Fig. 2: Network workflow and schematic diagram of each part structure.
From: Deep learning-driven likelihood-free parameter inference for 21-cm forest observations
a Depicts network workflow chart. b Depicts schematic diagram of the architecture of the NF. Both the GNF and INF are composed of RQ-NSF (AR). In the GNF, the conditions are parameters, and the output is a 1D power spectrum. Conversely, in the INF, the condition is a 1D power spectrum, and the outputs are parameters. The network consists of a sequence of flows, each of which consists of a sequence of autoregressive layers that share parameters. Input the condition and the sampling vector obtained from the base distribution into the NFs to generate the output. The simulated power spectra data set is divided into a training set, a validation set, and a testing set. The training set and validation set are used to train the GNF, which generates reasonable power spectra for any given parameters. The INF is trained using the 1D power spectra generated by the GNF to learn the parameter distribution corresponding to any given power spectrum. Different testing sets of simulated power spectra are used to verify the rationality of the GNF and the INF, respectively.
