Update models to deeplearnjs 0.5.0 (#769)

demos/imagenet/imagenet.ts

@@ -111,7 +111,7 @@ export class ImagenetDemo extends ImagenetDemoPolymer {
         imagenet_util.getRenderGrayscaleChannelsCollageShader(this.gpgpu);
 
     const cameraSetup = this.setupCameraInput();
-    this.squeezeNet = new SqueezeNet(this.math);
+    this.squeezeNet = new SqueezeNet();
 
     await Promise.all([this.squeezeNet.load(), cameraSetup]);
 

demos/package.json

@@ -53,8 +53,8 @@
     "@bower_components/webcomponentsjs": "webcomponents/webcomponentsjs#^0.7.24",
     "chart.js": "~2.7.1",
     "deeplearn": "file:../dist/",
-    "deeplearn-knn-image-classifier": "~0.2.3",
-    "deeplearn-squeezenet": "~0.1.9",
+    "deeplearn-knn-image-classifier": "~0.3.0",
+    "deeplearn-squeezenet": "~0.2.0",
     "tslint": "~5.8.0",
     "vue": "~2.5.9"
   },

demos/teachable_gaming/teachable_gaming.ts

@@ -203,8 +203,7 @@ export class TeachableGamingDemo extends TeachableGamingDemoPolymer {
           });
     }
     this.classifier = new KNNImageClassifier(
-        TeachableGamingDemo.maxControls, TeachableGamingDemo.knnKValue,
-        dl.ENV.math);
+        TeachableGamingDemo.maxControls, TeachableGamingDemo.knnKValue);
     this.classifier.load();
     this.predictedIndex = -1;
     this.selectedIndex = -1;

models/knn_image_classifier/demo.html

@@ -1,2 +1,44 @@
 <script src="https://unpkg.com/deeplearn"></script>
 <script src="dist/bundle.js"></script>
+
+<h2>Training images</h2>
+<img id="cat" src="images/cat.jpg"></img>
+<img id="dog1" src="images/dog1.jpg"></img>
+<h2>Input image</h2>
+<img id="dog2"  src="images/dog2.jpg"></img>
+<div id="result"></div>
+
+<script>
+  const cat = document.getElementById('cat');
+  const dog1 = document.getElementById('dog1');
+  const dog2 = document.getElementById('dog2');
+  const resultElement = document.getElementById('result');
+
+  window.onload = async () => {
+    resultElement.innerText = 'Loading classifier...';
+
+    // If dl isn't loaded, wait 1 second.
+    if (dl == null) {
+      await new Promise(resolve => setTimeout(resolve, 1000));
+    }
+
+    const knn = new knn_image_classifier.KNNImageClassifier(2, 1);
+    await knn.load();
+
+    resultElement.innerText = 'Training...';
+
+    const catPixels = dl.fromPixels(cat);
+    const dog1Pixels = dl.fromPixels(dog1);
+    const dog2Pixels = dl.fromPixels(dog2);
+
+    knn.addImage(catPixels, 0);
+    knn.addImage(dog1Pixels, 1);
+
+    resultElement.innerText = 'Predicting...';
+
+    const prediction = await knn.predictClass(dog2Pixels);
+
+    resultElement.innerText = 'Predicted to be class '+(prediction.classIndex+1);
+  };
+
+</script>

models/knn_image_classifier/knn_image_classifier.ts

@@ -15,22 +15,23 @@
  * =============================================================================
  */
 // tslint:disable-next-line:max-line-length
-import {Array1D, Array2D, Array3D, Model, NDArrayMath, Scalar} from 'deeplearn';
+import * as dl from 'deeplearn';
+import {Tensor1D, Tensor2D, Tensor3D} from 'deeplearn';
 import {SqueezeNet} from 'deeplearn-squeezenet';
 import * as model_util from '../util';
 
-export class KNNImageClassifier implements Model {
+export class KNNImageClassifier {
   private squeezeNet: SqueezeNet;
 
   // A concatenated matrix of all class logits matrices, lazily created and
   // used during prediction.
-  private trainLogitsMatrix: Array2D;
+  private trainLogitsMatrix: Tensor2D;
 
-  private classLogitsMatrices: Array2D[] = [];
+  private classLogitsMatrices: Tensor2D[] = [];
   private classExampleCount: number[] = [];
 
   private varsLoaded = false;
-  private squashLogitsDenominator = Scalar.new(300);
+  private squashLogitsDenominator = dl.scalar(300);
 
   /**
    * Contructor for the class.
@@ -39,15 +40,13 @@ export class KNNImageClassifier implements Model {
    * @param k The number of nearest neighbors to look at when predicting.
    * @param math A math implementation for performing the calculations.
    */
-  constructor(
-      private numClasses: number, private k: number,
-      private math: NDArrayMath) {
+  constructor(private numClasses: number, private k: number) {
     for (let i = 0; i < this.numClasses; i++) {
       this.classLogitsMatrices.push(null);
       this.classExampleCount.push(0);
     }
 
