Fix truncated normal and expose test_util to public API

package.json

@@ -1,6 +1,6 @@
 {
   "name": "deeplearn",
-  "version": "0.2.7",
+  "version": "0.2.8",
   "description": "Hardware-accelerated JavaScript library for machine intelligence",
   "private": false,
   "main": "dist/src/index.js",

src/index.ts

@@ -20,6 +20,7 @@ import * as conv_util from './math/conv_util';
 import * as gpgpu_util from './math/webgl/gpgpu_util';
 import * as render_ndarray_gpu_util from './math/webgl/render_ndarray_gpu_util';
 import * as webgl_util from './math/webgl/webgl_util';
+import * as test_util from './test_util';
 import * as util from './util';
 
 export {CheckpointLoader} from './data/checkpoint_loader';
@@ -29,9 +30,9 @@ export {InCPUMemoryShuffledInputProviderBuilder, InGPUMemoryShuffledInputProvide
 export {XhrDataset, XhrDatasetConfig, XhrModelConfig} from './data/xhr-dataset';
 export {ENV, Features} from './environment';
 export {Graph, Tensor} from './graph/graph';
+export {AdadeltaOptimizer} from './graph/optimizers/adadelta_optimizer';
 export {AdagradOptimizer} from './graph/optimizers/adagrad_optimizer';
 export {MomentumOptimizer} from './graph/optimizers/momentum_optimizer';
-export {AdadeltaOptimizer} from './graph/optimizers/adadelta_optimizer';
 export {Optimizer} from './graph/optimizers/optimizer';
 export {RMSPropOptimizer} from './graph/optimizers/rmsprop_optimizer';
 export {SGDOptimizer} from './graph/optimizers/sgd_optimizer';
@@ -51,6 +52,7 @@ export {
   conv_util,
   gpgpu_util,
   render_ndarray_gpu_util,
+  test_util,
   util,
   webgl_util,
   xhr_dataset

src/util.ts

@@ -15,8 +15,8 @@
  * =============================================================================
  */
 
-export type Vector = number[] | Float64Array | Float32Array | Int32Array |
-    Int8Array | Int16Array;
+export type Vector =
+    number[]|Float64Array|Float32Array|Int32Array|Int8Array|Int16Array;
 
 /** Shuffles the array using Fisher-Yates algorithm. */
 // tslint:disable-next-line:no-any
@@ -63,7 +63,7 @@ export function randGauss(mean = 0, stdDev = 1, truncated = false): number {
   } while (s > 1);
 
   const result = Math.sqrt(-2 * Math.log(s) / s) * v1;
-  if (truncated && result > 2) {
+  if (truncated && Math.abs(result) > 2) {
     return randGauss(mean, stdDev, true);
   }
   return mean + stdDev * result;
@@ -106,12 +106,7 @@ export function flatten(arr: any[], ret?: number[]): number[] {
 }
 
 export type ArrayData =
-    Float32Array |
-        number |
-        number[] |
-        number[][] |
-        number[][][] |
-        number[][][][];
+    Float32Array|number|number[]|number[][]|number[][][]|number[][][][];
 
 export function inferShape(arr: ArrayData): number[] {
   const shape: number[] = [];

src/util_test.ts

@@ -218,3 +218,29 @@ describe('util.inferFromImplicitShape', () => {
     expect(() => util.inferFromImplicitShape([2, 3, 4], 25)).toThrowError();
   });
 });
+
+describe('util.randGauss', () => {
+  it('standard normal', () => {
+    const a = util.randGauss();
+    expect(a != null);
+  });
+
+  it('truncated standard normal', () => {
+    const numSamples = 1000;
+    for (let i = 0; i < numSamples; ++i) {
+      const sample = util.randGauss(0, 1, true);
+      expect(Math.abs(sample) <= 2);
+    }
+  });
+
+  it('truncated normal, mu = 3, std=4', () => {
+    const numSamples = 1000;
+    const mean = 3;
+    const stdDev = 4;
+    for (let i = 0; i < numSamples; ++i) {
+      const sample = util.randGauss(mean, stdDev, true);
+      const normalizedSample = (sample - mean) / stdDev;
+      expect(Math.abs(normalizedSample) <= 2);
+    }
+  });
+});