在搞目標(biāo)檢測(cè)的時(shí)候，對(duì)目標(biāo)檢測(cè)的數(shù)據(jù)集處理有多疑惑。

比如：

目標(biāo)數(shù)據(jù)是如何產(chǎn)生的，也就是groundtruth是如何誕生的。

在讀取的過(guò)程中，數(shù)據(jù)集中又是如何展現(xiàn)的

在計(jì)算loss的過(guò)程中，又是如何進(jìn)行計(jì)算的

為什么yolov5中，很多細(xì)節(jié)沒(méi)有表現(xiàn)出來(lái)，我們應(yīng)該怎么閱讀yolov5的代碼？

以下內(nèi)容在直播中出現(xiàn)，但是不會(huì)在專欄投稿里面出現(xiàn)：

在目標(biāo)數(shù)據(jù)集中，數(shù)據(jù)格式為：標(biāo)簽 中心x 中心y 寬w?高h(yuǎn)

The image of targets:? tensor([0.00000e+00, 2.90000e+01, 5.88579e-01, 4.03505e-01, 2.85181e-02, 9.58207e-03])

target shape:? torch.Size([84, 6])

number of anchor:? 3 ; number of targets:? 84

gain shape:? torch.Size([7])

original ai shape:? torch.Size([3])

view ai shape:? torch.Size([3, 1])

repeat ai shape:? torch.Size([3, 84])

The none ai shape:? torch.Size([3, 84, 1])

targets shape:? torch.Size([3, 84, 7])

The length of predictions is? 3

The anchor:?

tensor([[1.25000, 1.62500],

? ? ? ? [2.00000, 3.75000],

? ? ? ? [4.12500, 2.87500]])

prediction shape:? torch.Size([8, 3, 80, 80, 85])

The anchors' shape:? torch.Size([3, 2])

the gain is? tensor([ 1.,? 1., 80., 80., 80., 80.,? 1.])

The shape of t:? torch.Size([3, 84, 7])

The r's shape:? torch.Size([3, 84, 2])

The j shape is? torch.Size([3, 84])

The t shape is? torch.Size([93, 7])

The gxy shape is? torch.Size([93, 2])

gain[[2,3]] is? tensor([80., 80.])

The gxi shape is? torch.Size([93, 2])

j, k are? torch.Size([93])? ?torch.Size([93])

l, m are? torch.Size([93])? ?torch.Size([93])

The stack j shape is? torch.Size([5, 93])

The repeat t is? torch.Size([5, 93, 7])

The filter t is? torch.Size([276, 7])

The offsets shape is? torch.Size([276, 2])

The shape of bc:? torch.Size([276, 2])

The a shape is? torch.Size([276])

The b shape is? torch.Size([276])

The c shape is? torch.Size([276])

The shape of gij is? torch.Size([276, 2])

The shape of gi is? torch.Size([276])

The shape of gj is? torch.Size([276])

The anchor:?

tensor([[1.87500, 3.81250],

? ? ? ? [3.87500, 2.81250],

? ? ? ? [3.68750, 7.43750]])

prediction shape:? torch.Size([8, 3, 40, 40, 85])

The anchors' shape:? torch.Size([3, 2])

the gain is? tensor([ 1.,? 1., 40., 40., 40., 40.,? 1.])

The shape of t:? torch.Size([3, 84, 7])

The r's shape:? torch.Size([3, 84, 2])

The j shape is? torch.Size([3, 84])

The t shape is? torch.Size([115, 7])

The gxy shape is? torch.Size([115, 2])

gain[[2,3]] is? tensor([40., 40.])

The gxi shape is? torch.Size([115, 2])

j, k are? torch.Size([115])? ?torch.Size([115])

l, m are? torch.Size([115])? ?torch.Size([115])

The stack j shape is? torch.Size([5, 115])

The repeat t is? torch.Size([5, 115, 7])

The filter t is? torch.Size([345, 7])

The offsets shape is? torch.Size([345, 2])

The shape of bc:? torch.Size([345, 2])

The a shape is? torch.Size([345])

The b shape is? torch.Size([345])

The c shape is? torch.Size([345])

The shape of gij is? torch.Size([345, 2])

The shape of gi is? torch.Size([345])

The shape of gj is? torch.Size([345])

The anchor:?

tensor([[ 3.62500,? 2.81250],

? ? ? ? [ 4.87500,? 6.18750],

? ? ? ? [11.65625, 10.18750]])

prediction shape:? torch.Size([8, 3, 20, 20, 85])

The anchors' shape:? torch.Size([3, 2])

the gain is? tensor([ 1.,? 1., 20., 20., 20., 20.,? 1.])

The shape of t:? torch.Size([3, 84, 7])

The r's shape:? torch.Size([3, 84, 2])

The j shape is? torch.Size([3, 84])

The t shape is? torch.Size([104, 7])

The gxy shape is? torch.Size([104, 2])

gain[[2,3]] is? tensor([20., 20.])

The gxi shape is? torch.Size([104, 2])

j, k are? torch.Size([104])? ?torch.Size([104])

l, m are? torch.Size([104])? ?torch.Size([104])

The stack j shape is? torch.Size([5, 104])

The repeat t is? torch.Size([5, 104, 7])

The filter t is? torch.Size([310, 7])

The offsets shape is? torch.Size([310, 2])

The shape of bc:? torch.Size([310, 2])

The a shape is? torch.Size([310])

The b shape is? torch.Size([310])

The c shape is? torch.Size([310])

The shape of gij is? torch.Size([310, 2])

The shape of gi is? torch.Size([310])

The shape of gj is? torch.Size([310])