export result while equations are evolving

[AKHCE]

Hi,
By passing equation_file and/or temp_equation_file to PySRRegressor, we can save the latest updated "hall of fame". For example,

The results can be saved as csv.

The output csv file gets updated while evolution. In other words, we always have one csv file which is being updated. I am wondering if there is a way to keep all intermediate csv files (or at least some of them). For example, can we store 1 separated csv file per niterations with its own unique name? What I want to finally achieve is to plot loss versus evolution for each Complexity. e.g. complexity = 5, then plotting the corresponding loss versus iterations/generations.

Any advice is much appreciated,

Thank you so much for your support,
Cheers

[MilesCranmer]

Hi @AKHCE,

Sorry for the late reply, I missed this post for some reason.

There is not a built-in mechanism for saving the csv file over time. However, there is a mechanism for saving the entire history of a search to a JSON file (including every single mutation that happens). It is a bit overkill and I'm not sure if it is useful, but you could try it out.

It's typically used from the Julia side so there's no built-in parameter in PySRRegressor. However, you could turn this on by adding to:

PySR/pysr/sr.py

Line 1608 in 6abcc15

options = SymbolicRegression.Options(

add a line

val_recorder=Main.eval("Val(true)")

and it should generate a JSON file pysr_recorder.json of the history of the search. Just keep in mind that it is going to be much slower. This JSON file is used only as a debugging tool (by myself) so there's not documentation on it.

Other than that, you could modify the backend, particularly this line, and get it to save a file given by the current time() rounded to some precision.

You can use a custom backend in PySR by following the instructions here: https://astroautomata.com/PySR/backend/

Cheers,
Miles

[AKHCE]

Thank you so much for giving me alternatives. I chose a different approach which seems working well. Simply, I pass warm_start = True to enable resuming the training, then I use a for-loop to move forward the model training. Please see screenshot below:

As you can see I set populations (I interpret it as generation) to 1 and the model is trained through a for-loop. Also, niterations which is number of re-running algorithm in each generation can be set any values, I chose 1 in the given example.

As you can see, I can save hall of fame for each generation. How do you think about this approach?
Thank you so much for your support,

[MilesCranmer]

Oh yeah, that’s way easier!!! Sorry for my overcomplicated approach, I’ve done so much dev on the internals that I forgot there was a simple approach like this!

[MilesCranmer]

Btw, when using pow, you should turn on constraints={"pow": (-1, 1)}, otherwise it will just start stacking power laws everywhere.

[AKHCE]

Absolutely. Thanks