Optimisation and Free up Memory

[tortijazz]

I’m working with a 1-D system and I want to approximate its ground state with an RBM and capture the correct phase diagram. The Hamiltonian has 5 free parameters, but I want to scan a grid of two of these parameters inside a for loop, say alpha and beta. I think that if I reuse variational states the optimisation will be better, that is, the first state will take completely random parameters (of the RBM) until the first optimisation is finished, however, I want that from the second optimisation the previous optimised state is taken and that is how the new optimisation begins because I think by changing alpha and beta in small steps the energy minima are not very far from each other.

First question is: How can I take previous optimised state to make next optimisation?

I have noticed that my computer's RAM memory saturates when I run a "heavy" Netket script, for example, when I do a grid of parameters in a for loop, and when that happens the run is interrupted.

Second question is: Is there any way to free up RAM memory on the fly?

[PhilipVinc]

First question is: How can I take previous optimised state to make next optimisation?

there are a few ways. You can either use a previously constructed variational state and pass it to a new VMC optimiser, something like

vs = nk.vqs.MCState(..)

ham_1 = nk.operator.LocalOperator(...)
optim_1 = nk.VMC(ham_1, ..., variational_state=vs)
optim_1.run(...)

ham_2 = nk.operator.LocalOperator(...)
optim_2 = nk.VMC(ham_2, ..., variational_state=vs)
optim_2.run(...)

you can also create a new variational state and pass it the parameters you obtained before

vs1 = nk.vqs.MCState(...)
# optimize vs1...
vs2 = nk.vqs.MCState(..., parameters=vs1.parameters)
# or also
vs2 - nk.vqs.MCState(...)
vs2.parameters = vs1.parameters

you can also save and reload the variational states..
Well, all those things will work.

Second question is: Is there any way to free up RAM memory on the fly?

It's a difficult question. It really depends why your RAM is occupied.
In general, if you destroy your objects python should be aggressively deallocating them.
For example

vs = nk.vqs.MCState(...)
# this takes up memory
vs = nk.vqs.MCState(...)
# Python should release the memory of the old object.

if instead you call your objects with different names, eg

vs1 = nk.vqs.MCState(...)
vs2 = nk.vqs.MCState(...)

python will keep both around. the way to deallocate one would be to do

del vs1
# and if you really really really want to oblige python do free memory
import gc
gc.collect()

In general, however, if you don't keep tons of objects around I don' see what fills your memory up. To be sure, however, you can add a import gc; gc.collect() call at the end of every loop iteration.

Netket computations that use a lot of samples might saturate your memory. In that case you can try to set vs.chunk_size = 1024 which should limit the amount of memory you consume at the expense of a small increase in compute time (which, for sufficiently large chunk size should be negligible).
You can try that as well.