Closed JinhangZhu closed 4 years ago
[x] 学习莫烦的PyTorch教程基础
[x] 上手YOLOv3检测过程的代码
3.30-4.2
考虑到第一次用官方的PyTorch入门文档根本就没怎么懂神经网络的构建方法,所以重新按照莫烦Python的教程,从脚踏实地的coding直接开始。
PyTorch基础,包括1) tensor和numpy转换,2) tensor的Variable(但是现版本不需要使用Variable API,见[Variable (deprecated)]()https://pytorch.org/docs/stable/autograd.html#variable-deprecated),3) 主要的激活函数。
Regression & Classification和一般的神经网络搭建。这一节挺重要,直接把我拉入神经网络coding的大门,值得反复回看。
MNIST手写数字的卷积神经网络,重点为:
4.3
4.4
yolov3.cfg
module_list
module
Darknet
__init__()
forward()
torch
4.5
torch.Size([1, 10647, 85])
BBox per cell -> index in (batch, number of cells * number of anchors per cell, number of attributes) 1st BBox -> (0, 000, 5+num_classes) 2nd BBox -> (0, 001, 5+num_classes) 3rd BBox -> (0, 002, 5+num_classes) 1st BBox -> (0, 010, 5+num_classes) 2nd BBox -> (0, 011, 5+num_classes) 3rd BBox -> (0, 012, 5+num_classes) ... 1st BBox -> (bs-1, [gs-1][gs-1]0, 5+num_classes) 2nd BBox -> (bs-1, [gs-1][gs-1]1, 5+num_classes) 3rd BBox -> (bs-1, [gs-1][gs-1]2, 5+num_classes)
How to get anchors from a dataset?
Done
[x] 学习莫烦的PyTorch教程基础
[x] 上手YOLOv3检测过程的代码
3.30-4.2
考虑到第一次用官方的PyTorch入门文档根本就没怎么懂神经网络的构建方法,所以重新按照莫烦Python的教程,从脚踏实地的coding直接开始。
PyTorch基础,包括1) tensor和numpy转换,2) tensor的Variable(但是现版本不需要使用Variable API,见[Variable (deprecated)]()https://pytorch.org/docs/stable/autograd.html#variable-deprecated),3) 主要的激活函数。
Regression & Classification和一般的神经网络搭建。这一节挺重要,直接把我拉入神经网络coding的大门,值得反复回看。
MNIST手写数字的卷积神经网络,重点为:
4.3
4.4
yolov3.cfg
文件,建立一个module_list
,包括每一层的module
实体,共106层,与论文👇一致。Darknet
的类定义,包括__init__()
和forward()
两个函数。其实整个网络能直接给输入按照torch
给定的功能输出的就convolutional卷积层和上upsample采样层,其余层(route layer, shortcut layer, yolo/detection layer)都是需要自己写的。4.5
torch.Size([1, 10647, 85])
。Next