The dataset consists of deep neural network series (DNNs) underpinning two incarnations of the R2D2 algorithm. The first incarnation, simply referred to as R2D2, takes U-Net as the architecture of its network components. The second incarnation, referred to as R3D3, takes an unrolled architecture dubbed R2D2-Net as the architecture of its network components. R2D2-Net is a novel DNN obtained by unrolling the R2D2 algorithm itself. Both R2D2 and R3D3 series are trained specifically to form radio images of size 512x512 from radio-interferometric data acquired by the Very Large Array, under specific considerations of the data-weighting scheme and angular resolution (i.e., pixel size).
The provided DNN series are as follows:
(1) R2D2 series composed of 15 U-Net components,
(2) R3D3 series composed of 7 R2D2-Net components, with each R2D2-Net consisting of 3 U-Net layers and 2 data consistency layers,
(3) R3D3 series composed of 8 R2D2-Net components, with each R2D2-Net consisting of 6 U-Net layers and 5 data consistency layers.
DNNs are available in two formats: PyTorch models generated in training and used for inference in Python, and ONNX models used for inference in MATLAB.