Operators API

神经网络层和激活函数。

---

基类

Operator

所有算子的抽象基类。

abstract class Operator {
  constructor(protected context: GPUContext);
  abstract computeOutputShape(inputShape: TensorShape, params?: OperatorParams): TensorShape;
  abstract forward(inputs: Tensor[], params?: OperatorParams): Promise<Tensor>;
  protected abstract compileShader(): string;
  destroy(): void;
}

---

卷积算子

Conv2dOperator

2D 卷积操作（NCHW 布局）。

import { Conv2dOperator } from 'tiny-dl-inference';

const conv = new Conv2dOperator(context);
const output = await conv.forward([input, weights]);

参数：

interface Conv2dParams {
  stride: [number, number];
  padding: [number, number];
}

形状：

• 输入：[batch, inChannels, inHeight, inWidth]
• 权重：[outChannels, inChannels, kernelHeight, kernelWidth]
• 输出：[batch, outChannels, outHeight, outWidth]

---

Conv2dBiasReLUOperator

融合卷积、偏置和 ReLU 激活。

import { Conv2dBiasReLUOperator } from 'tiny-dl-inference';

const convBiasRelu = new Conv2dBiasReLUOperator(context);
const output = await convBiasRelu.forward([input, weights, bias]);

性能： 比分离算子快约 1.8 倍。

---

池化算子

MaxPoolOperator

2D 最大池化。

import { MaxPoolOperator } from 'tiny-dl-inference';

const pool = new MaxPoolOperator(context);
const output = await pool.forward([input], {
  kernelSize: [2, 2],
  stride: [2, 2],
  padding: [0, 0]
});

参数：

interface MaxPoolParams {
  kernelSize: [number, number];
  stride: [number, number];
  padding: [number, number];
}

---

激活函数

ReLUOperator

修正线性单元激活。

import { ReLUOperator } from 'tiny-dl-inference';

const relu = new ReLUOperator(context);
const output = await relu.forward([input]);

公式： f(x) = max(0, x)

---

SoftmaxOperator

Softmax 归一化。

import { SoftmaxOperator } from 'tiny-dl-inference';

const softmax = new SoftmaxOperator(context);
const output = await softmax.forward([input]);

公式： f(x_i) = exp(x_i) / sum(exp(x_j))

---

全连接层

DenseOperator

全连接层（线性层）。

import { DenseOperator } from 'tiny-dl-inference';

const dense = new DenseOperator(context);
const output = await dense.forward([input, weights, bias]);

参数：

interface DenseParams {
  useBias: boolean;
}

---

形状操作

FlattenOperator

展平张量为二维。

import { FlattenOperator } from 'tiny-dl-inference';

const flatten = new FlattenOperator(context);
const output = await flatten.forward([input]);

示例：

• 输入：[2, 3, 4, 5]
• 输出：[2, 60]

---

使用示例

构建简单网络

// Conv → ReLU → MaxPool → Flatten → Dense
const conv = new Conv2dOperator(context);
const relu = new ReLUOperator(context);
const pool = new MaxPoolOperator(context);
const flatten = new FlattenOperator(context);
const dense = new DenseOperator(context);

let x = input;
x = await conv.forward([x, convWeights], { stride: [1, 1], padding: [0, 0] });
x = await relu.forward([x]);
x = await pool.forward([x], { kernelSize: [2, 2], stride: [2, 2], padding: [0, 0] });
x = await flatten.forward([x]);
x = await dense.forward([x, denseWeights, denseBias]);

const result = await x.download();

---

相关资源

• 算子指南 - 详细使用说明
• 内核融合 - 组合算子优化
• 自定义算子 - 创建新算子