@tensorflow/tfjs-core - npm Package Compare versions

Comparing version 0.11.3 to 0.11.4

dist-es6/io/local_storage_test.js

@@ -0,1 +1,37 @@
+var __awaiter = (this && this.__awaiter) || function (thisArg, _arguments, P, generator) {
+    return new (P || (P = Promise))(function (resolve, reject) {
+        function fulfilled(value) { try { step(generator.next(value)); } catch (e) { reject(e); } }
+        function rejected(value) { try { step(generator["throw"](value)); } catch (e) { reject(e); } }
+        function step(result) { result.done ? resolve(result.value) : new P(function (resolve) { resolve(result.value); }).then(fulfilled, rejected); }
+        step((generator = generator.apply(thisArg, _arguments || [])).next());
+    });
+};
+var __generator = (this && this.__generator) || function (thisArg, body) {
+    var _ = { label: 0, sent: function() { if (t[0] & 1) throw t[1]; return t[1]; }, trys: [], ops: [] }, f, y, t, g;
+    return g = { next: verb(0), "throw": verb(1), "return": verb(2) }, typeof Symbol === "function" && (g[Symbol.iterator] = function() { return this; }), g;
+    function verb(n) { return function (v) { return step([n, v]); }; }
+    function step(op) {
+        if (f) throw new TypeError("Generator is already executing.");
+        while (_) try {
+            if (f = 1, y && (t = y[op[0] & 2 ? "return" : op[0] ? "throw" : "next"]) && !(t = t.call(y, op[1])).done) return t;
+            if (y = 0, t) op = [0, t.value];
+            switch (op[0]) {
+                case 0: case 1: t = op; break;
+                case 4: _.label++; return { value: op[1], done: false };
+                case 5: _.label++; y = op[1]; op = [0]; continue;
+                case 7: op = _.ops.pop(); _.trys.pop(); continue;
+                default:
+                    if (!(t = _.trys, t = t.length > 0 && t[t.length - 1]) && (op[0] === 6 || op[0] === 2)) { _ = 0; continue; }
+                    if (op[0] === 3 && (!t || (op[1] > t[0] && op[1] < t[3]))) { _.label = op[1]; break; }
+                    if (op[0] === 6 && _.label < t[1]) { _.label = t[1]; t = op; break; }
+                    if (t && _.label < t[2]) { _.label = t[2]; _.ops.push(op); break; }
+                    if (t[2]) _.ops.pop();
+                    _.trys.pop(); continue;
+            }
+            op = body.call(thisArg, _);
+        } catch (e) { op = [6, e]; y = 0; } finally { f = t = 0; }
+        if (op[0] & 5) throw op[1]; return { value: op[0] ? op[1] : void 0, done: true };
+    }
+};
+var _this = this;
 import * as tf from '../index';
@@ -118,22 +154,22 @@ import { describeWithFlags } from '../jasmine_util';
     });
-    it('Save-load round trip succeeds', function (done) {
-        var handler1 = tf.io.getSaveHandlers('localstorage://FooModel')[0];
-        handler1.save(artifacts1)
-            .then(function (saveResult) {
-            var handler2 = tf.io.getLoadHandlers('localstorage://FooModel')[0];
-            handler2.load()
-                .then(function (loaded) {
-                expect(loaded.modelTopology).toEqual(modelTopology1);
-                expect(loaded.weightSpecs).toEqual(weightSpecs1);
-                expect(loaded.weightData).toEqual(weightData1);
-                done();
-            })
-                .catch(function (err) {
-                console.error(err.stack);
-            });
-        })
-            .catch(function (err) {
-            console.error(err.stack);
+    it('Save-load round trip succeeds', function () { return __awaiter(_this, void 0, void 0, function () {
+        var handler1, handler2, loaded;
+        return __generator(this, function (_a) {
+            switch (_a.label) {
+                case 0:
+                    handler1 = tf.io.getSaveHandlers('localstorage://FooModel')[0];
+                    return [4, handler1.save(artifacts1)];
+                case 1:
+                    _a.sent();
+                    handler2 = tf.io.getLoadHandlers('localstorage://FooModel')[0];
+                    return [4, handler2.load()];
+                case 2:
+                    loaded = _a.sent();
+                    expect(loaded.modelTopology).toEqual(modelTopology1);
+                    expect(loaded.weightSpecs).toEqual(weightSpecs1);
+                    expect(loaded.weightData).toEqual(weightData1);
+                    return [2];
+            }
         });
-    });
+    }); });
     it('Loading nonexistent model fails.', function (done) {
@@ -140,0 +176,0 @@ var handler = tf.io.getSaveHandlers('localstorage://NonexistentModel')[0];

dist-es6/kernels/backend_cpu.d.ts

@@ -95,2 +95,4 @@ import { Conv2DInfo } from '../ops/conv_util';
     depthwiseConv2D(x: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerInput(dy: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerFilter(x: Tensor4D, dy: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
     tile<T extends Tensor>(x: T, reps: number[]): T;
@@ -97,0 +99,0 @@ pad<T extends Tensor>(x: T, paddings: Array<[number, number]>, constantValue: number): T;

