A customized operation of PyTorch for implementing Kernel Prediction denoising networks.
See：
- Kernel-Predicting Convolutional Networks for Denoising Monte Carlo Renderings
- Neural Temporal Adaptive Sampling and Denoising
KernelFilterClass.forward(input, kernels, dilation=1) → Tensor
Applies a 2D Kernel Filter over an input image.
For an input size (N, C, H, W), the size of kernels must be (N, K_Size x K_Size, H, W). The output size is the same as input size.
Letting denote the value at position p, channel c of the input image and
denote the value at position p, channel n of the kernels, the Kernel Filter can be defined as:
bash ./install_kernel_filter.sh