TensorFlow.js 张量基础 #

什么是张量？ #

张量是 TensorFlow.js 中的核心数据结构，它是多维数组的泛化形式。张量可以表示标量、向量、矩阵以及更高维度的数据。

text

┌─────────────────────────────────────────────────────────────┐
│                      张量的维度                               │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  标量 (0D)     向量 (1D)      矩阵 (2D)       3D 张量       │
│                                                             │
│    [5]        [1,2,3]     [[1,2,3],      [[[1,2],          │
│                              [4,5,6]]       [3,4]],         │
│                                             [[5,6],         │
│                                              [7,8]]]        │
│                                                             │
│  shape: []    shape: [3]   shape: [2,3]    shape: [2,2,2]  │
│                                                             │
└─────────────────────────────────────────────────────────────┘

张量的属性 #

每个张量都有以下核心属性：

数据类型（dtype） #

javascript

const intTensor = tf.tensor([1, 2, 3]);
console.log(intTensor.dtype);

const floatTensor = tf.tensor([1.5, 2.5, 3.5]);
console.log(floatTensor.dtype);

const boolTensor = tf.tensor([true, false, true], [3], 'bool');
console.log(boolTensor.dtype);

形状（shape） #

javascript

const tensor = tf.tensor([[1, 2, 3], [4, 5, 6]]);
console.log(tensor.shape);

const tensor3d = tf.tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]]);
console.log(tensor3d.shape);

维度（rank） #

javascript

const scalar = tf.scalar(5);
console.log(scalar.rank);

const vector = tf.tensor1d([1, 2, 3]);
console.log(vector.rank);

const matrix = tf.tensor2d([[1, 2], [3, 4]]);
console.log(matrix.rank);

大小（size） #

javascript

const tensor = tf.tensor([[1, 2, 3], [4, 5, 6]]);
console.log(tensor.size);

创建张量 #

基本创建方法 #

tf.tensor() #

javascript

const t1 = tf.tensor([1, 2, 3, 4]);
t1.print();

const t2 = tf.tensor([1, 2, 3, 4], [2, 2]);
t2.print();

const t3 = tf.tensor([1, 2, 3, 4, 5, 6, 7, 8], [2, 2, 2]);
t3.print();

指定维度的创建方法 #

javascript

const scalar = tf.scalar(5);
scalar.print();

const vector = tf.tensor1d([1, 2, 3]);
vector.print();

const matrix = tf.tensor2d([[1, 2], [3, 4]]);
matrix.print();

const tensor3d = tf.tensor3d([[[1, 2], [3, 4]], [[5, 6], [7, 8]]]);
tensor3d.print();

const tensor4d = tf.tensor4d([[[[1, 2], [3, 4]], [[5, 6], [7, 8]]]]);
tensor4d.print();

特殊值张量 #

零张量 #

javascript

const zeros = tf.zeros([3, 3]);
zeros.print();

const zerosLike = tf.zerosLike(tf.tensor([1, 2, 3]));
zerosLike.print();

一张量 #

javascript

const ones = tf.ones([2, 3]);
ones.print();

const onesLike = tf.onesLike(tf.tensor([[1, 2], [3, 4]]));
onesLike.print();

填充张量 #

javascript

const filled = tf.fill([2, 3], 7);
filled.print();

单位矩阵 #

javascript

const eye = tf.eye(3);
eye.print();

随机张量 #

均匀分布随机 #

javascript

const random = tf.randomUniform([3, 3], 0, 1);
random.print();

正态分布随机 #

javascript

const normal = tf.randomNormal([3, 3], 0, 1);
normal.print();

截断正态分布 #

javascript

const truncated = tf.truncatedNormal([3, 3], 0, 1);
truncated.print();

序列张量 #

等差序列 #

javascript

const range = tf.range(0, 10, 2);
range.print();

const linspace = tf.linspace(0, 1, 5);
linspace.print();

从其他数据创建 #

从 DOM 元素 #

javascript

const img = document.getElementById('myImage');
const tensor = tf.browser.fromPixels(img);
tensor.print();

从像素数据 #

javascript

const canvas = document.getElementById('myCanvas');
const ctx = canvas.getContext('2d');
const imageData = ctx.getImageData(0, 0, 100, 100);
const tensor = tf.browser.fromPixels(imageData);

张量操作 #

形状变换 #

reshape #

javascript

const t = tf.tensor([1, 2, 3, 4, 5, 6]);
const reshaped = t.reshape([2, 3]);
reshaped.print();

flatten #

javascript

const t = tf.tensor([[1, 2, 3], [4, 5, 6]]);
const flattened = t.flatten();
flattened.print();

expandDims #

javascript

const t = tf.tensor([1, 2, 3]);
const expanded = t.expandDims(0);
console.log(expanded.shape);

squeeze #

javascript

const t = tf.tensor([[[1, 2, 3]]]);
const squeezed = t.squeeze();
console.log(squeezed.shape);

transpose #

javascript

const t = tf.tensor([[1, 2, 3], [4, 5, 6]]);
const transposed = t.transpose();
transposed.print();

切片与索引 #

slice #

javascript

const t = tf.tensor([1, 2, 3, 4, 5, 6]);
const sliced = t.slice([2], [3]);
sliced.print();

const t2d = tf.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]]);
const sliced2d = t2d.slice([1, 1], [2, 2]);
sliced2d.print();

stridedSlice #

javascript

const t = tf.tensor([1, 2, 3, 4, 5, 6, 7, 8]);
const strided = t.stridedSlice([1], [7], [2]);
strided.print();

gather #

javascript

const t = tf.tensor([[1, 2], [3, 4], [5, 6]]);
const gathered = t.gather([0, 2]);
gathered.print();

