Training tweaks

Author: generic-github-user

script.js

@@ -2,12 +2,16 @@
 
 // Define settings
 const numParameters = 3;
-const numLayers = 6;
+const numLayers = 10;
 // Size of input and output images in pixels (width and height)
-const imageSize = 8;
+const imageSize = 32;
 // Number of images to use when training the neural network
 const numTrainingImages = 15;
-const logData = false;
+const logData = true;
+const optimizer = {
+"generator": tf.train.adam(0.01),
+"discriminator": tf.train.adam(0.01)
+}
 
 // Automatically generated settings and parameters
 // Volume of image data, calculated by squaring imageSize to find the area of the image (total number of pixels) and multiplying by three for each color channel (RGB)
@@ -43,13 +47,13 @@ const generator = {
 // Evaluate the loss function given the output of the autoencoder network and the actual image
 return loss(
 discriminator.model.predict(
-generator.model.predict(parameters.training)
+generator.model.predict(parameters.training).clipByValue(0, 255)
 ),
 tf.ones([15, numParameters])
 );
 }
 ),
-"optimizer": tf.train.adam(0.01)
+"optimizer": optimizer.generator
 };
 
 if (logData) {
@@ -59,7 +63,7 @@ if (logData) {
 generator.model.add(tf.layers.dense({units: numParameters, inputShape: [numParameters]}));
 for (var i = 0; i < numLayers; i ++) {
 const layerSize = Math.round(imageVolume / (2 ** ((numLayers - 1) - i)));
-generator.model.add(tf.layers.dense({units: layerSize, activation: "tanh"}));
+generator.model.add(tf.layers.dense({units: layerSize, activation: "relu"}));
 if (logData) {
 console.log(layerSize);
 }
@@ -80,7 +84,7 @@ const discriminator = {
 );
 }
 ),
-"optimizer": tf.train.adam(0.01)
+"optimizer": optimizer.discriminator
 };
 
 if (logData) {
@@ -170,7 +174,7 @@ trainingData.images[trainingData.images.length - 1].onload = function () {
 // Uncaught Error: Constructing tensor of shape (92160) should match the length of values (46095)
 const generated =
 tf.tidy(
-() => generator.model.predict(parameters.display).dataSync()
+() => generator.model.predict(parameters.display).clipByValue(0, 255).dataSync()
 );
 const generatedArray = [];
 generated.forEach(
@@ -252,20 +256,20 @@ trainingData.images[trainingData.images.length - 1].onload = function () {
 // Decode the low-dimensional representation of the input data created by the encoder
 return generator.model.predict(parameters.display)
 // Clip pixel values to a 0 - 255 (int32) range
-.clipByValue(0, 1)
+.clipByValue(0, 255)
 // Reshape the output tensor into an image format (W * L * 3)
 .reshape(
 [imageSize, imageSize, 3]
 )
 }
 );
-// output.dtype = "int32";
+output.dtype = "int32";
 
 const discriminatorOutput =
 tf.tidy(
 () => {
 return discriminator.model.predict(
-generator.model.predict(parameters.display)
+generator.model.predict(parameters.display).clipByValue(0, 255)
 );
 }
 );