+ * Model is trained on FashionMnist dataset.
+ */
 fun main() {
     val (train, test) = Dataset.createTrainAndTestDatasets(
         FASHION_TRAIN_IMAGES_ARCHIVE,

examples/src/main/kotlin/examples/inference/production/LeNetMnist.kt → examples/src/main/kotlin/examples/inference/production/visualisation/LeNetMnistVisualisation.kt

@@ -10,12 +10,16 @@ import org.jetbrains.kotlinx.dl.api.core.metric.Metrics
 import org.jetbrains.kotlinx.dl.api.core.optimizer.Adam
 import org.jetbrains.kotlinx.dl.datasets.Dataset
 import org.jetbrains.kotlinx.dl.datasets.handlers.*
-import javax.swing.JFrame
 
 private const val EPOCHS = 1
 private const val TRAINING_BATCH_SIZE = 500
 private const val TEST_BATCH_SIZE = 1000
 
+/**
+ * This examples demonstrates model activations and Conv2D filters visualisation.
+ *
+ * Model is trained on Mnist dataset.
+ */
 fun main() {
     val (train, test) = Dataset.createTrainAndTestDatasets(
         TRAIN_IMAGES_ARCHIVE,
@@ -72,28 +76,3 @@ fun main() {
         println("Ground Truth: $maxIdx")
     }
 }
-
-fun drawActivations(activations: List<*>) {
-    val frame = JFrame("Visualise the matrix weights on Relu")
-    frame.contentPane.add(ReluGraphics(activations[0] as Array<Array<Array<FloatArray>>>))
-    frame.setSize(1500, 1500)
-    frame.isVisible = true
-    frame.defaultCloseOperation = JFrame.EXIT_ON_CLOSE
-    frame.isResizable = false
-
-    val frame2 = JFrame("Visualise the matrix weights on Relu_1")
-    frame2.contentPane.add(ReluGraphics2(activations[1] as Array<Array<Array<FloatArray>>>))
-    frame2.setSize(1500, 1500)
-    frame2.isVisible = true
-    frame2.defaultCloseOperation = JFrame.EXIT_ON_CLOSE
-    frame2.isResizable = false
-}
-
-fun drawFilters(filters: Array<*>, colorCoefficient: Double = 2.0) {
-    val frame = JFrame("Filters")
-    frame.contentPane.add(Conv2dJPanel(filters as Array<Array<Array<FloatArray>>>, colorCoefficient))
-    frame.setSize(1000, 1000)
-    frame.isVisible = true
-    frame.defaultCloseOperation = JFrame.EXIT_ON_CLOSE
-    frame.isResizable = false
-}

examples/src/main/kotlin/examples/inference/production/VisualisationHelpers.kt → examples/src/main/kotlin/examples/inference/production/visualisation/VisualisationHelpers.kt

@@ -7,6 +7,7 @@ package examples.inference.production
 
 import java.awt.Color
 import java.awt.Graphics
+import javax.swing.JFrame
 import javax.swing.JPanel
 import kotlin.math.max
 import kotlin.math.min
@@ -151,3 +152,28 @@ class ReluGraphics2(private val dst: Array<Array<Array<FloatArray>>>) : JPanel()
         }
     }
 }
+
+fun drawActivations(activations: List<*>) {
+    val frame = JFrame("Visualise the matrix weights on Relu")
+    frame.contentPane.add(ReluGraphics(activations[0] as Array<Array<Array<FloatArray>>>))
+    frame.setSize(1500, 1500)
+    frame.isVisible = true
+    frame.defaultCloseOperation = JFrame.EXIT_ON_CLOSE
+    frame.isResizable = false
+
+    val frame2 = JFrame("Visualise the matrix weights on Relu_1")
+    frame2.contentPane.add(ReluGraphics2(activations[1] as Array<Array<Array<FloatArray>>>))
+    frame2.setSize(1500, 1500)
+    frame2.isVisible = true
+    frame2.defaultCloseOperation = JFrame.EXIT_ON_CLOSE
+    frame2.isResizable = false
+}
+
+fun drawFilters(filters: Array<*>, colorCoefficient: Double = 2.0) {
+    val frame = JFrame("Filters")
+    frame.contentPane.add(Conv2dJPanel(filters as Array<Array<Array<FloatArray>>>, colorCoefficient))
+    frame.setSize(1000, 1000)
+    frame.isVisible = true
+    frame.defaultCloseOperation = JFrame.EXIT_ON_CLOSE
+    frame.isResizable = false
+}