Hyper-parameter tuning of genetic algorithm

[goldoak]

Hi, I just tried using the latest genetic algorithm on a customized cabinet CAD model, but the reconstruction results are still not satisfactory, even after enabling the prior distribution and posterior tree. Could you please provide some tips on how to tune the hyper-parameters in the genetic_settings.yaml file? I'm also curious if tuning is required for each object or category.

[vevenom]

Given the shape of the object, the optimization with vanilla genetic algorithm should find the right parameters. To answer your questions, in our experiments this exact set of parameters produced good results for all categories we tested.

If you try on cabinets from generated using generate_synthetic_dataset.py you should be getting good results. I double-checked, see the results below.

It appears this is something specific to your input CAD model. From this view it appears that the CAD model and reconstructed shape do not have the same center point. Can you confirm? If so, then the output shape from the program should be recentered.

[goldoak]

Thanks for your quick reply! I have re-centered my CAD models while generating the dataset, and now it works as expected. But I encountered one failure case where the width of the reconstructed cabinet was shorter than that of the input model. I am wondering if this issue is related to the width range predefined in the shape program. However, another cabinet with a longer side length works correctly. Do you have any insights on this?

[vevenom]

Cool!

It is an interesting failure. You can compare parameter range of cabinet program parameters. But I don't think this is the issue here.

pytorchgeonodes/configs_shape_programs/sp_cabinet.json

Lines 5 to 18 in 56f6682

	"Base": {
	"type": "nested",
	"params" : {
	"Width" : {
	"type" : "float",
	"range" : [0.3, 2.0]
	},
	"Height" : {
	"type" : "float",
	"range" : [0.3, 2.5]
	},
	"Depth" : {
	"type" : "float",
	"range" : [0.1, 0.6]

I would try to rerun for that failure example just to see if this is due to stochasticity of genetic algorithm or something else. On synthetic experiments, I rarely see errors of this magnitude with synthetic experiments but, theoretically, it can happen.

If you are not bothered by the run time you can always increase population, offsprings number, and init population size:

pytorchgeonodes/configs/genetic_settings.yaml

Lines 1 to 4 in 56f6682

	population_size: 32
	#num_parents: 64
	num_offsprings: 64
	init_population_size: 100