-    this.squeezeNet = new SqueezeNet(this.math);
+    this.squeezeNet = new SqueezeNet();
   }
 
   /**
@@ -75,7 +74,7 @@ export class KNNImageClassifier implements Model {
   /**
    * Adds the provided image to the specified class.
    */
-  addImage(image: Array3D, classIndex: number): void {
+  addImage(image: Tensor3D, classIndex: number): void {
     if (!this.varsLoaded) {
       console.warn('Cannot add images until vars have been loaded.');
       return;
@@ -85,7 +84,7 @@ export class KNNImageClassifier implements Model {
     }
     this.clearTrainLogitsMatrix();
 
-    this.math.scope((keep, track) => {
+    dl.tidy(() => {
       // Add the squeezenet logits for the image to the appropriate class
       // logits matrix.
       const logits = this.squeezeNet.predict(image);
@@ -95,14 +94,16 @@ export class KNNImageClassifier implements Model {
       if (this.classLogitsMatrices[classIndex] == null) {
         this.classLogitsMatrices[classIndex] = imageLogits.as2D(1, logitsSize);
       } else {
-        const newTrainLogitsMatrix = this.math.concat2D(
-            this.classLogitsMatrices[classIndex].as2D(
-                this.classExampleCount[classIndex], logitsSize),
-            imageLogits.as2D(1, logitsSize), 0);
+        const newTrainLogitsMatrix =
+            this.classLogitsMatrices[classIndex]
+                .as2D(this.classExampleCount[classIndex], logitsSize)
+                .concat(imageLogits.as2D(1, logitsSize), 0);
+
         this.classLogitsMatrices[classIndex].dispose();
         this.classLogitsMatrices[classIndex] = newTrainLogitsMatrix;
       }
-      keep(this.classLogitsMatrices[classIndex]);
+
+      dl.keep(this.classLogitsMatrices[classIndex]);
 
       this.classExampleCount[classIndex]++;
     });
@@ -123,12 +124,12 @@ export class KNNImageClassifier implements Model {
    * @param image The input image.
    * @returns cosine distances for each entry in the database.
    */
-  predict(image: Array3D): Array1D {
+  predict(image: Tensor3D): Tensor1D {
     if (!this.varsLoaded) {
       throw new Error('Cannot predict until vars have been loaded.');
     }
 
-    return this.math.scope((keep) => {
+    return dl.tidy(() => {
       const logits = this.squeezeNet.predict(image);
       const imageLogits = this.normalizeVector(logits);
       const logitsSize = imageLogits.shape[0];
@@ -149,13 +150,11 @@ export class KNNImageClassifier implements Model {
         return null;
       }
 
-      keep(this.trainLogitsMatrix);
+      dl.keep(this.trainLogitsMatrix);
 
       const numExamples = this.getNumExamples();
-      return this.math
-          .matMul(
-              this.trainLogitsMatrix.as2D(numExamples, logitsSize),
-              imageLogits.as2D(logitsSize, 1))
+      return this.trainLogitsMatrix.as2D(numExamples, logitsSize)
+          .matMul(imageLogits.as2D(logitsSize, 1))
           .as1D();
     });
   }
@@ -168,7 +167,7 @@ export class KNNImageClassifier implements Model {
    * @returns A dict of the top class for the image and an array of confidence
    * values for all possible classes.
    */
-  async predictClass(image: Array3D):
+  async predictClass(image: Tensor3D):
       Promise<{classIndex: number, confidences: number[]}> {
     let imageClass = -1;
     const confidences = new Array<number>(this.numClasses);
@@ -179,7 +178,7 @@ export class KNNImageClassifier implements Model {
     const knn = this.predict(image).asType('float32');
     const numExamples = this.getNumExamples();
     const kVal = Math.min(this.k, numExamples);
-    const topK = model_util.topK(await knn.data(), kVal);
+    const topK = model_util.topK(await knn.data() as Float32Array, kVal);
     knn.dispose();
     const topKIndices = topK.indices;
 
@@ -236,32 +235,30 @@ export class KNNImageClassifier implements Model {
     }
   }
 
-  private concatWithNulls(ndarray1: Array2D, ndarray2: Array2D): Array2D {
+  private concatWithNulls(ndarray1: Tensor2D, ndarray2: Tensor2D): Tensor2D {
     if (ndarray1 == null && ndarray2 == null) {
       return null;
     }
     if (ndarray1 == null) {
-      return this.math.clone(ndarray2);
+      return ndarray2.clone();
     } else if (ndarray2 === null) {
-      return this.math.clone(ndarray1);
+      return ndarray1.clone();
     }
-    return this.math.concat2D(ndarray1, ndarray2, 0);
+    return ndarray1.concat(ndarray2, 0);
   }
 
