conflict with the original paper

[yhao-z]

I don't think this implementation consistent with the original paper (RAKI).

In Network.py, the RAKI network are build by,

        net = nn.Sequential(
            nn.Conv2d(in_channels=2 * self.channels_in, out_channels=128, kernel_size=[3, 5], padding=[1, 2], padding_mode='replicate'),
            nn.ReLU(),
            nn.Conv2d(in_channels=128, out_channels=64, kernel_size=[1, 1], padding=[0, 0], padding_mode='replicate'),
            nn.ReLU(),
            nn.Conv2d(in_channels=64, out_channels=64, kernel_size=[1, 3], padding=[0, 1], padding_mode='replicate'),
            nn.ReLU(),
            nn.Conv2d(in_channels=64, out_channels=self.R * 2 * self.channels_in, kernel_size=[3, 3], padding=[1, 1], padding_mode='replicate'),
        ).to(self.device)

There is no conv dialation in your implement, while the paper has mensioned that

In our implementations, all layers use kernel dilation of size R in the ky direction to only process the acquired k‐space lines.

Also, the paper shows that the network only has three convs, but your code contains four. The kernel size is not consistent with the paper too.

I'm not sure whether your implement improves the reconstruction performance or not? Did u intentionally change the code for improving performance?