您好!
感谢您的源码。我在测试cdc模型的时候,有的类别得分输出大于1,导致评测AP和AR的时候会报错:Confidence score of obj1 should not be smaller than obj2. ""obj1['score'] = %s, obj2['score'] = %s"
后来我查看cdc模型和hgwi模型的decode结构时,发现二者有差异:
cdc模型decode最后没有加prelu激活:
x = self.prelu(self.convt1(x)) # (B, 256, W/4, 16, 16) -> (B, 128, W/2, 32, 32)
x = self.prelu(self.convt2(x)) # (B, 128, W/2, 32, 32) -> (B, 64, W, 64, 64)
x = self.convt3(x) # (B, 64, W, 64, 64) -> (B, 3, W, 128, 128)
return x
而hgwi模型最后有加prelu激活:
x = self.prelu(self.convt1(x + x3)) # (B, 256, W/4, 16, 16) -> (B, 128, W/2, 32, 32)
x = self.prelu(self.conv1(x))
x = self.prelu(self.convt2(x + x2)) # (B, 128, W/2, 32, 32) -> (B, 64, W, 64, 64)
x = self.prelu(self.conv2(x))
x = self.prelu(self.convt3(x + x1)) # (B, 64, W, 64, 64) -> (B, 3, W, 128, 128)
x = self.prelu(self.conv3(x))
return x
但是原论文的模型示意图上面,所有模型decode模块的最后都加了prelu激活,不知这是否是一个code bug?
感谢!