Undesirable behavior of LayerActivation in networks with inplace ReLUs

[mrsalehi]

Hi,
I was trying to use captum.attr._core.layer_activation.LayerActivation to get the activation of the first convolutional layer in a simple model. Here is my code:

torch.manual_seed(23)
np.random.seed(23)
model = nn.Sequential(nn.Conv2d(3, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                      nn.ReLU(inplace=True),
                      nn.Conv2d(4, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                      nn.ReLU(inplace=True))

layer_act = LayerActivation(model, model[0])
input = torch.randn(1, 3, 5, 5)
mylayer = model[0]
print(torch.norm(mylayer(input) - layer_act.attribute(input), p=2))

In fact, I have computed the activation in two different ways and compared them afterwards. Obviously, I expected a value close to zero to be printed as the output, however, this is what I got:

tensor(3.4646, grad_fn=<NormBackward0>)

I hypothesize that the inplace ReLU layer after the convolutional layer acts on its output since there were many zeros in the activation computed by Captum ( i.e. layer_act.attribute(input)). In fact, when I changed the architecture of the network to the following:

model = nn.Sequential(nn.Conv2d(3, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                      nn.ReLU(),
                      nn.Conv2d(4, 4, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                      nn.ReLU(inplace=True))

then the outputs matched.

System information

• Python 3.7.0
• torch 1.3.0
• Captum 0.1.0

[vivekmig]

Hi @mrsalehi, yes, this is a bug, thanks for pointing it out! We will push a fix for this soon.

[vivekmig]

Fix has been merged here: 5bf06ba