dist-es6/kernels/backend_webgl.d.ts

@@ -130,2 +130,4 @@ import { TimingInfo } from '../engine';
     depthwiseConv2D(x: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerInput(dy: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerFilter(x: Tensor4D, dy: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
     maxPool(x: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
@@ -132,0 +134,0 @@ avgPool(x: Tensor4D, convInfo: Conv2DInfo): Tensor4D;

dist-es6/kernels/backend_webgl.js

@@ -53,2 +53,3 @@ var __awaiter = (this && this.__awaiter) || function (thisArg, _arguments, P, generator) {
 import { Conv2DDerFilterProgram, Conv2DDerInputProgram } from './webgl/conv_backprop_gpu';
+import { DepthwiseConv2DDerFilterProgram, DepthwiseConv2DDerInputProgram } from './webgl/conv_backprop_gpu_depthwise';
 import { Conv2DProgram } from './webgl/conv_gpu';
@@ -730,2 +731,10 @@ import { DepthwiseConv2DProgram } from './webgl/conv_gpu_depthwise';
     };
+    MathBackendWebGL.prototype.depthwiseConv2DDerInput = function (dy, filter, convInfo) {
+        var program = new DepthwiseConv2DDerInputProgram(convInfo);
+        return this.compileAndRun(program, [dy, filter]);
+    };
+    MathBackendWebGL.prototype.depthwiseConv2DDerFilter = function (x, dy, convInfo) {
+        var program = new DepthwiseConv2DDerFilterProgram(convInfo);
+        return this.compileAndRun(program, [x, dy]);
+    };
     MathBackendWebGL.prototype.maxPool = function (x, convInfo) {
@@ -732,0 +741,0 @@ var program = new Pool2DProgram(convInfo, 'max', false);

dist-es6/kernels/backend.d.ts

@@ -94,2 +94,4 @@ import { Conv2DInfo } from '../ops/conv_util';
     depthwiseConv2D(input: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerInput(dy: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerFilter(x: Tensor4D, dY: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
     maxPool(x: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
@@ -96,0 +98,0 @@ maxPoolBackprop(dy: Tensor4D, x: Tensor4D, y: Tensor4D, convInfo: Conv2DInfo): Tensor4D;

dist-es6/ops/conv.js

@@ -165,7 +165,7 @@ var __decorate = (this && this.__decorate) || function (decorators, target, key, desc) {
         }
-        util.assert(x4D.rank === 4, "Error in depthwiseConv2D: input must be rank 4, but got " +
+        util.assert(x4D.rank === 4, "Error in depthwiseConv2d: input must be rank 4, but got " +
             ("rank " + x4D.rank + "."));
-        util.assert(filter.rank === 4, "Error in depthwiseConv2D: filter must be rank 4, but got rank " +
+        util.assert(filter.rank === 4, "Error in depthwiseConv2d: filter must be rank 4, but got rank " +
             (filter.rank + "."));
-        util.assert(x4D.shape[3] === filter.shape[2], "Error in depthwiseConv2D: number of input channels " +
+        util.assert(x4D.shape[3] === filter.shape[2], "Error in depthwiseConv2d: number of input channels " +
             ("(" + x4D.shape[3] + ") must match the inChannels dimension in ") +
@@ -179,7 +179,15 @@ ("filter " + filter.shape[2] + "."));
         if (dimRoundingMode != null) {
-            util.assert(util.isInt(pad), "Error in depthwiseConv2D: pad must be an integer when using, " +
+            util.assert(util.isInt(pad), "Error in depthwiseConv2d: pad must be an integer when using, " +
                 ("dimRoundingMode " + dimRoundingMode + " but got pad " + pad + "."));
         }
         var convInfo = conv_util.computeConv2DInfo(x4D.shape, filter.shape, strides, dilations, pad, dimRoundingMode, true);
-        var res = ENV.engine.runKernel(function (backend) { return backend.depthwiseConv2D(x4D, filter, convInfo); }, { x4D: x4D, filter: filter });
+        var grad = function (dy) {
+            util.assert(tupleValuesAreOne(dilations), 'Error in gradient of depthwiseConv2d: dilation rates greater than ' +
+                ("1 are not yet supported. Got dilations '" + dilations + "'"));
+            return {
+                x: function () { return depthwiseConv2dDerInput(x4D.shape, dy, filter, convInfo); },
+                filter: function () { return depthwiseConv2dDerFilter(x4D, dy, filter.shape, convInfo); },
+            };
+        };
+        var res = ENV.engine.runKernel(function (backend) { return backend.depthwiseConv2D(x4D, filter, convInfo); }, { x: x4D, filter: filter }, grad);
         if (reshapedTo4D) {
@@ -266,2 +274,26 @@ return res.as3D(res.shape[1], res.shape[2], res.shape[3]);
 }
+function depthwiseConv2dDerInput(xShape, dy, filter, convInfo) {
+    var dy4D = dy;
+    var reshapedTo4D = false;
+    if (dy.rank === 3) {
+        reshapedTo4D = true;
+        dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
+    }
+    var res = ENV.engine.runKernel(function (backend) { return backend.depthwiseConv2DDerInput(dy4D, filter, convInfo); }, { dy4D: dy4D });
+    if (reshapedTo4D) {
+        return res.as3D(res.shape[1], res.shape[2], res.shape[3]);
+    }
+    return res;
+}
+function depthwiseConv2dDerFilter(x, dy, filterShape, convInfo) {
+    var x4D = x;
+    if (x.rank === 3) {
+        x4D = x.as4D(1, x.shape[0], x.shape[1], x.shape[2]);
+    }
+    var dy4D = dy;
+    if (dy4D.rank === 3) {
+        dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
+    }
+    return ENV.engine.runKernel(function (backend) { return backend.depthwiseConv2DDerFilter(x4D, dy4D, convInfo); }, { x4D: x4D, dy4D: dy4D });
+}
 //# sourceMappingURL=conv.js.map

