commit unknown

LRP/lrp.py

@@ -7,10 +7,12 @@ class LRP():
  :param rule:str: name of used rule
  '''
 
- def __init__(self, model, rule):
+ def __init__(self, model, rule, input_lowest=-1, input_highest=1):
  self.model = copy.deepcopy(model)
  self.model = self.model.eval()
- self.model = utils.redefine_nn(self.model, rule=rule) #redefine each layer(module) of model, to set custom autograd func
+ self.model = utils.redefine_nn(self.model, rule=rule, 
+ input_lowest=input_lowest, 
+ input_highest=input_highest) #redefine each layer(module) of model, to set custom autograd func
  self.output = None
 
  def forward(self, input_):
@@ -35,7 +37,7 @@ def relprop(self, input_, R=None):
  C = self.local_input.grad.clone().detach()
  assert C is not None, 'obtained relevance is None'
  self.local_input.grad = None
- R = C*input_.clone().detach()
+ R = C#*input_.clone().detach()
  return R
 
  __call__ = relprop

LRP/rules.py

@@ -58,8 +58,9 @@ def z_rule(func, input, R, func_args, keep_bias=False):
  Z.backward(S)
  assert input.grad is not None
  C = input.grad
- R = input * C
- return R
+ Ri = input * C
+ print(Ri.sum())
+ return Ri
 
  @staticmethod
  def z_plus(func, input, R, func_args, keep_bias=False):
@@ -70,15 +71,16 @@ def z_plus(func, input, R, func_args, keep_bias=False):
  if func_args.get('bias', None) is not None:
  if not keep_bias:
  func_args['bias'] = None
- func_args['weight'].clamp_(0, float('inf'))
+ if func_args.get('weight', None) is not None:
+ func_args['weight'].clamp_(0, float('inf'))
  with torch.enable_grad():
  Z = func(input, **func_args)
  S = R /(Z + (Z==0).float()*np.finfo(np.float32).eps)
  Z.backward(S)
  assert input.grad is not None
  C = input.grad
- R = input * C
- return R
+ Ri = input * C
+ return Ri
 
  @staticmethod
  def z_box(func, input, R, func_args, lowest, highest, keep_bias=False):
@@ -111,14 +113,15 @@ def z_box(func, input, R, func_args, lowest, highest, keep_bias=False):
  if input.grad is not None: input.grad.zero_()
  input.retain_grad()
  with torch.enable_grad():
- Z = func(input, **ifunc_args) - func(L, **pfunc_args) - func(H, **nfunc_args) + np.finfo(np.float32).eps
+ Z = func(input, **ifunc_args) - func(L, **pfunc_args) - func(H, **nfunc_args)
  S = R / (Z + (Z==0).float()*np.finfo(np.float32).eps)
  Z.backward(S)
  assert input.grad is not None
  assert L.grad is not None
  assert H.grad is not None
- R = input * input.grad - L * L.grad - H * H.grad
- return R
+ import ipdb; ipdb.set_trace() # BREAKPOINT
+ Ri = input * input.grad - L * L.grad - H * H.grad
+ return Ri
 #
 #
 #def z_epsilon_rule(module, input_, R, keep_bias=True):

LRP/utils.py

@@ -29,7 +29,7 @@ def copy_module(module):
  module._backward_hooks.popitem() # remove hooks from module copy
  return module
 
-def redefine_nn(model, rule='z_rule', input_lowest=-1, input_highest=1):
+def redefine_nn(model, rule, input_lowest, input_highest):
  '''
  go over model layers and overload chosen instance methods (e.g. forward()).
  New methods come from classes of layers module