Is this right?

[XP-CHRC]

def concat_shuffle_split_abstract(x, y): x_cp_1 = x[:, :, :, 0::2] y_cp_1 = y[:, :, :, 0::2] x=tf.concat([x_cp_1, y_cp_1], axis=3)

x_cp_2 = x[:, :, :, 1::2]
y_cp_2 = y[:, :, :, 1::2]
y = tf.concat([x_cp_2, y_cp_2], axis=3)

return x, y

Why x_cp_2 = (local variable)x[:, :, :, 1::2] ? i dont know the reason. Looking forward to your reply.

[610265158]

Oh , that was modified for tensorflowlite, because tflite not support 5-d transpose, so i make it by slice. just take half of the x and y, stride 2 in channel , and concat them.

I think you can refer to the 'simple concat shuffle split' by transpose and split. as many othe implements.

And sorry it bothers you.There is no needs to concern that a lot, other net structures works the same, basicly.

good luck

[XP-CHRC]

Oh , that was modified for tensorflowlite, because tflite not support 5-d transpose, so i make it by slice. just take half of the x and y, stride 2 in channel , and concat them.

I think you can refer to the 'simple concat shuffle split' by transpose and split. as many othe implements.

And sorry it bothers you.There is no needs to concern that a lot, other net structures works the same, basicly.

good luck @610265158 啊啊啊，我懂你的意思，其实我想问的是x_cp_2 = x[:, :, :, 1::2]中x不是一个局部变量吗？它应该覆盖了输入变量x，所以想问一下这块是怎么回事？

[KarlKulator]

Oh , that was modified for tensorflowlite, because tflite not support 5-d transpose, so i make it by slice. just take half of the x and y, stride 2 in channel , and concat them.

I think you can refer to the 'simple concat shuffle split' by transpose and split. as many othe implements.

And sorry it bothers you.There is no needs to concern that a lot, other net structures works the same, basicly.

good luck

Hi, did you check if this line is correct?: (shufflenet.py:29) x_cp_2 = x[:, :, :, 1::2]

The x that is taken here comes from the local variable x (calculated in line 26) and not from the input variable x.

Additionally I get different results, when I exchange concat_shuffle_split_abstract with concat_shuffle_split. So these two do not seems to be the exact same implementation of the shuffling layer.

[610265158]

Yes, that is a mistake. It is better use the norm shuffle op (concat_shuffle_split).

And to correct this mistake need to retrain the model.

Now i intend to reimplement it based on mobilenetv3, when i get time.