Is it normal, or is there an explanation, for an error trace to reveal that negative input strides have been used? I read here that strides encode the number of steps needed in memory to step along a given axis in an array so it seemed odd to me that my error trace was showing negative input strides. See attached for full error trace with exception_verbosity=high TempMCMC_HighExceptionVerbosity.txt (2.5 MB).
I have also noticed that my model is using more memory than the error trace says it is. The storage map shown in the attached error trace tells me that barely 10MB of memory are being used, however, when I run the
free command in my Ubuntu terminal while the model is running I can see that upwards of 4GB are being used. Could this issue be linked to the negative input strides?
A quick summary about the model: each likelihood evaluation solves a time-dependent PDE (the heat equation) using theano scan to step forward in time. The physical problem has nonlinear boundary conditions so I have used custom Op to solve these at each time step. I have recorded temperature data at numerous time steps and I am trying to recover various thermal parameters such as conductivity and heat capacity.
Thanks in advance for any assistance!