deep wise convolution

Deep-wise 卷积是一种卷积神经网络中的特殊卷积操作。它是将卷积操作分为两个步骤：可分离卷积（Separable Conv）和点卷积（Point-wise Conv）。

简单来说，Deep-wise 卷积相对于传统的卷积操作，具有更少的浮点运算（FLOPs）。然而，在相同的计算复杂度（FLOPs）下，Deep-wise 卷积需要更多的输入输出（IO）读取次数。因此，尽管Deep-wise 卷积的尺寸较小，可以在相同的显存下使用更大的批次大小来充分利用GPU，但是速度的瓶颈已经由计算转变为IO。因此，Deep-wise 卷积无法实现传统卷积所具备的快速特性。

对于传统的卷积操作，例如对于一个输入特征图为[12,12,3]，输出特征图为[8,8,256]的卷积操作，需要使用256个[5,5,3]的卷积核。参数量为256 x 5 x 5 x 3 = 19200，乘法次数为256 x 5 x 5 x 3 x 8 x 8 = 1228800（也可以理解为FLOPs）。

相关问题

group-wise convolution

Group-wise convolution（群卷积或分组卷积）是一种卷积方法，它将输入的特征图在通道维度上进行分组，然后对每个分组分别进行卷积操作。这种方法可以减少计算量，并且在一定程度上提高模型的表达能力。在group-wise convolution中，输入特征图被分为多个子组，每个子组包含一部分输入特征图的通道。然后，对每个子组分别使用卷积核进行卷积操作，得到对应的输出子组。最后，将所有输出子组合并得到最终的输出特征图。这种方法通常用于深度卷积神经网络中，可以提升模型的参数效率和计算效率。<span class="em">1</span><span class="em">2</span><span class="em">3</span>
#### 引用[.reference_title]
- *1* *2* [深度学习之各种类型的卷积运算](https://blog.csdn.net/DreamHome_S/article/details/85233173)[target="_blank" data-report-click={"spm":"1018.2226.3001.9630","extra":{"utm_source":"vip_chatgpt_common_search_pc_result","utm_medium":"distribute.pc_search_result.none-task-cask-2~all~insert_cask~default-1-null.142^v93^chatsearchT3_1"}}] [.reference_item style="max-width: 50%"]
- *3* [group convolution (分组卷积)的计算量详解、卷积计算量特征图大小，池化特征图大小、深度通道deep-wise ...](https://blog.csdn.net/weixin_46838716/article/details/125649148)[target="_blank" data-report-click={"spm":"1018.2226.3001.9630","extra":{"utm_source":"vip_chatgpt_common_search_pc_result","utm_medium":"distribute.pc_search_result.none-task-cask-2~all~insert_cask~default-1-null.142^v93^chatsearchT3_1"}}] [.reference_item style="max-width: 50%"]
[ .reference_list ]

Axial-DeepLab

Axial-DeepLab is a deep learning architecture for semantic segmentation tasks that was introduced in 2019. It is based on the popular DeepLab architecture, which uses atrous convolution (also known as dilated convolution) to capture multi-scale contextual information for improved segmentation accuracy.

Axial-DeepLab extends DeepLab by incorporating axial attention modules, which enable the network to focus on relevant features in the axial (i.e., spatial) dimensions of the input image. This allows the network to better handle objects with elongated or thin structures, such as roads, trees, and poles.

Axial-DeepLab also incorporates a multi-scale fusion module that combines features from different scales to improve the accuracy of the segmentation. Additionally, it uses a depth-wise separable convolutional block to reduce the number of parameters and improve computational efficiency.

Overall, Axial-DeepLab achieves state-of-the-art performance on several benchmark datasets for semantic segmentation, including PASCAL VOC and Cityscapes.