pycaffe net.forward() assumes end blobs have the same name as their layers

[cdoersch]

If you call net.forward(end='somelayer') from python, where somelayer is a layer that has a different name for its output blob, _Net_forward() will attempt to access net.blobs['somelayer'] (and in fact, returns this as the output). This happens because this line appends the end layer name to the list of blobs to be output. If the appended name doesn't exist as a blob, the result is an exception.

Note that there may also be multiple outputs for the end layer, in which case the current behavior doesn't make sense either.

[vishakhhegde]

I too had the same issue. For my application, I just needed to comment out the line "return {out: self.blobs[out].data for out in outputs}" in pycaffe.py. Forward propagation now happens without returning any blob vectors.

[shelhamer]

Note that #5519 fixed the layer/top name issue! but still does not handle multiple tops as #2975 does.

[Noiredd]

I just checked #5519, called forward on Slice layer with 2 tops and it worked just fine. @shelhamer any example where it fails? I'll be glad to improve the code.
EDIT: forward with end param pointing to a Convolution with 4 bottoms and 4 tops, as well as blobs containing a list of 3 different blobs looks flawless. backward to Slice layer mentioned previously also seemed to be good.

[shelhamer]

@Noiredd sorry, I misspoke! It looks like #5519 handles everything 👍

[Noiredd]

@shelhamer Looks like this issue is fully resolved then :)