I am experimenting with the VLAE model in a graph setting (VGAE). The main change to the standard VLAE is a graph convolution-based encoder and a reconstruction loss term.
I noticed that my models tend to crash after a few epochs (<10) with the following error:
RuntimeError: torch.linalg.cholesky: (Batch element 22): The factorization could not be completed because the input is not positive-definite (the leading minor of order 50 is not positive-definite).
This also happens when the loss becomes largely negative. If it helps, I noticed that the p_x_z loss term tends to grow much quicker than the others (graph based on the mean):
Any advice is appreciated (can also provide more details if necessary).
Hi,
I am experimenting with the VLAE model in a graph setting (VGAE). The main change to the standard VLAE is a graph convolution-based encoder and a reconstruction loss term.
I noticed that my models tend to crash after a few epochs (<10) with the following error:
RuntimeError: torch.linalg.cholesky: (Batch element 22): The factorization could not be completed because the input is not positive-definite (the leading minor of order 50 is not positive-definite).
This also happens when the loss becomes largely negative. If it helps, I noticed that the p_x_z loss term tends to grow much quicker than the others (graph based on the mean):
Any advice is appreciated (can also provide more details if necessary).