dist-es6/ops/conv2d_depthwise_test.js

@@ -246,2 +246,127 @@ import * as tf from '../index';
 });
+describeWithFlags('depthwiseConv2d gradients', ALL_ENVS, function () {
+    var images;
+    var filter;
+    var result;
+    var stride = 1;
+    var pad = 'same';
+    beforeEach(function () {
+        images = tf.tensor4d([[
+                [[2, 3, 1], [3, 0, 2]],
+                [[0, 4, 1], [3, 1, 3]]
+            ], [
+                [[2, 1, 0], [0, 3, 3]],
+                [[4, 0, 1], [1, 4, 1]]
+            ]]);
+        filter = tf.tensor4d([[
+                [[1, 1], [1, 1], [0, 0]],
+                [[0, 1], [1, 1], [1, 1]]
+            ], [
+                [[1, 0], [1, 1], [0, 0]],
+                [[0, 1], [1, 0], [0, 0]]
+            ]]);
+        result = tf.tensor4d([[
+                [[2, 8, 8, 7, 2, 2], [6, 3, 1, 1, 0, 0]],
+                [[0, 3, 5, 5, 3, 3], [3, 3, 1, 1, 0, 0]]
+            ], [
+                [[6, 3, 8, 4, 3, 3], [1, 0, 7, 7, 0, 0]],
+                [[4, 5, 4, 4, 1, 1], [1, 1, 4, 4, 0, 0]]
+            ]]);
+    });
+    it('wrt input', function () {
+        var _a = tf.valueAndGrad(function (x) { return tf.depthwiseConv2d(x, filter, stride, pad); })(images), value = _a.value, grad = _a.grad;
+        expectArraysClose(value, result);
+        var expectedGrad = tf.tensor4d([[
+                [[2., 2., 0.], [3., 4., 2.]],
+                [[3., 4., 0.], [5., 7., 2.]]
+            ], [
+                [[2., 2., 0.], [3., 4., 2.]],
+                [[3., 4., 0.], [5., 7., 2.]]
+            ]]);
+        expectArraysClose(grad, expectedGrad);
+    });
+    it('wrt input, squared output', function () {
+        var grad = tf.grad(function (x) { return tf.square(tf.depthwiseConv2d(x, filter, stride, pad)); })(images);
+        var expectedGrad = tf.tensor4d([[
+                [[20., 30., 0.], [34., 34., 8.]],
+                [[10., 50., 0.], [46., 44., 12.]]
+            ], [
+                [[18., 24., 0.], [8., 52., 12.]],
+                [[30., 40., 0.], [22., 76., 4.]]
+            ]]);
+        expectArraysClose(grad, expectedGrad);
+    });
+    it('wrt filter', function () {
+        var _a = tf.valueAndGrad(function (f) { return tf.depthwiseConv2d(images, f, stride, pad); })(filter), value = _a.value, grad = _a.grad;
+        expectArraysClose(value, result);
+        var expectedGrad = tf.tensor4d([[
+                [[15., 15.], [16., 16.], [12., 12.]],
+                [[7., 7.], [8., 8.], [9., 9.]]
+            ], [
+                [[8., 8.], [9., 9.], [6., 6.]],
+                [[4., 4.], [5., 5.], [4., 4.]]
+            ]]);
+        expectArraysClose(grad, expectedGrad);
+    });
+    it('wrt filter, squared output', function () {
+        var grad = tf.grad(function (f) { return tf.square(tf.depthwiseConv2d(images, f, stride, pad)); })(filter);
+        var expectedGrad = tf.tensor4d([[
+                [[120., 122.], [180., 166.], [12., 12.]],
+                [[20., 76.], [90., 66.], [46., 46.]]
+            ], [
+                [[86., 42.], [122., 114.], [10., 10.]],
+                [[24., 54.], [80., 46.], [18., 18.]]
+            ]]);
+        expectArraysClose(grad, expectedGrad);
+    });
+    it('throws error on dilations > 1', function () {
+        var grad = tf.grad(function (x) { return tf.depthwiseConv2d(x, filter, stride, pad, 'NHWC', 2); });
+        expect(function () { return grad(images); })
+            .toThrowError(/dilation rates greater than 1 are not yet supported/);
+    });
+    it('wrt input, stride=2, pad=valid', function () {
+        var dx = tf.grad(function (x) { return tf.depthwiseConv2d(x, filter, 2, 'valid'); })(images);
+        expectArraysClose(dx, tf.tensor4d([[
+                [[2., 2., 0.], [1., 2., 2.]],
+                [[1., 2., 0.], [1., 1., 0.]]
+            ], [
+                [[2., 2., 0.], [1., 2., 2.]],
+                [[1., 2., 0.], [1., 1., 0.]]
+            ]]));
+    });
+    it('wrt filter, stride=2, pad=valid', function () {
+        var df = tf.grad(function (f) { return tf.depthwiseConv2d(images, f, 2, 'valid'); })(filter);
+        expectArraysClose(df, tf.tensor4d([[
+                [[4., 4.], [4., 4.], [1., 1.]],
+                [[3., 3.], [3., 3.], [5., 5.]]
+            ], [
+                [[4., 4.], [4., 4.], [2., 2.]],
+                [[4., 4.], [5., 5.], [4., 4.]]
+            ]]));
+    });
+    it('wrt input and filter, 1x3x3x1 and 2x2x1x1', function () {
+        var fSize = 2;
+        var pad = 'valid';
+        var stride = 1;
+        var chMul = 1;
+        var inDepth = 1;
+        var x = tf.tensor4d([
+            0.230664, 0.987388, 0.0685208, 0.419224, 0.887861, 0.731641,
+            0.0741907, 0.409265, 0.351377
+        ], [1, 3, 3, inDepth]);
+        var f = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
+        var _a = tf.grads(function (x, f) { return tf.depthwiseConv2d(x, f, stride, pad); })([x, f]), dx = _a[0], df = _a[1];
+        expectArraysClose(dx, tf.tensor4d([[
+                [[0.303873], [0.533096], [0.229223]],
+                [[0.448206], [1.480802], [1.032596]],
+                [[0.144333], [0.947706], [0.803373]]
+            ]]));
+        expectArraysClose(df, tf.tensor4d([
+            [[[2.525137]], [[2.6754108]]
+            ], [
+                [[1.7905407]], [[2.380144]]
+            ]
+        ]));
+    });
+});
 //# sourceMappingURL=conv2d_depthwise_test.js.map

