feat: Use categorical encoder in multiclass example

examples/classification/binary_iris.ts

@@ -16,7 +16,7 @@ import {
   ClassificationReport,
   // Split the dataset
   useSplit,
-} from "https://deno.land/x/vectorizer@v0.2.3/mod.ts";
+} from "https://deno.land/x/vectorizer@v0.3.6/mod.ts";
 
 // Define classes
 const classes = ["Setosa", "Versicolor"];

examples/classification/iris.ts

@@ -10,6 +10,7 @@ import {
   Sequential,
   setupBackend,
   SoftmaxLayer,
+  tensor,
   tensor1D,
   tensor2D,
 } from "../../mod.ts";
@@ -21,37 +22,39 @@ import {
   ClassificationReport,
   // Split the dataset
   useSplit,
-} from "https://deno.land/x/[email protected]/mod.ts";
-
-// Define classes
-const classes = ["setosa", "versicolor", "virginica"];
+  // One-hot encoding of targets
+  CategoricalEncoder,
+} from "https://deno.land/x/[email protected]/mod.ts";
 
 // Read the training dataset
 const _data = Deno.readTextFileSync("examples/classification/iris.csv");
 const data = parse(_data);
 
 // Get the predictors (x) and targets (y)
-const x = data.map((fl) => fl.slice(0, 2).map(Number));
-const y = data.map((fl) => {
-  const yi = new Array(classes.length).fill(0);
-  yi[classes.indexOf(fl[4])] = 1;
-  return yi;
-});
+const x = data.map((fl) => fl.slice(0, 4).map(Number));
+const y_pre = data.map((fl) => fl[4]);
+
+const encoder = new CategoricalEncoder()
+
+const y = encoder.fit(y_pre).transform(y_pre, "f32")
 
 // Split the dataset for training and testing
+// @ts-ignore Matrices can be split
 const [train, test] = useSplit({ ratio: [7, 3], shuffle: true }, x, y) as [
   [typeof x, typeof y],
-  [typeof x, typeof y],
+  [typeof x, typeof y]
 ];
 
 // Setup the CPU backend for Netsaur
 await setupBackend(CPU);
 
+console.log(train[1])
+
 // Create a sequential neural network
 const net = new Sequential({
   // Set number of minibatches to 4
   // Set size of output to 4
-  size: [4, 2],
+  size: [4, 4],
 
   // Disable logging during training
   silent: false,
@@ -80,11 +83,11 @@ net.train(
   [
     {
       inputs: tensor2D(train[0]),
-      outputs: tensor2D(train[1]),
+      outputs: tensor(train[1].data as Float32Array, train[1].shape),
     },
   ],
   // Train for 300 epochs
-  300,
+  400,
   1,
   0.02,
 );
@@ -95,8 +98,8 @@ console.log(`training time: ${performance.now() - time}ms`);
 const res = await Promise.all(
   test[0].map((input) => net.predict(tensor1D(input))),
 );
-const y1 = res.map((x) => x.data.indexOf(Math.max(...x.data)));
-const y0 = test[1].map((x) => x.indexOf(1));
+const y1 = res.map((x) => encoder.getOg(x.data.indexOf(Math.max(...x.data))));
+const y0 = encoder.untransform(test[1]);
 
 console.log(y1.map((x, i) => [y0[i], x]));
 const cMatrix = new ClassificationReport(y0, y1);

examples/classification/spam.ts

@@ -18,7 +18,7 @@ import {
   TextVectorizer,
   // Split the dataset
   useSplit,
-} from "https://deno.land/x/vectorizer@v0.2.3/mod.ts";
+} from "https://deno.land/x/vectorizer@v0.3.6/mod.ts";
 
 // Define classes
 const ymap = ["spam", "ham"];