Enhance InputLayer to support default values and update tests (#948)

Author: ishii-norimi

lib/model/nns/layer/index.js

@@ -187,7 +187,7 @@ export { default as VarLayer } from './variance.js'
  * { type: 'huber' } |
  * { type: 'identity' } |
  * { type: 'include', net: NeuralNetwork | object[], input_to?: string, train?: boolean } |
- * { type: 'input', name?: string, size?: number[] } |
+ * { type: 'input', name?: string, size?: number[], value?: number | number[] | number[][] | nunber[][][] | number[][][][] | Matrix | Tensor } |
  * { type: 'is_inf' } |
  * { type: 'is_nan' } |
  * { type: 'isigmoid', a?: number, alpha?: number } |

lib/model/nns/layer/input.js

@@ -10,11 +10,13 @@ export default class InputLayer extends Layer {
 	 * @param {object} config object
 	 * @param {string} [config.name] Name of the layer
 	 * @param {number[]} [config.size] Size of the layer
+	 * @param {number | number[] | number[][] | nunber[][][] | number[][][][] | Matrix | Tensor} [config.value] Default value
 	 */
-	constructor({ name = null, size = null, ...rest }) {
+	constructor({ name = null, size = null, value, ...rest }) {
 		super(rest)
 		this._name = name
 		this._size = size
+		this._value = value
 	}
 
 	bind({ input }) {
@@ -27,6 +29,9 @@ export default class InputLayer extends Layer {
 		) {
 			input = input[this._name]
 		}
+		if (input == null) {
+			input = this._value
+		}
 		if (Array.isArray(input)) {
 			this._o = Tensor.fromArray(input)
 			if (this._o.dimension === 2) {
@@ -57,6 +62,7 @@ export default class InputLayer extends Layer {
 			type: 'input',
 			name: this._name,
 			size: this._size?.concat(),
+			value: this._value instanceof Matrix || this._value instanceof Tensor ? this._value.toArray() : this._value,
 		}
 	}
 }

tests/lib/model/nns/layer/input.test.js

@@ -101,6 +101,26 @@ describe('layer', () => {
 				}
 			}
 		})
+
+		test('default value', () => {
+			const x = 1
+			const layer = new InputLayer({ value: x })
+
+			layer.bind({})
+			const y = layer.calc()
+			expect(y.sizes).toEqual([1, 1])
+			expect(y.at(0, 0)).toBe(1)
+		})
+
+		test('default value and bind value', () => {
+			const x = 1
+			const layer = new InputLayer({ value: x })
+
+			layer.bind({ input: 2 })
+			const y = layer.calc()
+			expect(y.sizes).toEqual([1, 1])
+			expect(y.at(0, 0)).toBe(2)
+		})
 	})
 
 	describe('grad', () => {
@@ -135,6 +155,14 @@ describe('layer', () => {
 			const obj = layer.toObject()
 			expect(obj).toEqual({ type: 'input', name: 'in', size: [null, 10] })
 		})
+
+		test('matrix value', () => {
+			const mat = Matrix.randn(1, 10)
+			const layer = new InputLayer({ name: 'in', size: [null, 10], value: mat })
+
+			const obj = layer.toObject()
+			expect(obj).toEqual({ type: 'input', name: 'in', size: [null, 10], value: mat.toArray() })
+		})
 	})
 
 	test('fromObject', () => {