def propagate_relevance(self,
module,
relevance_input,
relevance_output,
lrp_method,
lrp_params=None):
# todo how is batchnorm handled by default in innvestigate?
if lrp_method == 'identity':
R = relevance_output[0]
else:
input_ = module.input[0]
mean = module._buffers['running_mean']
var = module._buffers['running_var']
gamma = module._parameters['weight']
beta = module._parameters['bias']
w = (gamma / torch.sqrt(var + module.eps))[:, None, None]
b = (beta - (mean * gamma) / torch.sqrt(var + module.eps))[:, None,
None]
xw = input_ * w
R = util.safe_divide(torch.abs(xw), (torch.abs(xw) + torch.abs(b))) * \
relevance_output[0]
assert not torch.isnan(R.sum())
assert not torch.isinf(R.sum())
assert R.sum() != 0
module.zero_grad()
return R, relevance_input[1], relevance_input[2]
def propagate_relevance(self, module, relevance_input, relevance_output,
lrp_method, lrp_params=None):
input_ = module.input[0]
module_clone = self._clone_module(module)
with torch.enable_grad():
X = input_.clone().detach().requires_grad_(True)
Z = module_clone(X)
S = util.safe_divide(relevance_output[0].clone().detach(), Z)
Z.backward(S)
R = X * X.grad
assert not torch.isnan(R.sum())
assert not torch.isinf(R.sum())
module.zero_grad()
return (R, )