UNPKG

@tensorflow/tfjs-core

Version:

Hardware-accelerated JavaScript library for machine intelligence

829 lines (828 loc) 38 kB
"use strict"; Object.defineProperty(exports, "__esModule", { value: true }); var tf = require("../index"); var test_util_1 = require("../test_util"); var jasmine_util_1 = require("../jasmine_util"); jasmine_util_1.describeWithFlags('div', test_util_1.ALL_ENVS, function () { it('same shape', function () { var a = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]); var c = tf.tensor2d([1, 2, 3, 4, 2, 5], [2, 3]); var r = tf.div(a, c); test_util_1.expectArraysClose(r, [1, 1, 1, 1, 2.5, 6 / 5]); }); it('integer division implements floor divide', function () { var a = tf.tensor1d([-6, -6, -5, -4, -3, -3, 3, 3, 2], 'int32'); var c = tf.tensor1d([-2, 2, 3, 2, -3, 3, 2, 3, 2], 'int32'); var r = tf.div(a, c); expect(r.dtype).toEqual('int32'); test_util_1.expectArraysClose(r, [3, -3, -2, -2, 1, -1, 1, 1, 1]); }); it('integer division broadcasts', function () { var a = tf.tensor1d([-5, -4, 3, 2], 'int32'); var c = tf.scalar(2, 'int32'); var r = tf.div(a, c); expect(r.dtype).toEqual('int32'); test_util_1.expectArraysClose(r, [-3, -2, 1, 1]); }); it('propagates NaNs', function () { var a = tf.tensor2d([1, 2], [2, 1]); var c = tf.tensor2d([3, NaN], [2, 1]); var r = tf.div(a, c); test_util_1.expectArraysClose(r, [1 / 3, NaN]); }); it('broadcasting same rank Tensors different shape', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor2d([2, 3], [2, 1]); var result = tf.div(a, b); expect(result.shape).toEqual([2, 2]); var expected = [1 / 2, 1, -1, -4 / 3]; test_util_1.expectArraysClose(result, expected); }); it('broadcast 2D + 1D', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor1d([1, 2]); var result = tf.div(a, b); expect(result.shape).toEqual([2, 2]); var expected = [1, 1, -3, -2]; test_util_1.expectArraysClose(result, expected); }); it('throws when passed tensors of different types', function () { var a = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]); var b = tf.tensor2d([1, 2, 3, 4, 2, 5], [2, 3], 'int32'); expect(function () { return tf.div(a, b); }).toThrowError(); expect(function () { return tf.div(b, a); }).toThrowError(); }); it('throws when passed tensors of different shapes', function () { var a = tf.tensor2d([1, 2, -3, -4, 5, 6], [2, 3]); var b = tf.tensor2d([5, 3, 4, -7], [2, 2]); expect(function () { return tf.div(a, b); }).toThrowError(); expect(function () { return tf.div(b, a); }).toThrowError(); }); it('scalar divided by array', function () { var c = tf.scalar(2); var a = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]); var r = tf.div(c, a); test_util_1.expectArraysClose(r, [2 / 1, 2 / 2, 2 / 3, 2 / 4, 2 / 5, 2 / 6]); }); it('scalar divided by array propagates NaNs', function () { var c = tf.scalar(NaN); var a = tf.tensor2d([1, 2, 3], [1, 3]); var r = tf.div(c, a); test_util_1.expectArraysEqual(r, [NaN, NaN, NaN]); }); it('array divided by scalar', function () { var a = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]); var c = tf.scalar(2); var r = tf.div(a, c); test_util_1.expectArraysClose(r, [1 / 2, 2 / 2, 3 / 2, 4 / 2, 5 / 2, 6 / 2]); }); it('array divided by scalar propagates NaNs', function () { var a = tf.tensor2d([1, 2, NaN], [1, 3]); var c = tf.scalar(2); var r = tf.div(a, c); test_util_1.expectArraysClose(r, [1 / 2, 2 / 2, NaN]); }); it('gradient: Scalar', function () { var a = tf.scalar(5); var