mobilenetv3的block

### MobileNetV3 Block 结构和实现详解 #### 倒残差瓶颈 (Inverted Residual Bottleneck) MobileNetV3 的基础构建单元是倒残差瓶颈(Inverted Residuals)，这种设计继承自 MobileNetV2 并进行了改进。在每个 bottleneck 中，输入特征图首先通过逐点卷积(point-wise convolution)扩展到更高的维度，接着经过深度可分离卷积(depthwise separable convolution)，最后再压缩回较低的维度。 ```python def inverted_residual_block(inputs, expansion, filters, kernel_size, stride, se_ratio=None): shortcut = inputs # Pointwise Convolution to expand channels x = tf.keras.layers.Conv2D(expansion * int(inputs.shape[-1]), 1, padding='same')(inputs) # Depthwise Separable Convolution x = tf.keras.layers.DepthwiseConv2D(kernel_size=kernel_size, strides=stride, padding='same')(x) # Squeeze-and-Excitation Module if specified if se_ratio is not None: x = squeeze_excite(x, ratio=se_ratio) # Linear pointwise convolution x = tf.keras.layers.Conv2D(filters=filters, kernel_size=(1, 1), activation=None)(x) if all(s == 1 for s in stride) and int(shortcut.shape[-1]) == filters: return tf.keras.layers.Add()([shortcut, x]) else: return x ``` #### 挤压激励(Squeeze-and-Excitation, SE)模块 SE 模块被引入来增强模型的感受野并改善信息流动效率[^1]。具体来说，在完成一次完整的bottleneck操作之后，会计算全局平均池化(global average pooling)，从而获得各个通道上的统计特性；随后利用两个全连接层分别降低和恢复维度，最终产生的权重用于重新加权原始特征图中的每一通道。 ```python def squeeze_excite(tensor_input, ratio=0.25): num_channels = tensor_input.shape[-1] # Global Average Pooling across spatial dimensions pooled_features = tf.reduce_mean(tensor_input, axis=[1, 2], keepdims=True) # Fully connected layers with reduced number of units squeezed = tf.keras.layers.Dense(int(num_channels*ratio))(pooled_features) excited = tf.keras.layers.ReLU()(squeezed) expanded = tf.keras.layers.Dense(num_channels)(excited) scaled = tf.nn.sigmoid(expanded)*tensor_input return scaled ``` #### Hard-Swish 激活函数 为了进一步提升性能表现，MobileNetV3采用了新的激活函数——Hard-Swish替代传统的ReLU6作为默认选项之一[^3]。相比于后者，前者能够在保持轻量化的同时提供更好的表达能力。 ```python import tensorflow as tf @tf.function def hard_swish(x): return x * tf.nn.relu6(x + 3.) / 6. ```