2.4.2. Implementation Details

The number of unrolled iterations in the proposed GRDRN is set to 4. The weighting factor α and λ was, respectively, optimized to be 0.05 and 1 by a limited number of searches. The network performance reaches a plateau within 60 epochs. The training samples are shuffled at the beginning of each epoch and the undersampling masks are generated on-the-fly during training to reduce overfitting. We train the network with Adam optimizer with the initial learning rate of 1e − 4, which is reduced by half every 20 epochs. The network is trained on a single NVIDIA GeForce RTX 3090 graphics card. With batch size of 1, the network training took around 12 h and 19 GB GPU memory.