robbwu / tensorsvm Goto Github PK

fixed in commit a6700a1

The reason is that 400k dataset triggers chunking (NN=100k), but the C=Q'*W did not accumulate (in SGEMM(), the beta was 0, should have been 1).

Hardware/Software requirement

Short description

How to train

How to predict

iter 8, mu=1.311e-06, normdx=3.811e-03, normdy=4.306e-03 max/min(D)=1.925e+16 pobj=-5.396244462e+08 dobj=-5.396244536e+08
MPC elapsed time = 362 seconds
Writemodel #BSV 2500092, #SV 4608000
Segmentation fault

When running on SUSY with rank 512, here is a segmentation fault when writing the model

if the #features if very large (>10k) the dense matrix representation is not feasible both in terms of memory space and computation.

Need sparse matrix representation.

Using the whole dataset, the Chol(C) fails very easily with $k$ as small as 100.

right now, the size of data (n) must be divisble by the number of chunks (nb).

This only applies to large datasets, but nonetheless should be fixed.

Enable TensorCore in LRA.
Test the stability.

make it compatible with LIBSVM/ThunderSVM.

How do we choose good hyperparameters such as gamma, C, and rank k?