Fix overlapFactor bug and cleanup a bit (#97)

Cleanup the speech commands model:

- Fix an issue with `overlapFactor`, where if the user passed an overlap close to 1, the callback `f` in `startStream(f)` would still be called only once every 1s. Now that callback can be called as frequently as the frame duration (tested in the work-in-progress code lab). This makes the demo run in real-time, which is really nice.
- Remove `columnBufferLength` and `columnHopLength` as extractor config params since they were used to indirectly talk about `overlapFactor` between the recognizer and extractor. Thus, add `overlapFactor` as config instead.
- Change `DEFAULT_SUPPRESSION_TIME_MILLIS` from `1000ms` to `0ms`, so the user can get instant predictions.
- Remove unused params `filterSize` from `RecognizerParams`, and `minSamples` from `StreamingRecognitionConfig`
- Move top-level `utils/` folder to `training/utils` since only training uses it
- Remove `utils/util.ts` (all the util functions there were unused), other than `powerOfTwo` which got moved to `audio_util.ts`
- Make tests (`*_test.ts`) visible to the typescript compiler and linter.
- Improve `tslint.json`
- s/SpeechCommands/speechCommands/ in Rollup to align with the other models.
- minimize use of `any`

Author: dsmilkov

speech-commands/README.md

@@ -103,7 +103,7 @@ As the example above shows, you can specify optional parameters when calling
 `startStreaming()`. The supported parameters are:
 
 * `overlapFactor`: Controls how often the recognizer performs prediction on
-  spectrograms. Must be a number between 0 and 1 (default: 0.5). For example,
+  spectrograms. Must be >=0 and <1 (default: 0.5). For example,
   if each spectrogram is 1000 ms long and `overlapFactor` is set to 0.25,
   the prediction will happen every 250 ms.
 * `includeSpectrogram`: Let the callback function be invoked with the

speech-commands/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tensorflow-models/speech-commands",
-  "version": "0.1.3",
+  "version": "0.1.4",
   "description": "Speech-command recognizer in TensorFlow.js",
   "main": "dist/index.js",
   "unpkg": "dist/speech-commands.min.js",
@@ -19,28 +19,32 @@
     "@tensorflow/tfjs": "^0.13.1",
     "@tensorflow/tfjs-node": "^0.1.18",
     "@types/jasmine": "~2.8.8",
+    "@types/rimraf": "^2.0.2",
+    "@types/tempfile": "^2.0.0",
     "babel-core": "~6.26.0",
     "babel-plugin-transform-runtime": "~6.23.0",
+    "clang-format": "^1.2.4",
     "dct": "^0.0.3",
     "jasmine": "^3.2.0",
     "jasmine-core": "^3.2.1",
     "kissfft-js": "^0.1.8",
     "rimraf": "2.6.2",
     "rollup": "~0.60.7",
     "rollup-plugin-node-resolve": "~3.3.0",
-    "rollup-plugin-typescript2": "~0.15.0",
+    "rollup-plugin-typescript2": "~0.13.0",
     "rollup-plugin-uglify": "~3.0.0",
     "tempfile": "2.0.0",
     "ts-node": "~5.0.0",
     "tslint": "~5.10.0",
+    "tslint-no-circular-imports": "^0.6.1",
     "typescript": "2.9.2",
     "yalc": "~1.0.0-pre.21"
   },
   "scripts": {
-    "build": "rimraf dist && tsc && rollup -c",
+    "build": "tsc",
     "lint": "tslint -p . -t verbose",
-    "publish-local": "yarn build && yalc push",
-    "test": "yarn build && ts-node run_tests.ts"
+    "publish-local": "rimraf dist && yarn build && rollup -c && yalc push",
+    "test": "ts-node run_tests.ts"
   },
   "license": "Apache-2.0"
 }

speech-commands/rollup.config.js

@@ -19,8 +19,22 @@ import node from 'rollup-plugin-node-resolve';
 import typescript from 'rollup-plugin-typescript2';
 import uglify from 'rollup-plugin-uglify';
 
