A normalizing flow approach for the inference of star cluster properties from unresolved broadband photometry - I. Comparison to spectral energy distribution fitting

Open Access

Table A.1

Network architecture and training parameters.

parameter	explanation	fixed?	search range	final value(s)
number of coupling blocks	The total number of invertible conditional coupling blocks (see Sect. 2.1).	no	[4, 21]	12
internal network depth	The number of hidden layers in the coupling block subnetworks.	yes	−	2
internal network width	Number of neurons per layer in coupling block subnetworks.	no	[128, 2048]	256
activation function	Activation function of subnetworks.	yes	−	ReLU
initial learning rate	Step size of optimizer at first epoch.	no	[10⁻⁵, 10⁻²]	0.001876
learning rate decay	Factor by which learning rate decays after a fixed number of meta-epochs.	no	[0.1, 0.9]	0.6637
β₁, β₂	Parameters of Adam optimizer (see Kingma & Ba 2014).	no	(0.8, 0.8) or (0.9, 0.999)	(0.9, 0.999)
batch size	Number of (random) training samples in a single optimization step.	no	[64, 1024]	512
η	Regularization strength as described in Sect. 2.1.	no	[10⁻⁷, 10⁻⁴]	1.353 × 10⁻⁶

Notes. The third column indicates whether a parameter has been subject to the hyperparameter search. In that case, a search range is specified.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.