Add 'OneHot' op. (#36)

* Add relu op.

* Add multiply.

* Change type op attr helper method to dtype.

* Cleanup.

* basic math.

* use expectArraysClose()

* Add reverse()

* Add neg - add comment about supporting concat.

* sum()

* More ops.

* Fixup concat()

* wip

* Some cleanup in the backend.

* min and minimum()

* Add basic pow and TODO for handling tensor upcasting of types.

* More single-input ops.

* wip

* cleanup

Author: nkreeger

src/nodejs_kernel_backend.ts

@@ -127,11 +127,8 @@ export class NodeJSKernelBackend implements KernelBackend {
 
   slice<T extends Tensor<Rank>>(x: T, begin: number[], size: number[]): T {
     const opAttrs = [
-      this.createTypeOpAttr('T', x.dtype), {
-        name: 'Index',
-        type: this.binding.TF_ATTR_TYPE,
-        value: this.binding.TF_INT32
-      }
+      this.createTypeOpAttr('T', x.dtype),
+      this.createTypeOpAttr('Index', 'int32')
     ];
 
     // Bind tensor values
@@ -570,18 +567,15 @@ export class NodeJSKernelBackend implements KernelBackend {
   }
   pad<T extends Tensor<Rank>>(
       x: T, paddings: Array<[number, number]>, constantValue: number): T {
-    const opAttrs = [
-      this.createTypeOpAttr('T', x.dtype), {
-        name: 'Tpaddings',
-        type: this.binding.TF_ATTR_TYPE,
-        value: this.binding.TF_INT32
-      }
-    ];
-
     // Bind tensor values
     const paddingsTensor = tensor2d(paddings, [2, 2], 'int32');
     const constantTensor = scalar(constantValue, x.dtype);
 
+    const opAttrs = [
+      this.createTypeOpAttr('T', x.dtype),
+      this.createTypeOpAttr('Tpaddings', paddingsTensor.dtype)
+    ];
+
     return this.execute(
                'PadV2', opAttrs, [x, paddingsTensor, constantTensor]) as T;
   }
@@ -613,7 +607,19 @@ export class NodeJSKernelBackend implements KernelBackend {
   }
   oneHot(indices: Tensor1D, depth: number, onValue: number, offValue: number):
       Tensor2D {
-    throw new Error('Method not implemented.');
+    const depthTensor = scalar(depth, 'int32');
+    const onValueTensor = scalar(onValue, 'int32');
+    const offValueTensor = scalar(offValue, 'int32');
+
+    const opAttrs = [
+      {name: 'axis', type: this.binding.TF_ATTR_INT, value: -1},
+      this.createTypeOpAttr('T', indices.dtype),
+      this.createTypeOpAttr('TI', indices.dtype)
+    ];
+
+    return this.execute('OneHot', opAttrs, [
+      indices, depthTensor, onValueTensor, offValueTensor
+    ]) as Tensor2D;
   }
   dispose(): void {
     throw new Error('Method not implemented.');

src/nodejs_kernel_backend_test.ts

@@ -489,3 +489,13 @@ describe('tanh', () => {
     expectArraysClose(result, expected);
   });
 });
+
+describe('oneHot', () => {
+  it('should work', () => {
+    const indices = dl.tensor1d([0, 1], 'int32');
+    const res = dl.oneHot(indices, 2);
+
+    expect(res.shape).toEqual([2, 2]);
+    expectArraysClose(res, [1, 0, 0, 1]);
+  });
+});