NanodetPlus网络调试笔记_nanodet输入大小-优快云博客

本文链接：https://blog.youkuaiyun.com/zhqh100/article/details/123022332

根据

GitHub - RangiLyu/nanodet: NanoDet-Plus⚡Super fast and lightweight anchor-free object detection model. 🔥Only 980 KB(int8) / 1.8MB (fp16) and run 97FPS on cellphone🔥

调试得出

backbone 为 'ShuffleNetV2'

根据 nanodet_custom_xml_dataset.yml 的默认配置,默认输入尺寸为 320,320

然后我用笔记本电脑的默认图片尺寸为 480, 640

然后等比例缩放后,尺寸为 240, 320

然后因为240除以32不能整除,会变换为 256,320

这是我此次调试的输入尺寸

首先会进行backbone计算,

summary(self.backbone, input_size = (3, 256, 320))

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1         [-1, 24, 128, 160]             648
       BatchNorm2d-2         [-1, 24, 128, 160]              48
         LeakyReLU-3         [-1, 24, 128, 160]               0
         MaxPool2d-4           [-1, 24, 64, 80]               0
            Conv2d-5           [-1, 24, 32, 40]             216
       BatchNorm2d-6           [-1, 24, 32, 40]              48
            Conv2d-7           [-1, 58, 32, 40]           1,392
       BatchNorm2d-8           [-1, 58, 32, 40]             116
         LeakyReLU-9           [-1, 58, 32, 40]               0
           Conv2d-10           [-1, 58, 64, 80]           1,392
      BatchNorm2d-11           [-1, 58, 64, 80]             116
        LeakyReLU-12           [-1, 58, 64, 80]               0
           Conv2d-13           [-1, 58, 32, 40]             522
      BatchNorm2d-14           [-1, 58, 32, 40]             116
           Conv2d-15           [-1, 58, 32, 40]           3,364
      BatchNorm2d-16           [-1, 58, 32, 40]             116
        LeakyReLU-17           [-1, 58, 32, 40]               0
   ShuffleV2Block-18          [-1, 116, 32, 40]               0
           Conv2d-19           [-1, 58, 32, 40]           3,364
      BatchNorm2d-20           [-1, 58, 32, 40]             116
        LeakyReLU-21           [-1, 58, 32, 40]               0
           Conv2d-22           [-1, 58, 32, 40]             522
      BatchNorm2d-23           [-1, 58, 32, 40]             116
           Conv2d-24           [-1, 58, 32, 40]           3,364
      BatchNorm2d-25           [-1, 58, 32, 40]             116
        LeakyReLU-26           [-1, 58, 32, 40]               0
   ShuffleV2Block-27          [-1, 116, 32, 40]               0
           Conv2d-28           [-1, 58, 32, 40]           3,364
      BatchNorm2d-29           [-1, 58, 32, 40]             116
        LeakyReLU-30           [-1, 58, 32, 40]               0
           Conv2d-31           [-1, 58, 32, 40]             522
      BatchNorm2d-32           [-1, 58, 32, 40]             116
           Conv2d-33           [-1, 58, 32, 40]           3,364
      BatchNorm2d-34           [-1, 58, 32, 40]             116
        LeakyReLU-35           [-1, 58, 32, 40]               0
   ShuffleV2Block-36          [-1, 116, 32, 40]               0
           Conv2d-37           [-1, 58, 32, 40]           3,364
      BatchNorm2d-38           [-1, 58, 32, 40]             116
        LeakyReLU-39           [-1, 58, 32, 40]               0
           Conv2d-40           [-1, 58, 32, 40]             522
      BatchNorm2d-41           [-1, 58, 32, 40]             116
           Conv2d-42           [-1, 58, 32, 40]           3,364
      BatchNorm2d-43           [-1, 58, 32, 40]             116
        LeakyReLU-44           [-1, 58, 32, 40]               0
   ShuffleV2Block-45          [-1, 116, 32, 40]               0
           Conv2d-46          [-1, 116, 16, 20]           1,044
      BatchNorm2d-47          [-1, 116, 16, 20]             232
           Conv2d-48          [-1, 116, 16, 20]          13,456
      BatchNorm2d-49          [-1, 116, 16, 20]             232
        LeakyReLU-50          [-1, 116, 16, 20]               0
           Conv2d-51          [-1, 116, 32, 40]          13,456
      BatchNorm2d-52          [-1, 116, 32, 40]             232
        LeakyReLU-53          [-1, 116, 32, 40]               0
           Conv2d-54          [-1, 116, 16, 20]           1,044
      BatchNorm2d-55          [-1, 116, 16, 20]             232
           Conv2d-56          [-1, 116, 16, 20]          13,456
      BatchNorm2d-57          [-1, 116, 16, 20]             232
        LeakyReLU-58          [-1, 116, 16, 20]               0
   ShuffleV2Block-59          [-1, 232, 16, 20]               0
           Conv2d-60          [-1, 116, 16, 20]          13,456
      BatchNorm2d-61          [-1, 116, 16, 20]             232
        LeakyReLU-62          [-1, 116, 16, 20]               0
           Conv2d-63          [-1, 116, 16, 20]           1,044
      BatchNorm2d-64          [-1, 116, 16, 20]             232
           Conv2d-65          [-1, 116, 16, 20]          13,456
      BatchNorm2d-66          [-1, 116, 16, 20]             232
        LeakyReLU-67          [-1, 116, 16, 20]               0
   ShuffleV2Block-68          [-1, 232, 16, 20]               0
           Conv2d-69          [-1, 116, 16, 20]          13,456
      BatchNorm2d-70          [-1, 116, 16, 20]             232
        LeakyReLU-71          [-1, 116, 16, 20]               0
           Conv2d-72          [-1, 116, 16, 20]           1,044
      BatchNorm2d-73          [-1, 116, 16, 20]             232
           Conv2d-74          [-1, 116, 16, 20]          13,456
      BatchNorm2d-75          [-1, 116, 16, 20]             232
        LeakyReLU-76          [-1, 116, 16, 20]               0
   ShuffleV2Block-77          [-1, 232, 16, 20]               0
           Conv2d-78          [-1, 116, 16, 20]          13,456
      BatchNorm2d-79          [-1, 116, 16, 20]             232
        LeakyReLU-80          [-1, 116, 16, 20]               0