-const PREAMBLE =
-  `// @tensorflow/tfjs-models Copyright ${(new Date).getFullYear()} Google`;
+const PREAMBLE = `/**
+ * @license
+ * Copyright ${(new Date).getFullYear()} Google LLC. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */`;
 
 function minify() {
   return uglify({
@@ -48,6 +62,7 @@ function config({
     ],
     output: {
       banner: PREAMBLE,
+      sourcemap: true,
       globals: {
         '@tensorflow/tfjs': 'tf'
       },
@@ -61,15 +76,15 @@ export default [
   config({
     output: {
       format: 'umd',
-      name: 'SpeechCommands',
+      name: 'speechCommands',
       file: 'dist/speech-commands.js'
     }
   }),
   config({
     plugins: [minify()],
     output: {
       format: 'umd',
-      name: 'SpeechCommands',
+      name: 'speechCommands',
       file: 'dist/speech-commands.min.js'
     }
   }),

speech-commands/src/browser_fft_extractor.ts

@@ -20,8 +20,6 @@
  */
 
 import * as tf from '@tensorflow/tfjs';
-
-// tslint:disable-next-line:max-line-length
 import {getAudioContextConstructor, getAudioMediaStream, normalize} from './browser_fft_utils';
 import {FeatureExtractor, RecognizerParams} from './types';
 
@@ -61,6 +59,15 @@ export interface BrowserFftFeatureExtractorConfig extends RecognizerParams {
    * If `null` or `undefined`, will do no truncation.
    */
   columnTruncateLength?: number;
+
+  /**
+   * Overlap factor. Must be >=0 and <1.
+   * For example, if the model takes a frame length of 1000 ms,
+   * and if overlap factor is 0.4, there will be a 400ms
+   * overlap between two successive frames, i.e., frames
+   * will be taken every 600 ms.
+   */
+  overlapFactor: number;
 }
 
 /**
@@ -142,15 +149,12 @@ export class BrowserFftFeatureExtractor implements FeatureExtractor {
     this.fftSize = config.fftSize || 1024;
     this.frameDurationMillis = this.fftSize / this.sampleRateHz * 1e3;
     this.columnTruncateLength = config.columnTruncateLength || this.fftSize;
-    const columnBufferLength = config.columnBufferLength || this.fftSize;
-    const columnHopLength = config.columnHopLength || (this.fftSize / 2);
-    this.overlapFactor = columnHopLength / columnBufferLength;
+    this.overlapFactor = config.overlapFactor;
 
-    if (!(this.overlapFactor > 0)) {
-      throw new Error(
-          `Invalid overlapFactor: ${this.overlapFactor}. ` +
-          `Check your columnBufferLength and columnHopLength.`);
-    }
+    tf.util.assert(
+        this.overlapFactor >= 0 && this.overlapFactor < 1,
+        `Expected overlapFactor to be >= 0 and < 1, ` +
+            `but got ${this.overlapFactor}`);
 
     if (this.columnTruncateLength > this.fftSize) {
       throw new Error(
@@ -190,8 +194,10 @@ export class BrowserFftFeatureExtractor implements FeatureExtractor {
 
     this.frameCount = 0;
 
+    const period = Math.max(
+        1, Math.round(this.numFramesPerSpectrogram * (1 - this.overlapFactor)));
     this.tracker = new Tracker(
-        Math.round(this.numFramesPerSpectrogram * this.overlapFactor),
+        period,
         Math.round(this.suppressionTimeMillis / this.frameDurationMillis));
     this.frameIntervalTask = setInterval(
         this.onAudioFrame.bind(this), this.fftSize / this.sampleRateHz * 1e3);
@@ -200,7 +206,6 @@ export class BrowserFftFeatureExtractor implements FeatureExtractor {
   private async onAudioFrame() {
     this.analyser.getFloatFrequencyData(this.freqData);
     if (this.freqData[0] === -Infinity) {
-      console.warn(`No signal (frame #${this.frameCount})`);
       return;
     }