chore: Upgrade net to handle types

.eslintrc.json → .eslintrc

@@ -40,6 +40,7 @@
     "no-underscore-dangle": "off",
     "prettier/prettier": "error",
     "semi": "off",
-    "standard/no-callback-literal": "off"
+    "standard/no-callback-literal": "off",
+    "no-implied-eval": "off"
   }
 }

src/cross-validate.test.ts

@@ -9,7 +9,7 @@ import { LSTMTimeStep } from './recurrent/lstm-time-step';
 
 describe('CrossValidate', () => {
   describe('.train()', () => {
-    class FakeNN extends NeuralNetwork {
+    class FakeNN extends NeuralNetwork<number[], number[]> {
       constructor(
         options: Partial<
           INeuralNetworkOptions & INeuralNetworkTrainOptions
@@ -195,7 +195,7 @@ describe('CrossValidate', () => {
     });
   });
   describe('.fromJSON()', () => {
-    class FakeNN extends NeuralNetwork {}
+    class FakeNN extends NeuralNetwork<number[], number[]> {}
     it("creates a new instance of constructor from argument's sets.error", () => {
       const cv = new CrossValidate(FakeNN);
       const options = { inputSize: 1, hiddenLayers: [10], outputSize: 1 };
@@ -241,7 +241,7 @@ describe('CrossValidate', () => {
     });
   });
   describe('.toNeuralNetwork()', () => {
-    class FakeNN extends NeuralNetwork {}
+    class FakeNN extends NeuralNetwork<number[], number[]> {}
     it('creates a new instance of constructor from top .json sets.error', () => {
       const cv = new CrossValidate(FakeNN);
       const details = {

src/feed-forward.unit.test.ts

@@ -77,6 +77,7 @@ describe('FeedForward Class: Unit', () => {
                 {
                   filterHeight: 3,
                   filterWidth: 3,
+                  filterCount: 1,
                   padding: 2,
                   stride: 2,
                 },
@@ -100,6 +101,7 @@ describe('FeedForward Class: Unit', () => {
                   padding: 2,
                   filterWidth: 3,
                   filterHeight: 3,
+                  filterCount: 1,
                   stride: 3,
                 },
                 inputLayer
@@ -171,6 +173,7 @@ describe('FeedForward Class: Unit', () => {
                   {
                     filterWidth: 3, // TODO: setting height, width should behave same
                     filterHeight: 3,
+                    filterCount: 3,
                     padding: 2,
                     stride: 3,
                   },
@@ -187,6 +190,7 @@ describe('FeedForward Class: Unit', () => {
                         {
                           filterWidth: 3,
                           filterHeight: 3,
+                          filterCount: 16,
                           padding: 2,
                           stride: 2,
                         },

src/layer/target.test.ts

@@ -57,7 +57,10 @@ describe('Target Layer', () => {
   });
 
   test('uses compare2D when width > 1', () => {
-    const target = new Target({}, mockLayer({ height: 10, width: 10 }));
+    const target = new Target(
+      { height: 10, width: 10 },
+      mockLayer({ height: 10, width: 10 })
+    );
     target.setupKernels();
     expect(makeKernel).toHaveBeenCalledWith(compare2D, {
       output: [10, 10],

src/likely.test.ts

@@ -3,8 +3,6 @@ import { NeuralNetwork } from './neural-network';
 
 /**
  * Return 0 or 1 for '#'
- * @param character
- * @returns {number}
  */
 function integer(character: string): number {
   if (character === '#') return 1;
@@ -13,8 +11,6 @@ function integer(character: string): number {
 
 /**
  * Turn the # into 1s and . into 0s. for whole string
- * @param string
- * @returns {Array}
  */
 function character(string: string): number[] {
   return string.trim().split('').map(integer);

src/likely.ts

@@ -1,28 +1,24 @@
-import { INumberHash } from './lookup';
-import { NeuralNetwork } from './neural-network';
+import { INeuralNetworkData, NeuralNetwork } from './neural-network';
 
-/**
- *
- * @param {*} input
- * @param {brain.NeuralNetwork} net
- * @returns {*}
- */
-export function likely<T extends number[] | Float32Array | INumberHash>(
-  input: T,
-  net: NeuralNetwork
-): T | null {
+export function likely<
+  InputType extends INeuralNetworkData,
+  OutputType extends INeuralNetworkData
+>(
+  input: InputType,
+  net: NeuralNetwork<InputType, OutputType>
+): OutputType | null {
   if (!net) {
     throw new TypeError(
       `Required parameter 'net' is of type ${typeof net}. Must be of type 'brain.NeuralNetwork'`
     );
   }
 
-  const output = net.run<T>(input);
+  const output = net.run(input);
   let maxProp = null;
   let maxValue = -1;
 
   Object.entries(output).forEach(([key, value]) => {
-    if (value > maxValue) {
+    if (typeof value !== 'undefined' && value > maxValue) {
       maxProp = key;
       maxValue = value;
     }

src/lookup.ts

@@ -10,7 +10,7 @@ export interface INumberArray {
   [index: number]: number;
 }
 
-export type InputOutputValue = INumberArray | INumberHash;
+export type InputOutputValue = INumberArray | Partial<INumberHash>;
 
 export interface ITrainingDatum {
   input: InputOutputValue | InputOutputValue[] | KernelOutput;

src/neural-network-gpu.test.ts

@@ -1,6 +1,7 @@
+import { Texture } from 'gpu.js';
+
 import { NeuralNetwork } from './neural-network';
 import { NeuralNetworkGPU } from './neural-network-gpu';
-import { Texture } from 'gpu.js';
 
 describe('NeuralNetworkGPU', () => {
   const xorTrainingData = [
@@ -33,11 +34,11 @@ describe('NeuralNetworkGPU', () => {
     });
 
     it('can serialize from NeuralNetworkGPU & deserialize to NeuralNetwork', () => {
-      const net = new NeuralNetworkGPU();
+      const net = new NeuralNetworkGPU<number[], number[]>();
       net.train(xorTrainingData, { iterations: 1 });
       const target = xorTrainingData.map((datum) => net.run(datum.input));
       const json = net.toJSON();
-      const net2 = new NeuralNetwork();
+      const net2 = new NeuralNetwork<number[], number[]>();
       net2.fromJSON(json);
       for (let i = 0; i < xorTrainingData.length; i++) {
         // there is a wee bit of loss going from GPU to CPU
@@ -49,11 +50,11 @@ describe('NeuralNetworkGPU', () => {
     });
 
     it('can serialize from NeuralNetwork & deserialize to NeuralNetworkGPU', () => {
-      const net = new NeuralNetwork();
+      const net = new NeuralNetwork<number[], number[]>();
       net.train(xorTrainingData, { iterations: 1 });
       const target = xorTrainingData.map((datum) => net.run(datum.input));
       const json = net.toJSON();
-      const net2 = new NeuralNetworkGPU();
+      const net2 = new NeuralNetworkGPU<number[], number[]>();
       net2.fromJSON(json);
       for (let i = 0; i < xorTrainingData.length; i++) {
         // there is a wee bit of loss going from CPU to GPU