Test impact of neuron numbers in Architecture 0

GeorgiadouAntigoni / weak_lensing_machine_learning

1 stars 0 forks source link

Test impact of neuron numbers in Architecture 0 #2

Open bnord opened 5 years ago

bnord commented 5 years ago

loop over many "alp" constant values to change the number of neurons in all layers.
do (at least) 10 iterations of training and testing for each value of alp
change random seed each time to make sure it resets all weight values
plot sigma vs. alp value, with some kind of error bar on sigma related to the distribution of results from the iterations

GeorgiadouAntigoni commented 5 years ago

unknown Plot of sigma vs constant multiplying number of neurons with 3 iterations without changing the seed

bnord commented 5 years ago

But, we didn't plot all three iterations at each constant value, right?

GeorgiadouAntigoni commented 5 years ago

Looped the AE with the Architecture 0 over 5 alp's (constants multiplying the # of neurons) and for each alp, did 10 iterations for 10 different seeds. This is a plot of sigma vs alp value, with error bars on sigma related to the distribution of results from the iterations alp =5, # of iterations =10, epochs = 500 sigma_vs_alp_w_errors_seed10_alp5_ep500-2 copy

GeorgiadouAntigoni commented 5 years ago

Will loop over 100 iterations, instead of 10 and get back with the results.

GeorgiadouAntigoni commented 5 years ago

Plot showing sigma versus the constant multiplying the # of neurons of each layer. I did 50 iterations (for the 50 different seeds), for alp =5 and epochs = 500: sigma_vs_alp_w_errors_seed50_alp5_ep500 copy

bnord commented 5 years ago

So, to me it looks like this flattens out with increasing neurons. I could imagine that it would get better, but not by much.

What do you think?

GeorgiadouAntigoni commented 5 years ago

I think we can benefit if we increase the number of neurons but I don't see much difference between x2 and x5 implementation. Probably if we double the size we can gain in precision and not loose substantially in speed.

bnord commented 5 years ago

I agree with that.

Do you mean double from x5 or double from x2?

bnord commented 5 years ago

Also, are you recording the run times for each of the trainings?

GeorgiadouAntigoni commented 5 years ago

I mean to double the # of neurons in Arch0, which means to go with the x2 implementation.

GeorgiadouAntigoni commented 5 years ago

I don't record the run times... I'll start doing that from now on.