OpenPose 如何通过输出形状和真实值不匹配来实现其损失函数？

如何解决OpenPose 如何通过输出形状和真实值不匹配来实现其损失函数？

我最近一直在实现基于 OpenPose 的模型。在 OpenPose 中，它使用 VGG 作为其主干模型来提取特征图，但 VGG 包含最大池化层，这会将输出的形状减少到 1/4。这是 OpenPose 的模型结构：

VGGOpenPose(
  (model0): OpenPose_Feature(
    (model): Sequential(
      (0): Conv2d(3,64,kernel_size=(3,3),stride=(1,1),padding=(1,1))
      (1): ReLU(inplace=True)
      (2): Conv2d(64,1))
      (3): ReLU(inplace=True)
      (4): MaxPool2d(kernel_size=2,stride=2,padding=0,dilation=1,ceil_mode=False)
      (5): Conv2d(64,128,1))
      (6): ReLU(inplace=True)
      (7): Conv2d(128,1))
      (8): ReLU(inplace=True)
      (9): MaxPool2d(kernel_size=2,ceil_mode=False)
      (10): Conv2d(128,256,1))
      (11): ReLU(inplace=True)
      (12): Conv2d(256,1))
      (13): ReLU(inplace=True)
      (14): Conv2d(256,1))
      (15): ReLU(inplace=True)
      (16): Conv2d(256,1))
      (17): ReLU(inplace=True)
      (18): MaxPool2d(kernel_size=2,ceil_mode=False)
      (19): Conv2d(256,512,1))
      (20): ReLU(inplace=True)
      (21): Conv2d(512,1))
      (22): ReLU(inplace=True)
      (23): Conv2d(512,1))
      (24): ReLU(inplace=True)
      (25): Conv2d(256,1))
      (26): ReLU(inplace=True)
    )
  )
  (model1_1): Sequential(
    (0): Conv2d(128,1))
    (1): ReLU(inplace=True)
    (2): Conv2d(128,1))
    (3): ReLU(inplace=True)
    (4): Conv2d(128,1))
    (5): ReLU(inplace=True)
    (6): Conv2d(128,kernel_size=(1,1))
    (7): ReLU(inplace=True)
    (8): Conv2d(512,38,1))
  )
  (model2_1): Sequential(
    (0): Conv2d(185,kernel_size=(7,7),padding=(3,3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128,3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128,3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128,3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128,3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128,1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128,1))
  )
  (model3_1): Sequential(
    (0): Conv2d(185,1))
  )
  (model4_1): Sequential(
    (0): Conv2d(185,1))
  )
  (model5_1): Sequential(
    (0): Conv2d(185,1))
  )
  (model6_1): Sequential(
    (0): Conv2d(185,1))
  )
  (model1_2): Sequential(
    (0): Conv2d(128,19,1))
  )
  (model2_2): Sequential(
    (0): Conv2d(185,1))
  )
  (model3_2): Sequential(
    (0): Conv2d(185,1))
  )
  (model4_2): Sequential(
    (0): Conv2d(185,1))
  )
  (model5_2): Sequential(
    (0): Conv2d(185,1))
  )
  (model6_2): Sequential(
    (0): Conv2d(185,1))
  )
)

在原始论文中，它说groundtruth heatmap和paf与输入图像的宽度和高度相同。 OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields

我在 Python 中搜索了一些 OpenPose 的实现。他们大多使用element-wise loss函数来计算output和groundtruth label之间的loss，和论文中提到的函数一样：

loss function in openpose

我想知道 OpenPose 的输出是否与输入图像的大小不同，OpenPose 是如何计算输出和 groundtruth heatmap/paf 之间的损失函数的？