Returning KL divergence

[Wilco17]

Thank you for this fantastic work!

Could it be possible the fit_transform() method returns the KL divergence of the run?

Thx!

[stu-blair]

related question, @DavidMChan , is the Avg. Gradient Norm printed to the log the same/analogous to KLD ?

[DavidMChan]

The average gradient norm is basically the norm of the gradient of the KL-divergence with respect to the particle positions. Thus, it can be a proxy for how stable the optimization process is, but is not the same as the KL.

[stu-blair]

Thanks for the explanation! That makes sense.

So, just to confirm-- there's no way at all currently to see the KL/D value ?

[DavidMChan]

Currently, no - but I'll consider working it into the next version, and we're always welcome to PRs if anyone wants to contribute!

Here would be a good place in the code to start looking:

tsne-cuda/src/fit_tsne.cu

Line 513 in b740a7d

float grad_norm = tsnecuda::util::L2NormDeviceVector(old_forces_device);