Question about prop_C3 and prop_bottleneck

[MaxwellHogan]

I have a question about your implementation of the back propagation processes you have implemented for the C3 block and the bottleneck block.

From my understanding all the relevance should propagated backwards through the network - that's why in innvestigator.py you have rev_model = self.inverter.module_list[::-1]

However that logic doesn't seem to be strictly followed in the prop_c3 as you call the inverter on each block of the bottleneck in order, moreover, Conv3 seems to be ignored - as indicated below:

def prop_C3(*args):

    inverter, mod, relevance = args
    msg = relevance.scatter(which=-1)

    ## invert mod.cv3 

    c_ = msg.size(1)

    msg_cv1 = msg[:, : (c_ // 2), ...]
    msg_cv2 = msg[:, (c_ // 2) :, ...]

    for m1 in mod.m: <--- RIGHT HERE 
        msg_cv1 = inverter(m1, msg_cv1)
    
    msg = inverter(mod.cv1, msg_cv1) + inverter(mod.cv2, msg_cv2)

    relevance.gather([(-1, msg)])

    return relevance

The same is the case in prop_Bottleneck as you invert cv1, then cv2, also how do you deal with the addition as the forward function is:

y = x + conv2( conv1 ( x ))

def prop_Bottleneck(*args):

    inverter, mod, relevance_in = args

    ar = mod.cv2.conv.out_tensor.abs()
    ax = mod.cv1.conv.in_tensor.abs()

    relevance = relevance_in
    relevance = inverter(mod.cv1, relevance)
    relevance = inverter(mod.cv2, relevance)

    return relevance

My question is why is this the case, is it taken care of in another part of the code?