search

Verified-Intelligence / auto_LiRPA

auto_LiRPA: An Automatic Linear Relaxation based Perturbation Analysis Library for Neural Networks and General Computational Graphs
https://arxiv.org/pdf/2002.12920
Other
289 stars 75 forks source link

An problem when using relu and linear layer #70

Open fxh903 opened 6 months ago

fxh903 commented 6 months ago

when i use a simple net for testing

class simpleNet(nn.Module):
  def __init__(self) -> None:
    super(simpleNet,self).__init__()
    self.linear_input = nn.Linear(64,100)
    self.relu = nn.ReLU()

  def forward(self,x):
    x = self.linear_input(x)
    x = self.relu(x)
    return x

if __name__ == '__main__':

    model = simpleNet()
    my_input = torch.ones((1,1,200,64))

    crown_model = BoundedModule(model, torch.empty_like(my_input))
    crown_model.eval()
    ptb = PerturbationLpNorm(norm = 2,eps=1)
    bounded_x = BoundedTensor(my_input, ptb)
    pred = crown_model(bounded_x)

    lb, ub = crown_model.compute_bounds(x=(bounded_x,), method='backward')
    print(lb,ub)

then i found 

Traceback (most recent call last):
  File "/home/fengxianheng/ConvexCertify/test.py", line 117, in <module>
    lb, ub = crown_model.compute_bounds(x=(bounded_x,), method='backward')
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/bound_general.py", line 1318, in compute_bounds
    return self._compute_bounds_main(C=C,
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/bound_general.py", line 1416, in _compute_bounds_main
    self.check_prior_bounds(final)
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/bound_general.py", line 899, in check_prior_bounds
    self.compute_intermediate_bounds(
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/bound_general.py", line 983, in compute_intermediate_bounds
    node.lower, node.upper = self.backward_general(
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/backward_bound.py", line 338, in backward_general
    A, lower_b, upper_b = l.bound_backward(
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/operators/linear.py", line 891, in bound_backward
    results = super().bound_backward(last_lA, last_uA, *x, **kwargs)
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/operators/linear.py", line 355, in bound_backward
    lA_x, uA_x, lbias, ubias = multiply_with_weight(weight, set_l=True, set_u=True)
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/operators/linear.py", line 346, in multiply_with_weight
    lA_x, lbias = _bound_oneside(last_lA, weight_override=weight)
  File "/opt/anaconda3/envs/certified_watermark/lib/python3.9/site-packages/auto_LiRPA/operators/linear.py", line 252, in _bound_oneside
    -1, *last_A.input_shape[1:])
TypeError: 'NoneType' object is not subscriptable

i dont know what's the reason. Can someone explain this?

aknirala commented 2 months ago

If I change my_input as: my_input = torch.ones((1,200,64)) it works. Any reason why my_input was 4 dimensional? Ideally, for your model, it should be 2 dimensional. i.e, like: my_input = torch.ones((1,64))