This implements in pytorch the Convolutional Kernel Approach proposed by Mairal and al in their 2014 NIPS paper Convolutional Kernel Networks.
Some small modifications were made with respect to the original paper (see below). In short, the training algorithm works as follows:
- (1) extract patches from input map (by concatenating neighboring pixel input). Normalize the patches so that they all have unit norm. (Keep the orginal norms in memory, in Cell.norms).
- (2) apply dimensionality reduction to these large patches (RobustScaler+PCA), so that the whole thing remains computable
- (3) initalize new filters with Kmeans
- (4) train W and eta (parameters of the cell) on a subset of the data. The function check_convergence() allows to check how well we are doing in terms of approximating the kernel.
- (5) call get_activation_map() to compute the activation of each patch