ConvTranspose CROWN Bounds #61
Comments
|
CROWN with a single input (e.g. |
|
Hi @cherrywoods , could you please share the code for the model definition? |
|
Sure, sorry for not including it right away: generator = nn.Sequential(
nn.ConvTranspose2d(4, 49, kernel_size=4, stride=1, bias=False), # 49 x 4 x 4
nn.BatchNorm2d(49, affine=True),
nn.LeakyReLU(negative_slope=0.2),
nn.ConvTranspose2d(49, 12, kernel_size=4, stride=4, bias=False), # 12 x 16 x 16
nn.BatchNorm2d(12, affine=True),
nn.LeakyReLU(negative_slope=0.2),
nn.ConvTranspose2d(12, 1, kernel_size=13, stride=1, bias=False), # 1 x 28 x 28
nn.Sigmoid(),
) |
|
Hi @cherrywoods , the issue is that you need to update |
|
Hi @shizhouxing, the incorrect batch dimension was indeed a problem in the code I posted, however a very similar error persists also with fixed batch dimensions: import torch
from auto_LiRPA import PerturbationLpNorm, BoundedModule, BoundedTensor
net = torch.load("mnist_conv_generator.pyt")
net = BoundedModule(net, torch.zeros(1, 4, 1, 1))
ptb = PerturbationLpNorm(x_L=torch.zeros(10, 4, 1, 1), x_U=torch.ones(10, 4, 1, 1))
tensor = BoundedTensor(torch.zeros(10, 4, 1, 1), ptb)
net.compute_bounds(x=(tensor,), method="ibp") # works fine, output omitted
net.compute_bounds(x=(tensor,), method="crown") |
|
Hi @cherrywoods , You'll need to modify both |
|
Hi @shizhouxing, this was only a typo. I updated the code above. The error remains the same. |
|
Hi @cherrywoods , but I tried your code and it worked fine on my side. I see your output contains |
|
That indeed seemed to be the issue. I somehow messed up pulling the latest release from Github. Thanks for your patience and sorry for the inconvenience. I'm happy that I can now use ConvTranspose layers :) |
|
I reopen this because I keep getting errors in the actual code I'm using, which obviously uses different bounds than 0.0 and 1.0. I debugged through this for the past hour and couldn't find anything like the errors that we discussed above. To be on the safe side this time, I made a docker container that reproduces the issue: conv_transpose_issue.zip The container creates a conda environment, downloads and installs the latest auto_LiRPA commit and then runs the following script: import torch
from torch import nn
import auto_LiRPA
from auto_LiRPA import PerturbationLpNorm, BoundedModule, BoundedTensor
print(auto_LiRPA.__version__)
torch.manual_seed(0)
net = nn.Sequential(
nn.ConvTranspose2d(4, 49, kernel_size=4, stride=1, bias=False), # 49 x 4 x 4
nn.BatchNorm2d(49, affine=True),
nn.LeakyReLU(negative_slope=0.2),
nn.ConvTranspose2d(49, 12, kernel_size=4, stride=4, bias=False), # 12 x 16 x 16
nn.BatchNorm2d(12, affine=True),
nn.LeakyReLU(negative_slope=0.2),
nn.ConvTranspose2d(12, 1, kernel_size=13, stride=1, bias=False), # 1 x 28 x 28
nn.Sigmoid(),
)
net = BoundedModule(net, torch.empty(1, 4, 1, 1))
lb = torch.zeros(1, 4, 1, 1)
ub = torch.ones(1, 4, 1, 1)
ptb = PerturbationLpNorm(x_L=lb,x_U=ub)
tensor = BoundedTensor(lb, ptb)
print(lb.shape, ub.shape, tensor.shape)
print(lb, ub, tensor)
bounds = net.compute_bounds(x=(tensor,), method="crown") # works fine
print(bounds)
lb = lb.clone() - 1.0
ptb = PerturbationLpNorm(x_L=lb,x_U=ub)
tensor = BoundedTensor(lb, ptb)
print(lb.shape, ub.shape, tensor.shape)
print(lb, ub, tensor)
bounds = net.compute_bounds(x=(tensor,), method="crown") # fails
print(bounds)When I run this using: docker build . -t auto_lirpa
docker run -t auto_lirpaI get this output: I know this behaviour is extremely strange, but since I am only subtracting 1.0 from the lower bound for which CROWN works, I don't think it's a shape issue again. |
|
I also confirmed that the error persists when I use a batch size of |
|
Thanks for reporting the issue and sorry for delayed response. We have fixed it internally and will push the fix in the upcoming release soon. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Describe the bug
I was delighted to see that
auto_LiRPAcan bound ConvTranspose layers out of the box, but, unfortunately, CROWN in batch mode doesn't seem to work.To Reproduce
Code to reproduce with the attached network (zipped): mnist_conv_generator.zip
System configuration:
The text was updated successfully, but these errors were encountered: