Hourglass implementation vs paper

[jerome-auguste]

Hey there!

Looking at the Hourglass implementation, it looks like the out_convs (for intermediate supervision) are implemented using the ConvModule with no act_cfg argument. This means that they use the default argument which is ReLU. What purpose does it serve to have a ReLU activation at prediction stage? This technical choice has not been discussed in the Stacked Hourglass paper?

Moreover, it looks like the HourglassModule implementation uses standard ResLayer (as used in regular ResNet, with the default argument downsampling_first=True) instead of the Residual Module as in Figure 4 of Stacked Hourglass paper, even though the latter has been implemented in the ResLayer. Is there a specific reason for that technical choice? Then what did motivate you to choose the BasicBlock from ResNet as their

Thank you in advance for your answer and thank you for your hard work!