MaxOut Operator. (TODO, Not Assigned. Welcome people to implement it.)

[luotao1]

Reference:

• MaxOutLayer.cpp
• MaxOut api

This Op is TODO, welcome people to implement it.

[zhouxiao-coder]

I took a look at this layer and its related paper.
I'm not sure if I misunderstood its formula, but it seems to me like it can be implemented by reshape->maxPooling. Something similar to the following numpy code:

import numpy as np
n, c, h, w, group = 2, 4, 4, 4, 2
x = np.random.normal(size=(n, c, h, w))
# maxout: reshape the channel dimension to 2 dimensions(group, channel/group), 
# then max pooling on the channel/group dimension
y = np.reshape(x, newshape=(n, group, c/group, h, w)).max(axis=2)

[qingqing01]

@zhouxiao-coder Good! You are right, Reshape + reduce_max can do this. And the maxout in TensorFlow is more flexible than our implementation.

[chengduoZH]

@qingqing01
This operation is easy to implement if our top-k op supports the axis property.