Normalise the input data

src/lib/DataPreprocessor.mjs

@@ -30,10 +30,14 @@ class DataPreprocessor {
 		
 		let filenames = await fs.promises.readdir(dirpath);
 		for(let filename of filenames) {
-			let imagetensor = tf.node.decodeImage(
+			let imagebin = await fs.promises.readFile(path.join(dirpath, filename));
-				await fs.promises.readFile(path.join(dirpath, filename)),
+			let imagetensor = tf.tidy(() => {
-				3 // channels
+				// Normalise the data to be 0 - 1
-			);
+				return tf.node.decodeImage(
+					imagebin,
+					3 // channels
+				).cast("float32").div(255);
+			});
 			if(imagetensor.shape[0] == 256
 				&& imagetensor.shape[1] == 256
 				&& imagetensor.shape[0] == 3)

src/lib/model.mjs

@@ -13,7 +13,7 @@ export default function(cats_count, image_size, use_crossentropy = false) {
 		kernelSize: 5,
 		filters: 3,
 		strides: 1,
-		activation: "relu"
+		activation: "tanh"
 	}));
 	if(image_size > 32) {
 		model.add(tf.layers.conv2d({
@@ -22,7 +22,7 @@ export default function(cats_count, image_size, use_crossentropy = false) {
 			kernelSize: 5,
 			filters: 3,
 			strides: 2,
-			activation: "relu"
+			activation: "tanh"
 		}));
 		model.add(tf.layers.conv2d({
 			name: "conv2d_3",
@@ -30,7 +30,7 @@ export default function(cats_count, image_size, use_crossentropy = false) {
 			kernelSize: 5,
 			filters: 3,
 			strides: 2,
-			activation: "relu"
+			activation: "tanh"
 		}));
 	}