For advanced inference, especially with certain attacks, it would be beneficial to implement custom CUDA kernels for fast inference on Nvidia GPUs. For example, in the case of the One Pixel Attack, we could implement CUDA kernels that execute the Differential Evolution Algorithms. We welcome more ideas and examples in this area. This topic should be explored further through more discussions.
For advanced inference, especially with certain attacks, it would be beneficial to implement custom CUDA kernels for fast inference on Nvidia GPUs. For example, in the case of the One Pixel Attack, we could implement CUDA kernels that execute the Differential Evolution Algorithms. We welcome more ideas and examples in this area. This topic should be explored further through more discussions.