b = tf.scalar(2); var dy = tf.scalar(4); var grads = tf.grads(function (a, b) { return tf.div(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [4 / 2]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-4 * 5 / (2 * 2)]); }); it('gradient: Tensor1D', function () { var a = tf.tensor1d([1, 2, 3]); var b = tf.tensor1d([3, 4, 5]); var dy = tf.tensor1d([1, 10, 20]); var grads = tf.grads(function (a, b) { return tf.div(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [1 / 3, 10 / 4, 20 / 5]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-1 * 1 / 9, -10 * 2 / 16, -20 * 3 / 25]); }); it('gradient: Tensor1D with int32', function () { var a = tf.tensor1d([1, 2, 3], 'int32'); var b = tf.tensor1d([3, 4, 5], 'int32'); var dy = tf.tensor1d([1, 10, 20]); var grads = tf.grads(function (a, b) { return tf.div(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [1 / 3, 10 / 4, 20 / 5]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-1 * 1 / 9, -10 * 2 / 16, -20 * 3 / 25]); }); it('gradient: 1d<int32> with 1d<bool> ', function () { var a = tf.tensor1d([true, false, true], 'bool'); var b = tf.tensor1d([1, 2, 3], 'int32'); var dy = tf.tensor1d([1, 19, 20]); var grads = tf.grads(function (a, b) { return tf.div(a.toInt(), b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [1, 19 / 2, 20 / 3]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-1 / 1, 0, -20 / 9]); }); it('gradient: Tensor2D', function () { var a = tf.tensor2d([3, 1, 2, 3], [2, 2]); var b = tf.tensor2d([1, 3, 4, 5], [2, 2]); var dy = tf.tensor2d([1, 10, 15, 20], [2, 2]); var grads = tf.grads(function (a, b) { return tf.div(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [1 / 1, 10 / 3, 15 / 4, 20 / 5]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-1 * 3 / 1, -10 * 1 / 9, -15 * 2 / 16, -20 * 3 / 25]); }); it('gradient: scalar / Tensor1D', function () { var a = tf.scalar(2); var b = tf.tensor1d([3, 4, 5]); var dy = tf.tensor1d([6, 7, 8]); var grads = tf.grads(function (a, b) { return tf.div(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [6 / 3 + 7 / 4 + 8 / 5]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-6 * 2 / 9, -7 * 2 / 16, -8 * 2 / 25]); }); it('gradient: Tensor2D / scalar', function () { var a = tf.tensor2d([[2, 3], [4, 5]], [2, 2]); var b = tf.scalar(2); var dy = tf.tensor2d([[6, 7], [8, 9]], [2, 2]); var grads = tf.grads(function (a, b) { return tf.div(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [6 / 2, 7 / 2, 8 / 2, 9 / 2]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-6 * 2 / 4 + -7 * 3 / 4 + -8 * 4 / 4 + -9 * 5 / 4]); }); it('gradient: Tensor2D / Tensor2D w/ broadcast', function () { var a = tf.tensor2d([3, 4], [2, 1]); var b = tf.tensor2d([[2, 3], [4, 5]], [2, 2]); var dy = tf.tensor2d([[6, 7], [8, 9]], [2, 2]); var grads = tf.grads(function (a, b) { return tf.div(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [6 / 2 + 7 / 3, 8 / 4 + 9 / 5]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-6 * 3 / 4, -7 * 3 / 9, -8 * 4 / 16, -9 * 4 / 25]); }); it('throws when passed a as a non-tensor', function () { expect(function () { return tf.div({}, tf.scalar(1)); }) .toThrowError(/Argument 'a' passed to 'div' must be a Tensor/); }); it('throws when passed b as a non-tensor', function () { expect(function () { return tf.div(tf.scalar(1), {}); }) .toThrowError(/Argument 'b' passed to 'div' must be a Tensor/); }); }); jasmine_util_1.describeWithFlags('mul', test_util_1.ALL_ENVS, function () { it('strict same-shaped tensors', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor2d([5, 3, 4, -7], [2, 2]); var expected = [5, 6, -12, 28]; var result = tf.mulStrict(a, b); expect(result.shape).toEqual([2, 2]); expect(result.dtype).toBe('float32'); test_util_1.expectArraysClose(result, expected); }); it('strict propagates NaNs', function () { var a = tf.tensor2d([1, 3, 4, 0], [2, 2]); var b = tf.tensor2d([NaN, 3, NaN, 3], [2, 2]); var result = tf.mulStrict(a, b); expect(result.dtype).toBe('float32'); test_util_1.expectArraysClose(result, [NaN, 9, NaN, 0]); }); it('strict throws when passed tensors of different shapes', function () { var a = tf.tensor2d([1, 2, -3, -4, 5, 6], [2, 3]); var b = tf.tensor2d([5, 3, 4, -7], [2, 2]); expect(function () { return tf.mulStrict(a, b); }).toThrowError(); expect(function () { return tf.mulStrict(b, a); }).toThrowError(); }); it('strict throws when dtypes do not match', function () { var a = tf.tensor2d([1, 2, -3, -4, 5, 6], [2, 3], 'float32'); var b = tf.tensor2d([5, 3, 4, -7], [2, 2], 'int32'); expect(function () { return tf.mulStrict(a, b); }).toThrowError(); expect(function () { return tf.mulStrict(b, a); }).toThrowError(); }); it('strict int32 * int32', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2], 'int32'); var b = tf.tensor2d([2, 1, 3, -4], [2, 2], 'int32'); var res = tf.mulStrict(a, b); expect(res.dtype).toBe('int32'); test_util_1.expectArraysClose(res, [2, 2, -9, 16]); }); it('same-shaped tensors', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor2d([5, 3, 4, -7], [2, 2]); var expected = [5, 6, -12, 28]; var result = tf.mul(a, b); expect(result.shape).toEqual([2, 2]); test_util_1.expectArraysClose(result, expected); }); it('broadcasting tensors', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.scalar(2); var expected = [2, 4, -6, -8]; var result = tf.mul(a, b); expect(result.shape).toEqual([2, 2]); test_util_1.expectArraysClose(result, expected); }); it('broadcasting same rank Tensors different shape', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor2d([2, 3], [2, 1]); var result = tf.mul(a, b); expect(result.shape).toEqual([2, 2]); var expected = [2, 4, -9, -12]; test_util_1.expectArraysClose(result, expected); }); it('broadcast 2D + 1D', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor1d([1, 2]); var result = tf.mul(a, b); expect(result.shape).toEqual([2, 2]); var expected = [1, 4, -3, -8]; test_util_1.expectArraysClose(result, expected); }); it('gradient: Scalar', function () { var a = tf.scalar(5); var b = tf.scalar(2); var dy = tf.scalar(4); var grads = tf.grads(function (a, b) { return tf.mul(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [b.get() * dy.get()]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [a.get() * dy.get()]); }); it('gradient: Tensor1D', function () { var a = tf.tensor1d([1, 2, 3]); var b = tf.tensor1d([3, 4, 5]); var dy = tf.tensor1d([1, 10, 20]); var grads = tf.grads(function (a, b) { return tf.mul(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [3 * 1, 4 * 10, 5 * 20]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [1 * 1, 2 * 10, 3 * 20]); }); it('gradient: Tensor1D with dtype int32', function () { var a = tf.tensor1d([1, 2, 3], 'int32'); var b = tf.tensor1d([3, 4, 5], 'int32'); var dy = tf.tensor1d([1, 10, 20]); var grads = tf.grads(function (a, b) { return tf.mul(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [3 * 1, 4 * 10, 5 * 20]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [1 * 1, 2 * 10, 3 * 20]); }); it('gradient: Tensor2D', function () { var a = tf.tensor2d([3, 1, 2, 3], [2, 2]); var b = tf.tensor2d([1, 3, 4, 5], [2, 2]); var dy = tf.tensor2d([1, 10, 15, 20], [2, 2]); var grads = tf.grads(function (a, b) { return tf.mul(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [1 * 1, 3 * 10, 4 * 15, 5 * 20]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [3 * 1, 1 * 10, 2 * 15, 3 * 20]); }); it('gradient: scalar * Tensor1D', function () { var a = tf.scalar(2); var b = tf.tensor1d([3, 4, 5]); var dy = tf.tensor1d([6, 7, 8]); var grads = tf.grads(function (a, b) { return tf.mul(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [3 * 6 + 4 * 7 + 5 * 8]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [2 * 6, 2 * 7, 2 * 8]); }); it('gradient: Tensor2D * scalar', function () { var a = tf.tensor2d([[2, 3], [4, 5]], [2, 2]); var b = tf.scalar(2); var dy = tf.tensor2d([[6, 7], [8, 9]], [2, 2]); var grads = tf.grads(function (a, b) { return tf.mul(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [2 * 6, 2 * 7, 2 * 8, 2 * 9]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [2 * 6 + 3 * 7 + 4 * 8 + 5 * 9]); }); it('gradient: Tensor2D * Tensor2D w/ broadcast', function () { var a = tf.tensor2d([3, 4], [2, 1]); var b = tf.tensor2d([[2, 3], [4, 5]], [2, 2]); var dy = tf.tensor2d([[6, 7], [8, 9]], [2, 2]); var grads = tf.grads(function (a, b) { return tf.mul(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [2 * 6 + 3 * 7, 4 * 8 + 5 * 9]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [6 * 3, 7 * 3, 8 * 4, 9 * 4]); }); it('throws when passed a as a non-tensor', function () { expect(function () { return tf.mul({}, tf.scalar(1)); }) .toThrowError(/Argument 'a' passed to 'mul' must be a Tensor/); }); it('throws when passed b as a non-tensor', function () { expect(function () { return tf.mul(tf.scalar(1), {}); }) .toThrowError(/Argument 'b' passed to 'mul' must be a Tensor/); }); }); jasmine_util_1.describeWithFlags('pow', test_util_1.ALL_ENVS, function () { it('same-shaped tensors', function () { var a = tf.tensor2d([1, -2, -3, 0, 7, 1], [2, 3]); var b = tf.tensor2d([5, 3, 4, 5, 2, -3], [2, 3], 'int32'); var expected = [1, -8, 81, 0, 49, 1]; var result = tf.pow(a, b); expect(result.shape).toEqual([2, 3]); test_util_1.expectArraysClose(result, expected, 0.01); }); it('int32^int32 returns int32', function () { var a = tf.tensor1d([1, 2, 3], 'int32'); var exp = tf.scalar(2, 'int32'); var result = tf.pow(a, exp); expect(result.shape).toEqual([3]); expect(result.dtype).toBe('int32'); test_util_1.expectArraysEqual(result, [1, 4, 9]); }); it('different-shaped tensors', function () { var a = tf.tensor2d([1, -2, -3, 0, 7, 1], [2, 3]); var b = tf.scalar(2, 'int32'); var expected = [1, 4, 9, 0, 49, 1]; var result = tf.pow(a, b); expect(result.shape).toEqual([2, 3]); test_util_1.expectArraysClose(result, expected, 0.05); }); it('propagates NaNs', function () { var a = tf.tensor2d([NaN, 3, NaN, 0], [2, 2]); var b = tf.tensor2d([1, 3, 2, 3], [2, 2], 'int32'); var result = tf.pow(a, b); test_util_1.expectArraysClose(result, [NaN, 27, NaN, 0], 0.05); }); it('handles non int32 exponent param', function () { var a = tf.tensor1d([2, 4]); var b = tf.tensor1d([.5, 1.2]); var result = tf.pow(a, b); var expected = [Math.pow(2, 0.5), Math.pow(4, 