shufflenet unit原理和实现

目录

前言

论文

shufflenet unit原理

shufflenet unit实现

举例说明 

---

前言

解决的问题：分组卷积的输出之和前面部分输入相关。

提出shufflenet unit. 在1x1的卷积后，把组内的特征，再分成g组，随机关联到后面的g组，然后对每一组都如此。

论文

ShufflfleNet: An Extremely Effificient Convolutional Neural Network for Mobile

Devices

https://arxiv.org/pdf/1707.01083v1.pdf

shufflenet unit原理

图片(a) 正常的group conv，输出之和前面部分输入相关;

图(b) 红色部分的特征，三块分别与下面三块相关，其他颜色一样。

图(c) channel shuffle红色部分随机与下面的三部分相关，其他颜色一样。

shufflenet unit实现

mmclassification上的实现形式如下

def channel_shuffle(x, groups):
    """Channel Shuffle operation.

    This function enables cross-group information flow for multiple groups
    convolution layers.

    Args:
        x (Tensor): The input tensor.
        groups (int): The number of groups to divide the input tensor
            in the channel dimension.

    Returns:
        Tensor: The output tensor after channel shuffle operation.
    """

    batch_size, num_channels, height, width = x.size()  # (128,64,224,224)
    assert (num_channels % groups == 0), ('num_channels should be '
                                          'divisible by groups')
    channels_per_group = num_channels // groups  # group = 2
    
    x = x.view(batch_size, groups, channels_per_group, height, width)  # (128, 2, 32, 224, 224)
    x = torch.transpose(x, 1, 2).contiguous()   # (128, 32, 2, 224, 224)
    x = x.view(batch_size, -1, height, width)   # (128, 64, 224, 224)

    return x

举例说明 

其实就是三步，详细解释下

假如input shape是(1, 15, 1, 1)，值是

1, 2, 3, 4, 5,     6, 7, 8, 9, 10,    11, 12, 13, 14, 15

(1) x = x.view(batch_size, 3, 5, height, width)
[1, 2, 3, 4, 5], 
[6, 7, 8, 9, 10], 
[11, 12, 13, 14, 15]

(2) x = torch.transpose(x, 1, 2).contiguous()
[1, 6, 11],
[2, 7, 12],
[3, 8, 13],
[4, 9, 14],
[5, 10, 15]

(3) x = x.view(batch_size, -1, height, width)  # 先列后行
1, 6, 11, 2, 7,    12, 3, 8, 13, 4,   9, 14, 5, 10, 15