soft_argmax

https://github.com/Fdevmsy/PyTorch-Soft-Argmax/blob/master/soft-argmax.py

 

在姿态估计算法中，微软开发的 SimpleBaseline 是精度高而又轻量级的典范，昨天一篇论文Simple and Lightweight Human Pose Estimation，在该架构基础上做了少许改进，取得了更快的速度和更小的模型Size。

 

论文作者信息：

import torch
import torch.nn as nn

def soft_argmax(voxels):
	"""
	Arguments: voxel patch in shape (batch_size, channel, H, W, depth)
	Return: 3D coordinates in shape (batch_size, channel, 3)
	"""
	assert voxels.dim()==5
	# alpha is here to make the largest element really big, so it
	# would become very close to 1 after softmax
	alpha = 1000.0 
	N,C,H,W,D = voxels.shape
	soft_max = nn.functional.softmax(voxels.v