   /**
    * Normalize the provided vector to unit length.
    */
-  private normalizeVector(vec: Array1D) {
+  private normalizeVector(vec: Tensor1D) {
     // This hack is here for numerical stability on devices without floating
     // point textures. We divide by a constant so that the sum doesn't overflow
     // our fixed point precision. Remove this once we use floating point
     // intermediates with proper dynamic range quantization.
-    const squashedVec = this.math.divide(vec, this.squashLogitsDenominator);
+    const squashedVec = dl.div(vec, this.squashLogitsDenominator);
+    const sqrtSum = squashedVec.square().sum().sqrt();
 
-    const squared = this.math.multiplyStrict(squashedVec, squashedVec);
-    const sum = this.math.sum(squared);
-    const sqrtSum = this.math.sqrt(sum);
-    return this.math.divide(squashedVec, sqrtSum);
+    return dl.div(squashedVec, sqrtSum);
   }
 
   private getNumExamples() {

models/knn_image_classifier/package.json

@@ -1,15 +1,15 @@
 {
   "name": "deeplearn-knn-image-classifier",
-  "version": "0.2.3",
+  "version": "0.3.0",
   "description": "A KNN Image Classifier model in deeplearn.js",
   "main": "dist/knn_image_classifier/index.js",
   "unpkg": "dist/bundle.js",
   "types": "dist/knn_image_classifier/index.d.ts",
   "peerDependencies": {
-    "deeplearn": "0.4.0"
+    "deeplearn": "0.5.0"
   },
   "dependencies": {
-    "deeplearn-squeezenet": "~0.1.9"
+    "deeplearn-squeezenet": "~0.2.0"
   },
   "repository": {
     "type": "git",
@@ -18,7 +18,7 @@
   "devDependencies": {
     "browserify": "~14.5.0",
     "browserify-shim": "~3.8.14",
-    "deeplearn": "0.4.0",
+    "deeplearn": "0.5.0",
     "mkdirp": "~0.5.1",
     "tsify": "~3.0.3",
     "tslint": "~5.8.0",

models/knn_image_classifier/yarn.lock

@@ -564,11 +564,12 @@ deeplearn-squeezenet@~0.1.9:
   version "0.1.9"
   resolved "https://registry.yarnpkg.com/deeplearn-squeezenet/-/deeplearn-squeezenet-0.1.9.tgz#a4194c31156cd2fa7421aaffcc6e3d4b3e3ef1c1"
 
-deeplearn@0.4.0:
-  version "0.4.0"
-  resolved "https://registry.yarnpkg.com/deeplearn/-/deeplearn-0.4.0.tgz#3181789795b28c1c8d0311fb2209cea8dfe5ad67"
+deeplearn@0.5.0:
+  version "0.5.0"
+  resolved "https://registry.yarnpkg.com/deeplearn/-/deeplearn-0.5.0.tgz#015c25fb35fcbf40cf9f31d9fff7760af7a436ec"
   dependencies:
     seedrandom "~2.4.3"
+    utf8 "~2.1.2"
 
 defined@^1.0.0:
   version "1.0.0"
@@ -1993,6 +1994,10 @@ url@~0.11.0:
     punycode "1.3.2"
     querystring "0.2.0"
 
+utf8@~2.1.2:
+  version "2.1.2"
+  resolved "https://registry.yarnpkg.com/utf8/-/utf8-2.1.2.tgz#1fa0d9270e9be850d9b05027f63519bf46457d96"
+
 util-deprecate@~1.0.1:
   version "1.0.2"
   resolved "https://registry.yarnpkg.com/util-deprecate/-/util-deprecate-1.0.2.tgz#450d4dc9fa70de732762fbd2d4a28981419a0ccf"

models/squeezenet/demo.html

@@ -21,7 +21,7 @@
 
     resultElement.innerText = 'Predicting...';
 
-    const pixels = dl.Array3D.fromPixels(cat);
+    const pixels = dl.fromPixels(cat);
 
     const result = squeezeNet.predict(pixels);
 

models/squeezenet/package.json

@@ -1,12 +1,12 @@
 {
   "name": "deeplearn-squeezenet",
-  "version": "0.1.9",
+  "version": "0.2.0",
   "description": "Pretrained SqueezeNet model in deeplearn.js",
   "main": "dist/squeezenet/index.js",
   "unpkg": "dist/bundle.js",
   "types": "dist/squeezenet/index.d.ts",
   "peerDependencies": {
-    "deeplearn": "0.4.0"
+    "deeplearn": "0.5.0"
   },
   "repository": {
     "type": "git",
@@ -15,7 +15,7 @@
   "devDependencies": {
     "browserify": "~14.5.0",
     "browserify-shim": "~3.8.14",
-    "deeplearn": "0.4.0",
+    "deeplearn": "0.5.0",
     "mkdirp": "~0.5.1",
     "tsify": "~3.0.3",
     "tslint": "~5.8.0",