1.2)]; test_util_1.expectArraysClose(result, expected); }); it('broadcasting same rank Tensors different shape', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor2d([2, 1], [2, 1], 'int32'); var result = tf.pow(a, b); expect(result.shape).toEqual([2, 2]); var expected = [1, 4, -3, -4]; test_util_1.expectArraysClose(result, expected); }); it('broadcast 2D + 1D', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor1d([1, 2], 'int32'); var result = tf.pow(a, b); expect(result.shape).toEqual([2, 2]); var expected = [1, 4, -3, 16]; test_util_1.expectArraysClose(result, expected); }); it('powStrict same-shaped tensors', function () { var a = tf.tensor2d([1, -2, -3, 0, 7, 1], [2, 3]); var b = tf.tensor2d([5, 3, 4, 5, 2, -3], [2, 3], 'int32'); var expected = [1, -8, 81, 0, 49, 1]; var result = tf.powStrict(a, b); expect(result.shape).toEqual([2, 3]); test_util_1.expectArraysClose(result, expected, 0.01); }); it('powStrict throws when passed tensors of different shapes', function () { var a = tf.tensor2d([1, 2, -3, -4, 5, 6], [2, 3]); var b = tf.tensor2d([5, 3, 4, -7], [2, 2], 'int32'); expect(function () { return tf.powStrict(a, b); }).toThrowError(); }); it('powStrict handles non int32 exponent param', function () { var a = tf.tensor1d([2, 4]); var b = tf.tensor1d([.5, 1.2]); var result = tf.powStrict(a, b); var expected = [Math.pow(2, 0.5), Math.pow(4, 1.2)]; test_util_1.expectArraysClose(result, expected); }); it('gradients: Scalar ^ Scalar', function () { var a = tf.scalar(5); var b = tf.scalar(2, 'int32'); var dy = tf.scalar(3); var grads = tf.grads(function (a, b) { return tf.pow(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [2 * 5 * 3]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [3 * Math.pow(5, 2) * Math.log(5)]); }); it('gradients: Scalar ^ Scalar fractional exponent', function () { var a = tf.scalar(4.0); var b = tf.scalar(1.5); var dy = tf.scalar(3.0); var grads = tf.grads(function (a, b) { return tf.pow(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [1.5 * Math.pow(4, 0.5) * 3]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [3.0 * Math.pow(4, 1.5) * Math.log(4.0)]); }); it('gradients: Tensor ^ Tensor', function () { var a = tf.tensor1d([-1, .5, 2]); var b = tf.tensor1d([3, 2, -1], 'int32'); var dy = tf.tensor1d([1, 5, 10]); var grads = tf.grads(function (a, b) { return tf.pow(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [ 3 * Math.pow(-1, 2) * 1, 2 * Math.pow(.5, 1) * 5, -1 * Math.pow(2, -2) * 10 ], 1e-1); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [ NaN, 5 * Math.pow(.5, 2) * Math.log(.5), 10 * Math.pow(2, -1) * Math.log(2) ]); }); it('gradient: scalar / Tensor1D', function () { var a = tf.scalar(2); var b = tf.tensor1d([3, 4, 5]); var dy = tf.tensor1d([6, 7, 8]); var grads = tf.grads(function (a, b) { return tf.pow(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [ 6 * 3 * Math.pow(2, 2) + 7 * 4 * Math.pow(2, 3) + 8 * 5 * Math.pow(2, 4) ]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [ 6 * Math.pow(2, 3) * Math.log(2), 7 * Math.pow(2, 4) * Math.log(2), 8 * Math.pow(2, 5) * Math.log(2) ]); }); it('gradient: Tensor2D / scalar', function () { var a = tf.tensor2d([[2, 3], [4, 5]], [2, 2]); var b = tf.scalar(2); var dy = tf.tensor2d([[6, 7], [8, 9]], [2, 2]); var grads = tf.grads(function (a, b) { return tf.pow(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [ 6 * 2 * Math.pow(2, 1), 7 * 2 * Math.pow(3, 1), 8 * 2 * Math.pow(4, 1), 9 * 2 * Math.pow(5, 1) ]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [6 * Math.pow(2, 2) * Math.log(2) + 7 * Math.pow(3, 2) * Math.log(3) + 8 * Math.pow(4, 2) * Math.log(4) + 9 * Math.pow(5, 2) * Math.log(5)]); }); it('gradient: Tensor2D / Tensor2D w/ broadcast', function () { var a = tf.tensor2d([3, 4], [2, 1]); var b = tf.tensor2d([[2, 3], [4, 5]], [2, 2]); var dy = tf.tensor2d([[6, 7], [8, 9]], [2, 2]); var grads = tf.grads(function (a, b) { return tf.pow(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [ 6 * 2 * Math.pow(3, 1) + 7 * 3 * Math.pow(3, 2), 8 * 4 * Math.pow(4, 3) + 9 * 5 * Math.pow(4, 4) ]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [ 6 * Math.pow(3, 2) * Math.log(3), 7 * Math.pow(3, 3) * Math.log(3), 8 * Math.pow(4, 4) * Math.log(4), 9 * Math.pow(4, 5) * Math.log(4) ]); }); it('throws when passed base as a non-tensor', function () { expect(function () { return tf.pow({}, tf.scalar(1)); }) .toThrowError(/Argument 'base' passed to 'pow' must be a Tensor/); }); it('throws when passed exp as a non-tensor', function () { expect(function () { return tf.pow(tf.scalar(1), {}); }) .toThrowError(/Argument 'exp' passed to 'pow' must be a Tensor/); }); }); jasmine_util_1.describeWithFlags('add', test_util_1.ALL_ENVS, function () { it('c + A', function () { var c = tf.scalar(5); var a = tf.tensor1d([1, 2, 3]); var result = tf.add(c, a); test_util_1.expectArraysClose(result, [6, 7, 8]); }); it('c + A propagates NaNs', function () { var c = tf.scalar(NaN); var a = tf.tensor1d([1, 2, 3]); var res = tf.add(c, a); test_util_1.expectArraysEqual(res, [NaN, NaN, NaN]); }); it('A + B broadcasting same rank Tensors different shape', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor2d([2, 3], [2, 1]); var result = tf.add(a, b); expect(result.shape).toEqual([2, 2]); var expected = [3, 4, 0, -1]; test_util_1.expectArraysClose(result, expected); }); it('A + B broadcast 2D + 1D', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor1d([1, 2]); var result = tf.add(a, b); expect(result.shape).toEqual([2, 2]); var expected = [2, 4, -2, -2]; test_util_1.expectArraysClose(result, expected); }); it('A + B', function () { var a = tf.tensor1d([2, 5, 1]); var b = tf.tensor1d([4, 2, -1]); var result = tf.add(a, b); var expected = [6, 7, 0]; test_util_1.expectArraysClose(result, expected); }); it('A + B propagates NaNs', function () { var a = tf.tensor1d([2, 5, NaN]); var b = tf.tensor1d([4, 2, -1]); var res = tf.add(a, b); test_util_1.expectArraysClose(res, [6, 7, NaN]); }); it('A + B throws when passed tensors with different shape', function () { var a = tf.tensor1d([2, 5, 1, 5]); var b = tf.tensor1d([4, 2, -1]); expect(function () { return tf.add(a, b); }).toThrowError(); expect(function () { return tf.add(b, a); }).toThrowError(); }); it('2D+scalar broadcast', function () { var a = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]); var b = tf.scalar(2); var res = tf.add(a, b); expect(res.shape).toEqual([2, 3]); test_util_1.expectArraysClose(res, [3, 4, 5, 6, 7, 8]); }); it('scalar+1D broadcast', function () { var a = tf.scalar(2); var b = tf.tensor1d([1, 2, 3, 4, 5, 6]); var res = tf.add(a, b); expect(res.shape).toEqual([6]); test_util_1.expectArraysClose(res, [3, 4, 5, 6, 7, 8]); }); it('2D+2D broadcast each with 1 dim', function () { var a = tf.tensor2d([1, 2, 5], [1, 3]); var b = tf.tensor2d([7, 3], [2, 1]); var res = tf.add(a, b); expect(res.shape).toEqual([2, 