ERROR with ReLU, inplace_unfold cause the bad shape. #73
Description
Describe the bug
I tried to use auto_LiRPA to verify the robustness of my audio processing model. Unfortunately, when I use model.compute_bounds, a RuntimeError: The following operation failed in the TorchScript interpreter. occurred.
The situation directly caused by the unfitted shape of A_pos([64, 4, 25, 20, 64, 9, 9]) and d_pos([1, 4, 24, 19, 64, 9, 9]) in A_new = d_pos * A_pos + d_neg * A_neg, which happens during clamp_mutiply_forward operation. I tried to trace back the cause of this difference, and find something happened in the inplace_unfold step of maybe_unfold_patches, in the relu.py.
Here follows my code. I make a emulation dataset to provide random data with the same attributes of real data, which are Mel Spectrograms of some audio, to avoid unnecessary uploadings. And I used an adjusted Resnet18 to extract the features, to make it work correctly on [B, 1, 100, 80] shaped data.
To Reproduce
System configuration:
- OS: Ubuntu 22.04
- Python version: Python 3.10
- Pytorch Version: Pytorch 2.2
- Hardware: RTX 4090
- Have you tried to reproduce the problem in a cleanly created conda/virtualenv environment using official installation instructions and the latest code on the main branch?: Yes