拼接与分割 #

concat #

javascript

const a = tf.tensor([[1, 2]]);
const b = tf.tensor([[3, 4]]);
const concatenated = a.concat(b, 0);
concatenated.print();

split #

javascript

const t = tf.tensor([1, 2, 3, 4, 5, 6]);
const [a, b] = tf.split(t, 2);
a.print();
b.print();

stack #

javascript

const a = tf.tensor([1, 2, 3]);
const b = tf.tensor([4, 5, 6]);
const stacked = tf.stack([a, b]);
stacked.print();

unstack #

javascript

const t = tf.tensor([[1, 2, 3], [4, 5, 6]]);
const [a, b] = tf.unstack(t);
a.print();
b.print();

填充与复制 #

pad #

javascript

const t = tf.tensor([[1, 2], [3, 4]]);
const padded = t.pad([[1, 1], [1, 1]]);
padded.print();

tile #

javascript

const t = tf.tensor([1, 2]);
const tiled = t.tile([3]);
tiled.print();

repeat #

javascript

const t = tf.tensor([1, 2, 3]);
const repeated = t.repeat(2);
repeated.print();

张量数据获取 #

同步获取 #

javascript

const t = tf.tensor([1, 2, 3, 4]);
const data = t.dataSync();
console.log(data);

const array = t.arraySync();
console.log(array);

异步获取 #

javascript

const t = tf.tensor([1, 2, 3, 4]);
t.data().then(data => {
  console.log(data);
});

t.array().then(array => {
  console.log(array);
});

获取单个值 #

javascript

const t = tf.tensor([1, 2, 3, 4]);
const value = t.dataSync()[0];
console.log(value);

const scalar = tf.scalar(5);
const scalarValue = scalar.dataSync()[0];
console.log(scalarValue);

变量（Variable） #

变量是可变的张量，用于存储模型参数。

创建变量 #

javascript

const initial = tf.zeros([3, 3]);
const variable = tf.variable(initial);
variable.print();

更新变量 #

javascript

const variable = tf.variable(tf.zeros([3]));
variable.assign(tf.tensor([1, 2, 3]));
variable.print();

变量属性 #

javascript

const variable = tf.variable(tf.zeros([3, 3]));
console.log(variable.shape);
console.log(variable.dtype);
console.log(variable.name);

内存管理 #

TensorFlow.js 使用 WebGL 纹理存储张量，需要手动管理内存。

手动释放 #

javascript

const t = tf.tensor([1, 2, 3]);
t.dispose();

tf.tidy 自动清理 #

javascript

const result = tf.tidy(() => {
  const a = tf.tensor([1, 2, 3]);
  const b = tf.tensor([4, 5, 6]);
  const c = a.add(b);
  return c;
});
result.print();

tf.dispose 批量释放 #

javascript

const a = tf.tensor([1, 2, 3]);
const b = tf.tensor([4, 5, 6]);
tf.dispose([a, b]);

内存监控 #

javascript

console.log(tf.memory());

内存管理最佳实践 #

text

┌─────────────────────────────────────────────────────────────┐
│                    内存管理最佳实践                           │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  1. 使用 tf.tidy 包装中间计算                                │
│                                                             │
│     const result = tf.tidy(() => {                         │
│       const a = tf.tensor([1, 2, 3]);                      │
│       const b = a.square();                                │
│       return b;                                            │
│     });                                                    │
│                                                             │
│  2. 及时释放不需要的张量                                     │
│                                                             │
│     const t = expensiveOperation();                        │
│     useTensor(t);                                          │
│     t.dispose();                                           │
│                                                             │
│  3. 避免在循环中创建张量                                     │
│                                                             │
│     for (let i = 0; i < 1000; i++) {                       │
│       tf.tidy(() => {                                      │
│         const t = tf.tensor(i);                            │
│       });                                                  │
│     }                                                      │
│                                                             │
│  4. 监控内存使用                                            │
│                                                             │
│     setInterval(() => {                                    │
│       console.log(tf.memory());                            │
│     }, 1000);                                              │
│                                                             │
└─────────────────────────────────────────────────────────────┘

张量打印与可视化 #

print() #

javascript

const t = tf.tensor([[1, 2, 3], [4, 5, 6]]);
t.print();
t.print(true);

toString() #

javascript

const t = tf.tensor([1, 2, 3]);
console.log(t.toString());

浏览器渲染 #

javascript

const t = tf.randomNormal([100, 100]);
await tf.browser.toPixels(t, document.getElementById('canvas'));

张量类型转换 #

cast #

javascript

const t = tf.tensor([1.5, 2.7, 3.9]);
const intT = t.cast('int32');
intT.print();

asType #

javascript

const t = tf.tensor([1, 2, 3]);
const floatT = t.asType('float32');

实用工具函数 #

获取张量信息 #

javascript

const t = tf.tensor([[1, 2, 3], [4, 5, 6]]);

console.log('Shape:', t.shape);
console.log('Rank:', t.rank);
console.log('Size:', t.size);
console.log('Dtype:', t.dtype);

判断是否为张量 #

javascript

const t = tf.tensor([1, 2, 3]);
console.log(t instanceof tf.Tensor);
console.log(tf.util.isTensor(t));

克隆张量 #

javascript

const original = tf.tensor([1, 2, 3]);
const cloned = original.clone();
cloned.print();

下一步 #

现在你已经掌握了张量的基础知识，接下来学习张量运算，了解如何对张量进行数学运算！