           Conv2d-81          [-1, 116, 16, 20]           1,044
      BatchNorm2d-82          [-1, 116, 16, 20]             232
           Conv2d-83          [-1, 116, 16, 20]          13,456
      BatchNorm2d-84          [-1, 116, 16, 20]             232
        LeakyReLU-85          [-1, 116, 16, 20]               0
   ShuffleV2Block-86          [-1, 232, 16, 20]               0
           Conv2d-87          [-1, 116, 16, 20]          13,456
      BatchNorm2d-88          [-1, 116, 16, 20]             232
        LeakyReLU-89          [-1, 116, 16, 20]               0
           Conv2d-90          [-1, 116, 16, 20]           1,044
      BatchNorm2d-91          [-1, 116, 16, 20]             232
           Conv2d-92          [-1, 116, 16, 20]          13,456
      BatchNorm2d-93          [-1, 116, 16, 20]             232
        LeakyReLU-94          [-1, 116, 16, 20]               0
   ShuffleV2Block-95          [-1, 232, 16, 20]               0
           Conv2d-96          [-1, 116, 16, 20]          13,456
      BatchNorm2d-97          [-1, 116, 16, 20]             232
        LeakyReLU-98          [-1, 116, 16, 20]               0
           Conv2d-99          [-1, 116, 16, 20]           1,044
     BatchNorm2d-100          [-1, 116, 16, 20]             232
          Conv2d-101          [-1, 116, 16, 20]          13,456
     BatchNorm2d-102          [-1, 116, 16, 20]             232
       LeakyReLU-103          [-1, 116, 16, 20]               0
  ShuffleV2Block-104          [-1, 232, 16, 20]               0
          Conv2d-105          [-1, 116, 16, 20]          13,456
     BatchNorm2d-106          [-1, 116, 16, 20]             232
       LeakyReLU-107          [-1, 116, 16, 20]               0
          Conv2d-108          [-1, 116, 16, 20]           1,044
     BatchNorm2d-109          [-1, 116, 16, 20]             232
          Conv2d-110          [-1, 116, 16, 20]          13,456
     BatchNorm2d-111          [-1, 116, 16, 20]             232
       LeakyReLU-112          [-1, 116, 16, 20]               0
  ShuffleV2Block-113          [-1, 232, 16, 20]               0
          Conv2d-114          [-1, 116, 16, 20]          13,456
     BatchNorm2d-115          [-1, 116, 16, 20]             232
       LeakyReLU-116          [-1, 116, 16, 20]               0
          Conv2d-117          [-1, 116, 16, 20]           1,044
     BatchNorm2d-118          [-1, 116, 16, 20]             232
          Conv2d-119          [-1, 116, 16, 20]          13,456
     BatchNorm2d-120          [-1, 116, 16, 20]             232
       LeakyReLU-121          [-1, 116, 16, 20]               0
  ShuffleV2Block-122          [-1, 232, 16, 20]               0
          Conv2d-123           [-1, 232, 8, 10]           2,088
     BatchNorm2d-124           [-1, 232, 8, 10]             464
          Conv2d-125           [-1, 232, 8, 10]          53,824
     BatchNorm2d-126           [-1, 232, 8, 10]             464
       LeakyReLU-127           [-1, 232, 8, 10]               0
          Conv2d-128          [-1, 232, 16, 20]          53,824
     BatchNorm2d-129          [-1, 232, 16, 20]             464
       LeakyReLU-130          [-1, 232, 16, 20]               0
          Conv2d-131           [-1, 232, 8, 10]           2,088
     BatchNorm2d-132           [-1, 232, 8, 10]             464
          Conv2d-133           [-1, 232, 8, 10]          53,824
     BatchNorm2d-134           [-1, 232, 8, 10]             464
       LeakyReLU-135           [-1, 232, 8, 10]               0
  ShuffleV2Block-136           [-1, 464, 8, 10]               0
          Conv2d-137           [-1, 232, 8, 10]          53,824
     BatchNorm2d-138           [-1, 232, 8, 10]             464
       LeakyReLU-139           [-1, 232, 8, 10]               0
          Conv2d-140           [-1, 232, 8, 10]           2,088
     BatchNorm2d-141           [-1, 232, 8, 10]             464
          Conv2d-142           [-1, 232, 8, 10]          53,824
     BatchNorm2d-143           [-1, 232, 8, 10]             464
       LeakyReLU-144           [-1, 232, 8, 10]               0
  ShuffleV2Block-145           [-1, 464, 8, 10]               0
          Conv2d-146           [-1, 232, 8, 10]          53,824
     BatchNorm2d-147           [-1, 232, 8, 10]             464
       LeakyReLU-148           [-1, 232, 8, 10]               0
          Conv2d-149           [-1, 232, 8, 10]           2,088
     BatchNorm2d-150           [-1, 232, 8, 10]             464
          Conv2d-151           [-1, 232, 8, 10]          53,824
     BatchNorm2d-152           [-1, 232, 8, 10]             464
       LeakyReLU-153           [-1, 232, 8, 10]               0
  ShuffleV2Block-154           [-1, 464, 8, 10]               0
          Conv2d-155           [-1, 232, 8, 10]          53,824
     BatchNorm2d-156           [-1, 232, 8, 10]             464
       LeakyReLU-157           [-1, 232, 8, 10]               0
          Conv2d-158           [-1, 232, 8, 10]           2,088
     BatchNorm2d-159           [-1, 232, 8, 10]             464
          Conv2d-160           [-1, 232, 8, 10]          53,824
     BatchNorm2d-161           [-1, 232, 8, 10]             464
       LeakyReLU-162           [-1, 232, 8, 10]               0
  ShuffleV2Block-163           [-1, 464, 8, 10]               0
================================================================
Total params: 776,420
Trainable params: 776,420
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.94
Forward/backward pass size (MB): 76.38
Params size (MB): 2.96
Estimated Total Size (MB): 80.28
----------------------------------------------------------------

也是常规的FPN输出,会得出三个尺寸的feature

torch.Size([1, 116, 32, 40])
torch.Size([1, 232, 16, 20])
torch.Size([1, 464, 8, 10])

然后送到FPN进行计算,FPN类型为 'GhostPAN', 输入的x的三个元素,分别经过 reduce_layers 的三个元素的计算

ModuleList(
  (0): ConvModule(
    (conv): Conv2d(116, 96, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (bn): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (act): LeakyReLU(negative_slope=0.1, inplace=True)
  )
  (1): ConvModule(
    (conv): Conv2d(232, 96, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (bn): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (act): LeakyReLU(negative_slope=0.1, inplace=True)
  )
  (2): ConvModule(
    (conv): Conv2d(464, 96, kernel_size=(1, 1), stride=(1, 1), bias=False)
    (bn): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (act): LeakyReLU(negative_slope=0.1, inplace=True)
  )
)

得到三个输出

torch.Size([1, 96, 32, 40])
torch.Size([1, 96, 16, 20])
torch.Size([1, 96, 8, 10])

inner_outs 先获取了最后一个,也就是 torch.Size([1, 96, 8, 10])

然后,总之经过一系列计算,最终会输出 outs,有四个元素,尺寸分别为:

torch.Size([1, 96, 32, 40])
torch.Size([1, 96, 16, 20])
torch.Size([1, 96, 8, 10])
torch.Size([1, 96, 4, 5])

经过head计算,head的类型为 'NanoDetPlusHead',最终的输出是

torch.Size([1, 1700, 52])

其中52是 20 + 4 * (7 + 1)

20是class个数,7是reg_max

1700 = 1280 + 320 + 80 + 20

然后,这个32的长度的数据,经过softmax和一个全连接层,再和步长相乘,获取到了所有box的尺寸,进而获取到所有box的大小,共(1, 1700,4),然后送到nms里去

也就差不多得出最终结果了