dist-es6/ops/logical_ops.js

@@ -41,3 +41,3 @@ var __decorate = (this && this.__decorate) || function (decorators, target, key, desc) {
         util.assertArgumentsAreTensors({ condition: condition, a: a, b: b }, 'where');
-        util.assert(condition.dtype === 'bool' || a.dtype === 'bool' || b.dtype === 'bool', 'Error Array must be of type bool.');
+        util.assert(condition.dtype === 'bool', 'Error Condition must be of type bool.');
         util.assertShapesMatch(a.shape, b.shape, 'Error in where: ');
@@ -44,0 +44,0 @@ if (condition.rank === 1) {

dist-es6/ops/logicalop_test.js

@@ -266,2 +266,11 @@ import * as tf from '../index';
     });
+    it('Invalid condition type', function () {
+        var c = tf.tensor1d([1, 0, 1, 0], 'int32');
+        var a = tf.tensor1d([10, 10, 10, 10], 'bool');
+        var b = tf.tensor1d([20, 20, 20, 20], 'bool');
+        var f = function () {
+            tf.where(c, a, b);
+        };
+        expect(f).toThrowError();
+    });
     it('Tensor1D', function () {
@@ -268,0 +277,0 @@ var c = tf.tensor1d([1, 0, 1, 0], 'bool');

dist-es6/version.d.ts

@@ -1,2 +0,2 @@
-declare const version = "0.11.3";
+declare const version = "0.11.4";
 export { version };

dist-es6/version.js

@@ -1,3 +0,3 @@
-var version = '0.11.3';
+var version = '0.11.4';
 export { version };
 //# sourceMappingURL=version.js.map

