'BoundTranspose' object has no attribute 'lower'

[Cli212]

Hi,

Thanks for your great work first!

I'm trying to run auto_LiRPA on an MLP based model, but I have run into some problems. I tried to minimize the code I have used and here is the code that I can reproduce the issue:

import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from auto_LiRPA import BoundedModule, BoundedTensor, PerturbationLpNorm

# Define model architecture

class SetConv(nn.Module):
    def __init__(self, sample_feats, hid_units):
        super(SetConv, self).__init__()
        self.sample_mlp1 = nn.Linear(sample_feats, hid_units)
        self.sample_mlp2 = nn.Linear(hid_units, hid_units)

    def forward(self, samples, sample_mask):
        # samples has shape [batch_size x num_joins+1 x sample_feats]
        hid_sample = F.relu(self.sample_mlp1(samples))
        hid_sample = F.relu(self.sample_mlp2(hid_sample))
        hid_sample = hid_sample * sample_mask  # Mask
        hid_sample = torch.sum(hid_sample, dim=1, keepdim=False)
        sample_norm = sample_mask.sum(1, keepdim=False)
        hid_sample = hid_sample / sample_norm  # Calculate average only over non-masked parts
        return hid_sample
model =SetConv(3, 10)
samples = torch.rand((1,2,3), requires_grad=True)
sample_mask = torch.rand((1,2,1), requires_grad=True)
bounded_model = BoundedModule(model, (torch.zeros_like(samples), torch.zeros_like(sample_mask)))
bounded_model.eval()
ptb = PerturbationLpNorm(norm=np.inf, eps=0.1)
my_input = (BoundedTensor(samples, ptb), BoundedTensor(sample_mask, ptb))
outputs = bounded_model(my_input)
lb, ub = bounded_model.compute_bounds(x=(my_input,), method="CROWN")

The error goes as "AttributeError: 'BoundTranspose' object has no attribute 'lower'" and it happens in the last line of the code when I try to compute the bound. I have tried to debug this issue but couldn't find any way to fix that. Could you take a look at this any time you feel comfortable?

Thanks!

[shizhouxing]

Hi @Cli212 , thanks for raising this issue, and it has been fixed on the latest master branch.

Also, since your input looks like sequence data (or similar input shape), and auto_LiRPA's default "patches" mode has issues with that right now, please add the following option (bound_opts) to disable the patches mode for now:

bounded_model = BoundedModule(model, (torch.zeros_like(samples), torch.zeros_like(sample_mask)),
    bound_opts={'conv_mode': 'matrix'})

[Cli212]

Great to know that! Thanks for your excellent work!