Group Shuffle

Group Conv

Channel Shuffle

Details could be checked here.

https://github.com/mindspore-lab/mindcv/blob/main/mindcv/models/shufflenetv1.py

def channel_shuffle(self, x: Tensor) -> Tensor:
    batch_size, num_channels, height, width = x.shape

    group_channels = num_channels // self.group
    x = ops.reshape(x, (batch_size, group_channels, self.group, height, width))
    x = ops.transpose(x, (0, 2, 1, 3, 4))
    x = ops.reshape(x, (batch_size, num_channels, height, width))
    return x

Memory Layout

123 456 789 (3 groups)

reordered as

147 258 369

assuming num_channels % self.group != 0 here.

Last modified: 10 March 2024