dist/io/local_storage_test.js

 "use strict";
+var __awaiter = (this && this.__awaiter) || function (thisArg, _arguments, P, generator) {
+    return new (P || (P = Promise))(function (resolve, reject) {
+        function fulfilled(value) { try { step(generator.next(value)); } catch (e) { reject(e); } }
+        function rejected(value) { try { step(generator["throw"](value)); } catch (e) { reject(e); } }
+        function step(result) { result.done ? resolve(result.value) : new P(function (resolve) { resolve(result.value); }).then(fulfilled, rejected); }
+        step((generator = generator.apply(thisArg, _arguments || [])).next());
+    });
+};
+var __generator = (this && this.__generator) || function (thisArg, body) {
+    var _ = { label: 0, sent: function() { if (t[0] & 1) throw t[1]; return t[1]; }, trys: [], ops: [] }, f, y, t, g;
+    return g = { next: verb(0), "throw": verb(1), "return": verb(2) }, typeof Symbol === "function" && (g[Symbol.iterator] = function() { return this; }), g;
+    function verb(n) { return function (v) { return step([n, v]); }; }
+    function step(op) {
+        if (f) throw new TypeError("Generator is already executing.");
+        while (_) try {
+            if (f = 1, y && (t = y[op[0] & 2 ? "return" : op[0] ? "throw" : "next"]) && !(t = t.call(y, op[1])).done) return t;
+            if (y = 0, t) op = [0, t.value];
+            switch (op[0]) {
+                case 0: case 1: t = op; break;
+                case 4: _.label++; return { value: op[1], done: false };
+                case 5: _.label++; y = op[1]; op = [0]; continue;
+                case 7: op = _.ops.pop(); _.trys.pop(); continue;
+                default:
+                    if (!(t = _.trys, t = t.length > 0 && t[t.length - 1]) && (op[0] === 6 || op[0] === 2)) { _ = 0; continue; }
+                    if (op[0] === 3 && (!t || (op[1] > t[0] && op[1] < t[3]))) { _.label = op[1]; break; }
+                    if (op[0] === 6 && _.label < t[1]) { _.label = t[1]; t = op; break; }
+                    if (t && _.label < t[2]) { _.label = t[2]; _.ops.push(op); break; }
+                    if (t[2]) _.ops.pop();
+                    _.trys.pop(); continue;
+            }
+            op = body.call(thisArg, _);
+        } catch (e) { op = [6, e]; y = 0; } finally { f = t = 0; }
+        if (op[0] & 5) throw op[1]; return { value: op[0] ? op[1] : void 0, done: true };
+    }
+};
+var _this = this;
 Object.defineProperty(exports, "__esModule", { value: true });
@@ -120,22 +156,22 @@ var tf = require("../index");
     });
-    it('Save-load round trip succeeds', function (done) {
-        var handler1 = tf.io.getSaveHandlers('localstorage://FooModel')[0];
-        handler1.save(artifacts1)
-            .then(function (saveResult) {
-            var handler2 = tf.io.getLoadHandlers('localstorage://FooModel')[0];
-            handler2.load()
-                .then(function (loaded) {
-                expect(loaded.modelTopology).toEqual(modelTopology1);
-                expect(loaded.weightSpecs).toEqual(weightSpecs1);
-                expect(loaded.weightData).toEqual(weightData1);
-                done();
-            })
-                .catch(function (err) {
-                console.error(err.stack);
-            });
-        })
-            .catch(function (err) {
-            console.error(err.stack);
+    it('Save-load round trip succeeds', function () { return __awaiter(_this, void 0, void 0, function () {
+        var handler1, handler2, loaded;
+        return __generator(this, function (_a) {
+            switch (_a.label) {
+                case 0:
+                    handler1 = tf.io.getSaveHandlers('localstorage://FooModel')[0];
+                    return [4, handler1.save(artifacts1)];
+                case 1:
+                    _a.sent();
+                    handler2 = tf.io.getLoadHandlers('localstorage://FooModel')[0];
+                    return [4, handler2.load()];
+                case 2:
+                    loaded = _a.sent();
+                    expect(loaded.modelTopology).toEqual(modelTopology1);
+                    expect(loaded.weightSpecs).toEqual(weightSpecs1);
+                    expect(loaded.weightData).toEqual(weightData1);
+                    return [2];
+            }
         });
-    });
+    }); });
     it('Loading nonexistent model fails.', function (done) {
@@ -142,0 +178,0 @@ var handler = tf.io.getSaveHandlers('localstorage://NonexistentModel')[0];

dist/kernels/backend_cpu.d.ts

@@ -95,2 +95,4 @@ import { Conv2DInfo } from '../ops/conv_util';
     depthwiseConv2D(x: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerInput(dy: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerFilter(x: Tensor4D, dy: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
     tile<T extends Tensor>(x: T, reps: number[]): T;
@@ -97,0 +99,0 @@ pad<T extends Tensor>(x: T, paddings: Array<[number, number]>, constantValue: number): T;

dist/kernels/backend_webgl.d.ts

@@ -130,2 +130,4 @@ import { TimingInfo } from '../engine';
     depthwiseConv2D(x: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerInput(dy: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerFilter(x: Tensor4D, dy: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
     maxPool(x: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
@@ -132,0 +134,0 @@ avgPool(x: Tensor4D, convInfo: Conv2DInfo): Tensor4D;