3]); test_util_1.expectArraysClose(res, [8, 9, 12, 4, 5, 8]); }); it('2D+2D broadcast inner dim of b', function () { var a = tf.tensor2d([1, 2, 5, 4, 5, 6], [2, 3]); var b = tf.tensor2d([7, 3], [2, 1]); var res = tf.add(a, b); expect(res.shape).toEqual([2, 3]); test_util_1.expectArraysClose(res, [8, 9, 12, 7, 8, 9]); }); it('3D+scalar', function () { var a = tf.tensor3d([1, 2, 3, 4, 5, 6], [2, 3, 1]); var b = tf.scalar(-1); var res = tf.add(a, b); expect(res.shape).toEqual([2, 3, 1]); test_util_1.expectArraysClose(res, [0, 1, 2, 3, 4, 5]); }); it('gradient: scalar + 1D broadcast', function () { var a = tf.scalar(2); var b = tf.tensor1d([3, 4, 5]); var dy = tf.tensor1d([7, 8, 9]); var grads = tf.grads(function (a, b) { return tf.add(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [7 + 8 + 9]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [7, 8, 9]); }); it('gradient: 2D + 2D broadcast', function () { var a = tf.tensor2d([2, 3], [2, 1]); var b = tf.tensor2d([4, 5, 6, 7], [2, 2]); var dy = tf.tensor2d([5, 4, 3, 2], [2, 2]); var grads = tf.grads(function (a, b) { return tf.add(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [5 + 4, 3 + 2]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [5, 4, 3, 2]); }); it('throws when passed a as a non-tensor', function () { expect(function () { return tf.add({}, tf.scalar(1)); }) .toThrowError(/Argument 'a' passed to 'add' must be a Tensor/); }); it('throws when passed b as a non-tensor', function () { expect(function () { return tf.add(tf.scalar(1), {}); }) .toThrowError(/Argument 'b' passed to 'add' must be a Tensor/); }); }); jasmine_util_1.describeWithFlags('sub', test_util_1.ALL_ENVS, function () { it('c - A', function () { var c = tf.scalar(5); var a = tf.tensor1d([7, 2, 3]); var result = tf.sub(c, a); test_util_1.expectArraysClose(result, [-2, 3, 2]); }); it('A - c', function () { var a = tf.tensor1d([1, 2, -3]); var c = tf.scalar(5); var result = tf.sub(a, c); test_util_1.expectArraysClose(result, [-4, -3, -8]); }); it('A - c propagates NaNs', function () { var a = tf.tensor1d([1, NaN, 3]); var c = tf.scalar(5); var res = tf.sub(a, c); test_util_1.expectArraysClose(res, [-4, NaN, -2]); }); it('A - B', function () { var a = tf.tensor1d([2, 5, 1]); var b = tf.tensor1d([4, 2, -1]); var result = tf.sub(a, b); var expected = [-2, 3, 2]; test_util_1.expectArraysClose(result, expected); }); it('A - B propagates NaNs', function () { var a = tf.tensor1d([2, 5, 1]); var b = tf.tensor1d([4, NaN, -1]); var res = tf.sub(a, b); test_util_1.expectArraysClose(res, [-2, NaN, 2]); }); it('A - B throws when passed tensors with different shape', function () { var a = tf.tensor1d([2, 5, 1, 5]); var b = tf.tensor1d([4, 2, -1]); expect(function () { return tf.sub(a, b); }).toThrowError(); expect(function () { return tf.sub(b, a); }).toThrowError(); }); it('A - B broadcasting same rank Tensors different shape', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor2d([2, 3], [2, 1]); var result = tf.sub(a, b); expect(result.shape).toEqual([2, 2]); var expected = [-1, 0, -6, -7]; test_util_1.expectArraysClose(result, expected); }); it('A - B broadcast 2D + 1D', function () { var a = tf.tensor2d([1, 2, -3, -4], [2, 2]); var b = tf.tensor1d([1, 2]); var result = tf.sub(a, b); expect(result.shape).toEqual([2, 2]); var expected = [0, 0, -4, -6]; test_util_1.expectArraysClose(result, expected); }); it('2D-scalar broadcast', function () { var a = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]); var b = tf.scalar(2); var res = tf.sub(a, b); expect(res.shape).toEqual([2, 3]); test_util_1.expectArraysClose(res, [-1, 0, 1, 2, 3, 4]); }); it('scalar-1D broadcast', function () { var a = tf.scalar(2); var b = tf.tensor1d([1, 2, 3, 4, 5, 6]); var res = tf.sub(a, b); expect(res.shape).toEqual([6]); test_util_1.expectArraysClose(res, [1, 0, -1, -2, -3, -4]); }); it('2D-2D broadcast each with 1 dim', function () { var a = tf.tensor2d([1, 2, 5], [1, 3]); var b = tf.tensor2d([7, 3], [2, 1]); var res = tf.sub(a, b); expect(res.shape).toEqual([2, 3]); test_util_1.expectArraysClose(res, [-6, -5, -2, -2, -1, 2]); }); it('2D-2D broadcast inner dim of b', function () { var a = tf.tensor2d([1, 2, 5, 4, 5, 6], [2, 3]); var b = tf.tensor2d([7, 3], [2, 1]); var res = tf.sub(a, b); expect(res.shape).toEqual([2, 3]); test_util_1.expectArraysClose(res, [-6, -5, -2, 1, 2, 3]); }); it('3D-scalar', function () { var a = tf.tensor3d([1, 2, 3, 4, 5, 6], [2, 3, 1]); var b = tf.scalar(-1); var res = tf.sub(a, b); expect(res.shape).toEqual([2, 3, 1]); test_util_1.expectArraysClose(res, [2, 3, 4, 5, 6, 7]); }); it('gradients: basic 1D arrays', function () { var a = tf.tensor1d([1, 2, 3]); var b = tf.tensor1d([3, 2, 1]); var dy = tf.tensor1d([1, 10, 20]); var grads = tf.grads(function (a, b) { return tf.sub(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [1, 10, 20]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-1, -10, -20]); }); it('gradients: basic 2D arrays', function () { var a = tf.tensor2d([0, 1, 2, 3], [2, 2]); var b = tf.tensor2d([3, 2, 1, 0], [2, 2]); var dy = tf.tensor2d([1, 10, 15, 20], [2, 2]); var grads = tf.grads(function (a, b) { return tf.sub(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [1, 10, 15, 20]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-1, -10, -15, -20]); }); it('gradient: 1D - scalar broadcast', function () { var a = tf.tensor1d([3, 4, 5]); var b = tf.scalar(2); var dy = tf.tensor1d([7, 8, 9]); var grads = tf.grads(function (a, b) { return tf.sub(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [7, 8, 9]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-7 - 8 - 9]); }); it('gradient: scalar - 1D broadcast', function () { var a = tf.scalar(2); var b = tf.tensor1d([3, 4, 5]); var dy = tf.tensor1d([7, 8, 9]); var grads = tf.grads(function (a, b) { return tf.sub(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [7 + 8 + 9]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-7, -8, -9]); }); it('gradient: 2D - 2D broadcast', function () { var a = tf.tensor2d([4, 5, 6, 7], [2, 2]); var b = tf.tensor2d([2, 3], [2, 1]); var dy = tf.tensor2d([5, 4, 3, 2], [2, 2]); var grads = tf.grads(function (a, b) { return tf.sub(a, b); }); var _a = grads([a, b], dy), da = _a[0], db = _a[1]; expect(da.shape).toEqual(a.shape); expect(da.dtype).toEqual('float32'); test_util_1.expectArraysClose(da, [5, 4, 3, 2]); expect(db.shape).toEqual(b.shape); expect(db.dtype).toEqual('float32'); test_util_1.expectArraysClose(db, [-5 - 4, -3 - 2]); }); it('throws when passed a as a non-tensor', function () { expect(function () { return tf.sub({}, tf.scalar(1)); }) .toThrowError(/Argument 'a' passed to 'sub' must be a Tensor/); }); it('throws when passed b as a non-tensor', function () { expect(function () { return tf.sub(tf.scalar(1), {}); }) .toThrowError(/Argument 'b' passed to 'sub' must be a Tensor/); }); });