Issue about using Hourglass as backbone

[seekFire]

Due to the hourglass network has multiple output (According to "intermediate supervision", the number of hourglass block is equal to the number of output block), suppose that I use hourglass network as backbone, which output should be regarded as the final output in test stage?

I noticed that in "large_hourglass.py", the output of the class "exkp"(inherited by class "HourglassNet") is multiple; while in "base_trainer.py", the class "ModuleWithLoss" return "output[-1]", does this mean that you only use the last output of hourglass to train the network?
Shouldn't we use all output to train, and the last output to test?

[xingyizhou]

We do use all output to train and the last output to test.

[seekFire]

@xingyizhou Okay! Thank you very much!