dist/kernels/backend_webgl.js

@@ -55,2 +55,3 @@ "use strict";
 var conv_backprop_gpu_1 = require("./webgl/conv_backprop_gpu");
+var conv_backprop_gpu_depthwise_1 = require("./webgl/conv_backprop_gpu_depthwise");
 var conv_gpu_1 = require("./webgl/conv_gpu");
@@ -732,2 +733,10 @@ var conv_gpu_depthwise_1 = require("./webgl/conv_gpu_depthwise");
     };
+    MathBackendWebGL.prototype.depthwiseConv2DDerInput = function (dy, filter, convInfo) {
+        var program = new conv_backprop_gpu_depthwise_1.DepthwiseConv2DDerInputProgram(convInfo);
+        return this.compileAndRun(program, [dy, filter]);
+    };
+    MathBackendWebGL.prototype.depthwiseConv2DDerFilter = function (x, dy, convInfo) {
+        var program = new conv_backprop_gpu_depthwise_1.DepthwiseConv2DDerFilterProgram(convInfo);
+        return this.compileAndRun(program, [x, dy]);
+    };
     MathBackendWebGL.prototype.maxPool = function (x, convInfo) {
@@ -734,0 +743,0 @@ var program = new pool_gpu_1.Pool2DProgram(convInfo, 'max', false);

dist/kernels/backend.d.ts

@@ -94,2 +94,4 @@ import { Conv2DInfo } from '../ops/conv_util';
     depthwiseConv2D(input: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerInput(dy: Tensor4D, filter: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
+    depthwiseConv2DDerFilter(x: Tensor4D, dY: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
     maxPool(x: Tensor4D, convInfo: Conv2DInfo): Tensor4D;
@@ -96,0 +98,0 @@ maxPoolBackprop(dy: Tensor4D, x: Tensor4D, y: Tensor4D, convInfo: Conv2DInfo): Tensor4D;

dist/ops/conv.js

@@ -167,7 +167,7 @@ "use strict";
         }
-        util.assert(x4D.rank === 4, "Error in depthwiseConv2D: input must be rank 4, but got " +
+        util.assert(x4D.rank === 4, "Error in depthwiseConv2d: input must be rank 4, but got " +
             ("rank " + x4D.rank + "."));
-        util.assert(filter.rank === 4, "Error in depthwiseConv2D: filter must be rank 4, but got rank " +
+        util.assert(filter.rank === 4, "Error in depthwiseConv2d: filter must be rank 4, but got rank " +
             (filter.rank + "."));
-        util.assert(x4D.shape[3] === filter.shape[2], "Error in depthwiseConv2D: number of input channels " +
+        util.assert(x4D.shape[3] === filter.shape[2], "Error in depthwiseConv2d: number of input channels " +
             ("(" + x4D.shape[3] + ") must match the inChannels dimension in ") +
@@ -181,7 +181,15 @@ ("filter " + filter.shape[2] + "."));
         if (dimRoundingMode != null) {
-            util.assert(util.isInt(pad), "Error in depthwiseConv2D: pad must be an integer when using, " +
+            util.assert(util.isInt(pad), "Error in depthwiseConv2d: pad must be an integer when using, " +
                 ("dimRoundingMode " + dimRoundingMode + " but got pad " + pad + "."));
         }
         var convInfo = conv_util.computeConv2DInfo(x4D.shape, filter.shape, strides, dilations, pad, dimRoundingMode, true);
-        var res = environment_1.ENV.engine.runKernel(function (backend) { return backend.depthwiseConv2D(x4D, filter, convInfo); }, { x4D: x4D, filter: filter });
+        var grad = function (dy) {
+            util.assert(tupleValuesAreOne(dilations), 'Error in gradient of depthwiseConv2d: dilation rates greater than ' +
+                ("1 are not yet supported. Got dilations '" + dilations + "'"));
+            return {
+                x: function () { return depthwiseConv2dDerInput(x4D.shape, dy, filter, convInfo); },
+                filter: function () { return depthwiseConv2dDerFilter(x4D, dy, filter.shape, convInfo); },
+            };
+        };
+        var res = environment_1.ENV.engine.runKernel(function (backend) { return backend.depthwiseConv2D(x4D, filter, convInfo); }, { x: x4D, filter: filter }, grad);
         if (reshapedTo4D) {
@@ -268,2 +276,26 @@ return res.as3D(res.shape[1], res.shape[2], res.shape[3]);
 }
+function depthwiseConv2dDerInput(xShape, dy, filter, convInfo) {
+    var dy4D = dy;
+    var reshapedTo4D = false;
+    if (dy.rank === 3) {
+        reshapedTo4D = true;
+        dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
+    }
+    var res = environment_1.ENV.engine.runKernel(function (backend) { return backend.depthwiseConv2DDerInput(dy4D, filter, convInfo); }, { dy4D: dy4D });
+    if (reshapedTo4D) {
+        return res.as3D(res.shape[1], res.shape[2], res.shape[3]);
+    }
+    return res;
+}
+function depthwiseConv2dDerFilter(x, dy, filterShape, convInfo) {
+    var x4D = x;
+    if (x.rank === 3) {
+        x4D = x.as4D(1, x.shape[0], x.shape[1], x.shape[2]);
+    }
+    var dy4D = dy;
+    if (dy4D.rank === 3) {
+        dy4D = dy.as4D(1, dy.shape[0], dy.shape[1], dy.shape[2]);
+    }
+    return environment_1.ENV.engine.runKernel(function (backend) { return backend.depthwiseConv2DDerFilter(x4D, dy4D, convInfo); }, { x4D: x4D, dy4D: dy4D });
+}
 //# sourceMappingURL=conv.js.map

