netty源码分析之线程模型

MultithreadEventExecutorGroup里面是EventExecutor的数组,该类的构造器最数组进行初始化,初始化过程交给newChild抽象方法

该抽象方法是由具体实现,下面是NioEventLoopGroup的具体实现如下

 @Override
    protected EventLoop newChild(Executor executor, Object... args) throws Exception {
        return new NioEventLoop(this, executor, (SelectorProvider) args[0],
            ((SelectStrategyFactory) args[1]).newSelectStrategy(), (RejectedExecutionHandler) args[2]);
    }

NioEventLoopGroup其实就是一组NioEventLoop的管理,具体执行执行任务时都是调用相应的某个EventExecutor实现类来执行

那这里到底选择哪个EventExecutor来执行呢?原来MultithreadEventExecutorGroup有个选择器chooser,netty支持选择器如下


/**
 * Default implementation which uses simple round-robin to choose next {@link EventExecutor}.
 *默认的实现是是简单的轮询选择器
 */
@UnstableApi
public final class DefaultEventExecutorChooserFactory implements EventExecutorChooserFactory {

    public static final DefaultEventExecutorChooserFactory INSTANCE = new DefaultEventExecutorChooserFactory();

    private DefaultEventExecutorChooserFactory() { }

    @SuppressWarnings("unchecked")
    @Override
    public EventExecutorChooser newChooser(EventExecutor[] executors) {
        if (isPowerOfTwo(executors.length)) {
            return new PowerOfTwoEventExecutorChooser(executors);
        } else {
            return new GenericEventExecutorChooser(executors);
        }
    }

    private static boolean isPowerOfTwo(int val) {
        return (val & -val) == val;
    }

    private static final class PowerOfTwoEventExecutorChooser implements EventExecutorChooser {
        private final AtomicInteger idx = new AtomicInteger();
        private final EventExecutor[] executors;

        PowerOfTwoEventExecutorChooser(EventExecutor[] executors) {
            this.executors = executors;
        }

        @Override
        public EventExecutor next() {
            return executors[idx.getAndIncrement() & executors.length - 1];
        }
    }

    private static final class GenericEventExecutorChooser implements EventExecutorChooser {
        private final AtomicInteger idx = new AtomicInteger();
        private final EventExecutor[] executors;

        GenericEventExecutorChooser(EventExecutor[] executors) {
            this.executors = executors;
        }

        @Override
        public EventExecutor next() {
            return executors[Math.abs(idx.getAndIncrement() % executors.length)];
        }
    }
}

该类是EventExecutorChooser的工厂类,根据数组大小,当数量是2的n次方的时候,用PowerOfTwoEventExecutorChooser,其他用GenericEventExecutorChooser。二个都是简单轮询的方式,只是当数量是2的n次方的时候可以用&的方式比取模的时候性能高而已。

### 使用Transformer模型进行图像分类的方法 #### 方法概述 为了使Transformer能够应用于图像分类任务,一种有效的方式是将图像分割成固定大小的小块(patches),这些小块被线性映射为向量,并加上位置编码以保留空间信息[^2]。 #### 数据预处理 在准备输入数据的过程中,原始图片会被切分成多个不重叠的patch。假设一张尺寸为\(H \times W\)的RGB图像是要处理的对象,则可以按照设定好的宽度和高度参数来划分该图像。例如,对于分辨率为\(224\times 224\)像素的图像,如果选择每边切成16个部分的话,那么最终会得到\((224/16)^2=196\)个小方格作为单独的特征表示单元。之后,每一个这样的补丁都会通过一个简单的全连接层转换成为维度固定的嵌入向量。 ```python import torch from torchvision import transforms def preprocess_image(image_path, patch_size=16): transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), # 假设目标分辨率是224x224 transforms.ToTensor(), ]) image = Image.open(image_path).convert('RGB') tensor = transform(image) patches = [] for i in range(tensor.shape[-2] // patch_size): # 高度方向上的循环 row_patches = [] for j in range(tensor.shape[-1] // patch_size): # 宽度方向上的循环 patch = tensor[:, :, i*patch_size:(i+1)*patch_size, j*patch_size:(j+1)*patch_size].flatten() row_patches.append(patch) patches.extend(row_patches) return torch.stack(patches) ``` #### 构建Transformer架构 构建Vision Transformer (ViT),通常包括以下几个组成部分: - **Patch Embedding Layer**: 将每个图像块转化为低维向量; - **Positional Encoding Layer**: 添加绝对或相对位置信息给上述获得的向量序列; - **Multiple Layers of Self-Attention and Feed Forward Networks**: 多层自注意机制与前馈神经网络交替堆叠而成的核心模块; 最后,在顶层附加一个全局平均池化层(Global Average Pooling)以及一个多类别Softmax回归器用于预测类标签。 ```python class VisionTransformer(nn.Module): def __init__(self, num_classes=1000, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=False, drop_rate=0.): super().__init__() self.patch_embed = PatchEmbed(embed_dim=embed_dim) self.pos_embed = nn.Parameter(torch.zeros(1, self.patch_embed.num_patches + 1, embed_dim)) self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) dpr = [drop_rate for _ in range(depth)] self.blocks = nn.Sequential(*[ Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, drop=dpr[i], ) for i in range(depth)]) self.norm = nn.LayerNorm(embed_dim) self.head = nn.Linear(embed_dim, num_classes) def forward(self, x): B = x.shape[0] cls_tokens = self.cls_token.expand(B, -1, -1) x = self.patch_embed(x) x = torch.cat((cls_tokens, x), dim=1) x += self.pos_embed x = self.blocks(x) x = self.norm(x) return self.head(x[:, 0]) ```
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值