Contrast source inversion images, scale equivariance, and frequency hopping for learned inverse medium scattering beyond the training contrast

Synthetic far-field data sets, network inputs used for evaluation, trained networks for every seed, and the records from which every table, figure and quoted number of the paper "Contrast source inversion images, scale equivariance and frequency hopping for learned inverse medium scattering beyond the training contrast" (G. Pelekanos and K. H. Leem) are taken. The setting is two-dimensional transverse magnetic scattering at wavenumber k = 6, with the contrast on a 64 x 64 grid of the square [-1, 1]^2 and 32 incident and 32 observation directions. Arrays are NumPy files, networks are PyTorch state dictionaries, and records are JSON files. The README in the archive maps every table and figure of the paper to its record. The code is available from the authors on request. The archive is split into the parts csi_v2_part01.zip to csi_v2_part08.zip, which unzip into one folder, csi_learned_scattering_data. README.md is also given separately. Version 2 adds the networks trained on the full contrast range (Section 6.6 and Table 8) with their training and validation data, logs and evaluation records, the comparison with published network inputs (Section 6.11, Table 14 and Figures 8 and 9), the test of the CSI initialization (Section 7), the timing records of Sections 5 and 6.3, and the current figures. All files of version 1 are unchanged except one timing record, whose earlier run is kept, and the figures.