dist/ops/conv2d_depthwise_test.js

@@ -248,2 +248,127 @@ "use strict";
 });
+jasmine_util_1.describeWithFlags('depthwiseConv2d gradients', test_util_1.ALL_ENVS, function () {
+    var images;
+    var filter;
+    var result;
+    var stride = 1;
+    var pad = 'same';
+    beforeEach(function () {
+        images = tf.tensor4d([[
+                [[2, 3, 1], [3, 0, 2]],
+                [[0, 4, 1], [3, 1, 3]]
+            ], [
+                [[2, 1, 0], [0, 3, 3]],
+                [[4, 0, 1], [1, 4, 1]]
+            ]]);
+        filter = tf.tensor4d([[
+                [[1, 1], [1, 1], [0, 0]],
+                [[0, 1], [1, 1], [1, 1]]
+            ], [
+                [[1, 0], [1, 1], [0, 0]],
+                [[0, 1], [1, 0], [0, 0]]
+            ]]);
+        result = tf.tensor4d([[
+                [[2, 8, 8, 7, 2, 2], [6, 3, 1, 1, 0, 0]],
+                [[0, 3, 5, 5, 3, 3], [3, 3, 1, 1, 0, 0]]
+            ], [
+                [[6, 3, 8, 4, 3, 3], [1, 0, 7, 7, 0, 0]],
+                [[4, 5, 4, 4, 1, 1], [1, 1, 4, 4, 0, 0]]
+            ]]);
+    });
+    it('wrt input', function () {
+        var _a = tf.valueAndGrad(function (x) { return tf.depthwiseConv2d(x, filter, stride, pad); })(images), value = _a.value, grad = _a.grad;
+        test_util_1.expectArraysClose(value, result);
+        var expectedGrad = tf.tensor4d([[
+                [[2., 2., 0.], [3., 4., 2.]],
+                [[3., 4., 0.], [5., 7., 2.]]
+            ], [
+                [[2., 2., 0.], [3., 4., 2.]],
+                [[3., 4., 0.], [5., 7., 2.]]
+            ]]);
+        test_util_1.expectArraysClose(grad, expectedGrad);
+    });
+    it('wrt input, squared output', function () {
+        var grad = tf.grad(function (x) { return tf.square(tf.depthwiseConv2d(x, filter, stride, pad)); })(images);
+        var expectedGrad = tf.tensor4d([[
+                [[20., 30., 0.], [34., 34., 8.]],
+                [[10., 50., 0.], [46., 44., 12.]]
+            ], [
+                [[18., 24., 0.], [8., 52., 12.]],
+                [[30., 40., 0.], [22., 76., 4.]]
+            ]]);
+        test_util_1.expectArraysClose(grad, expectedGrad);
+    });
+    it('wrt filter', function () {
+        var _a = tf.valueAndGrad(function (f) { return tf.depthwiseConv2d(images, f, stride, pad); })(filter), value = _a.value, grad = _a.grad;
+        test_util_1.expectArraysClose(value, result);
+        var expectedGrad = tf.tensor4d([[
+                [[15., 15.], [16., 16.], [12., 12.]],
+                [[7., 7.], [8., 8.], [9., 9.]]
+            ], [
+                [[8., 8.], [9., 9.], [6., 6.]],
+                [[4., 4.], [5., 5.], [4., 4.]]
+            ]]);
+        test_util_1.expectArraysClose(grad, expectedGrad);
+    });
+    it('wrt filter, squared output', function () {
+        var grad = tf.grad(function (f) { return tf.square(tf.depthwiseConv2d(images, f, stride, pad)); })(filter);
+        var expectedGrad = tf.tensor4d([[
+                [[120., 122.], [180., 166.], [12., 12.]],
+                [[20., 76.], [90., 66.], [46., 46.]]
+            ], [
+                [[86., 42.], [122., 114.], [10., 10.]],
+                [[24., 54.], [80., 46.], [18., 18.]]
+            ]]);
+        test_util_1.expectArraysClose(grad, expectedGrad);
+    });
+    it('throws error on dilations > 1', function () {
+        var grad = tf.grad(function (x) { return tf.depthwiseConv2d(x, filter, stride, pad, 'NHWC', 2); });
+        expect(function () { return grad(images); })
+            .toThrowError(/dilation rates greater than 1 are not yet supported/);
+    });
+    it('wrt input, stride=2, pad=valid', function () {
+        var dx = tf.grad(function (x) { return tf.depthwiseConv2d(x, filter, 2, 'valid'); })(images);
+        test_util_1.expectArraysClose(dx, tf.tensor4d([[
+                [[2., 2., 0.], [1., 2., 2.]],
+                [[1., 2., 0.], [1., 1., 0.]]
+            ], [
+                [[2., 2., 0.], [1., 2., 2.]],
+                [[1., 2., 0.], [1., 1., 0.]]
+            ]]));
+    });
+    it('wrt filter, stride=2, pad=valid', function () {
+        var df = tf.grad(function (f) { return tf.depthwiseConv2d(images, f, 2, 'valid'); })(filter);
+        test_util_1.expectArraysClose(df, tf.tensor4d([[
+                [[4., 4.], [4., 4.], [1., 1.]],
+                [[3., 3.], [3., 3.], [5., 5.]]
+            ], [
+                [[4., 4.], [4., 4.], [2., 2.]],
+                [[4., 4.], [5., 5.], [4., 4.]]
+            ]]));
+    });
+    it('wrt input and filter, 1x3x3x1 and 2x2x1x1', function () {
+        var fSize = 2;
+        var pad = 'valid';
+        var stride = 1;
+        var chMul = 1;
+        var inDepth = 1;
+        var x = tf.tensor4d([
+            0.230664, 0.987388, 0.0685208, 0.419224, 0.887861, 0.731641,
+            0.0741907, 0.409265, 0.351377
+        ], [1, 3, 3, inDepth]);
+        var f = tf.tensor4d([0.303873, 0.229223, 0.144333, 0.803373], [fSize, fSize, inDepth, chMul]);
+        var _a = tf.grads(function (x, f) { return tf.depthwiseConv2d(x, f, stride, pad); })([x, f]), dx = _a[0], df = _a[1];
+        test_util_1.expectArraysClose(dx, tf.tensor4d([[
+                [[0.303873], [0.533096], [0.229223]],
+                [[0.448206], [1.480802], [1.032596]],
+                [[0.144333], [0.947706], [0.803373]]
+            ]]));
+        test_util_1.expectArraysClose(df, tf.tensor4d([
+            [[[2.525137]], [[2.6754108]]
+            ], [
+                [[1.7905407]], [[2.380144]]
+            ]
+        ]));
+    });
+});
 //# sourceMappingURL=conv2d_depthwise_test.js.map

