memory error on medium scale about misvm

[garydoranjr]

Hello,

Kernel methods require computing an n-by-n kernel matrix between n training examples. With 120,000 examples, this is a matrix with 14.4 billion entries, which is ~115 GB if each entry is represented with a double-precision (8-byte) floating point number. If the kernel matrix is sparse, you might get away with using a sparse matrix representation, but that is not implemented here. There are also other optimization procedures (again, not implemented here unfortunately) that don't require computing the full kernel matrix in memory, but then you trade-off memory for CPU and it might take significantly longer to train the model.

Another approach I suggest (if you are ultimately interested in bag-level labels) is to use a bag-level classifier such as the MI-Kernel method. This only requires computing a matrix that is O(n_bags^2) instead of O(n_instances^2) and actually tends to have better performance on the bag-labeling task.