Dynamic Convolutions: Exploiting Spatial Sparsity for Faster Inference

Começar. É Gratuito
ou inscrever-se com seu endereço de e-mail
Dynamic Convolutions: Exploiting Spatial Sparsity for Faster Inference por Mind Map: Dynamic Convolutions: Exploiting Spatial Sparsity for Faster Inference

1. Abstract

1.1. 传统的卷积神经网络操作对一张图片里的每个pixel一视同仁,然鹅,图片里并不是所有的region都是同样重要的。因此,本文作者提出了一种针对输入图像动态的有条件的卷积方法。另外,作者还引入了一种针对cuda的有效的动态卷积实现方法,利用的是gater-scatter方法。

1.2. Method1:

1.2.1. 引入了residual block,一个小的gating 分支,用来学习哪个空间位置需要被evaluate。

1.2.1.1. 这个离散的gating decision 是通过端到端训出来的,并且用到了一个Gumbel-Softmax的小trick,和稀疏规则合并起来。

1.2.1.1.1. 数据集

1.3. Methord2:

1.3.1. 为mobilenetV2和shufflenetv2提速,在人体姿态估计上做了实验

2. Introduction

2.1. 在本文中,作者主要focus在通过对空间域有条件的执行来减小计算量。作者提出一种方法,是可以端到端训练的,并且没有给出的清晰的spatial supervisor。让卷积核只在重要的image区域上执行。对于每一个residual block,有一个小的gating network用来生成mask,应用到dynamic卷积上去。Gate Decision用的是Gumbel-Softmax trick来训的。ps. 需要注意的是,dynamic convolution 的input是一个dense matrics

3. Method

3.1. 首先,描述pixel-wise masks如何使用Gumbel-softmax trick 去学习

3.1.1. Trainable Mask

3.1.1.1. Block Arch

3.1.1.1.1. The operation of the residual block

3.1.1.1.2. a binary Gumbel-Softmax trick on each element of Mb

3.1.1.1.3. The residual block with sparsity mask

3.1.1.2. Binary Gumbel-Softmax

3.1.1.2.1. Discrete samples Z canbe drawn using

3.1.1.2.2. use Gumbel-Softmax trick to defines acontinuous differentialbe approximation by replacing the argmax operation with a softmax

3.1.1.2.3. converted to a probability pi1 indicating the probability that a pixel should be executed, using a sigmoid function.

3.1.1.2.4. The probability that a pixel is not executed

3.1.1.2.5. substituting pi1 and pi2 in equation 5

3.1.1.2.6. use a straight-through estimator, where hard samples are used during the forward pass and gradients are obtained from soft samples during the backward pass

3.2. 如何实现dynamic convolutions 去减小inference time

3.2.1. Efficient inference implementation

3.2.1.1. mask dilation

3.2.1.2. masked gather operation

3.2.1.3. modified 3*3 depthwise convolution

3.3. 最后提出一种sparsity criterion的标准