dist/ops/logical_ops.js

@@ -43,3 +43,3 @@ "use strict";
         util.assertArgumentsAreTensors({ condition: condition, a: a, b: b }, 'where');
-        util.assert(condition.dtype === 'bool' || a.dtype === 'bool' || b.dtype === 'bool', 'Error Array must be of type bool.');
+        util.assert(condition.dtype === 'bool', 'Error Condition must be of type bool.');
         util.assertShapesMatch(a.shape, b.shape, 'Error in where: ');
@@ -46,0 +46,0 @@ if (condition.rank === 1) {

dist/ops/logicalop_test.js

@@ -268,2 +268,11 @@ "use strict";
     });
+    it('Invalid condition type', function () {
+        var c = tf.tensor1d([1, 0, 1, 0], 'int32');
+        var a = tf.tensor1d([10, 10, 10, 10], 'bool');
+        var b = tf.tensor1d([20, 20, 20, 20], 'bool');
+        var f = function () {
+            tf.where(c, a, b);
+        };
+        expect(f).toThrowError();
+    });
     it('Tensor1D', function () {
@@ -270,0 +279,0 @@ var c = tf.tensor1d([1, 0, 1, 0], 'bool');

dist/version.d.ts

@@ -1,2 +0,2 @@
-declare const version = "0.11.3";
+declare const version = "0.11.4";
 export { version };

dist/version.js

 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-var version = '0.11.3';
+var version = '0.11.4';
 exports.version = version;
 //# sourceMappingURL=version.js.map

package.json

 {
   "name": "@tensorflow/tfjs-core",
-  "version": "0.11.3",
+  "version": "0.11.4",
   "description": "Hardware-accelerated JavaScript library for machine intelligence",
@@ -22,3 +22,3 @@ "private": false,
     "clang-format": "~1.2.2",
-    "jasmine-core": "~2.6.4",
+    "jasmine-core": "~3.1.0",
     "karma": "~1.7.0",
@@ -25,0 +25,0 @@ "karma-browserstack-launcher": "~1.3.0",

New alerts

License Policy Violation

License

This package is not allowed per your license policy. Review the package's license to ensure compliance.

Found 1 instance in 1 package

Fixed alerts

License Policy Violation

License

This package is not allowed per your license policy. Review the package's license to ensure compliance.

Found 1 instance in 1 package

Improved metrics

Total package byte prevSize

6710277

increased by1.08%

Number of package files

961

increased by0.52%

Lines of code

93414

increased by0.92%

No dependency changes