Open nanmehta opened 3 years ago
The forward of DDF is implemented in CUDA, the autograd is not suitable for this case. So, I also implement the backward function.
sir, can you please explain the concept of the relation between dynamic filters and attention as shown in Figure 3 of your paper. How can we relate applying dynamic concept (where every filter is applied to each pixel) with attention
sir, can you please let me know, why have you use backward compatible code also in the implementation. I am talking about the class defined below.. @staticmethod def backward(ctx, grad_output): assert grad_output.is_cuda