Intermediate supervision

[bearpaw]

Intermediate supervision

[bearpaw]

It's end-to-end training. The gradient of the lower stack comes from two sources: the gradients from its own intermediate loss, and the gradients back-propagated from the higher stacks.

[bearpaw]

Hi @visonpon , what do you mean by adding the intermediate loss the single hourglass model? The model in this code is exactly the same as the original ECCV16 paper, where intermediate losses are added to the end of each stack of hourglass. For example, 8-stack hourglass model has 8 losses.

[visonpon]

@bearpaw I know that every hourglass have a loss, i mean is every hourglass updated independently according to the their loss or all this losses added together and then do BP to update the whole stacked netwroks?thanks~

[bearpaw]

Each hourglass should be updated according to its own loss, as in https://github.com/bearpaw/pytorch-pose/blob/master/example/mpii.py#L146-L148

[visonpon]

@bearpaw So every hourglass trained independently and then stacked to output the final result? so this is not an end-to-end training?

[bearpaw]

Closed for no more discussion.