CODEWITHSUNDEEP
pythonkeras
Dims
Dims

Reputation: 51209

How to make convolution with maxout activation?

I want my activation function select maximal value, generated by N filters of M x M convolution. This layer would convert X channel image to 1-channel one.

How to do that?

First I wrote

classifier.add(Conv2D(3, (5, 5), activation='linear')
classifier.add(MaxPooling2D(pool_size=1, strides=1))

but then thought it doesn't return 1-channel image, but returns 3 channel.

How to do the job?

Upvotes: 5

Views: 2371

Answers (3)

Daniel Möller
Daniel Möller

Reputation: 86620

This is not a direct answer, but, if you just want the result to be 1 channel, you can create a convolutional layer with only one filter. You can add it after your existing convolution, or simply change your existing convolution. 

classifier.add(Conv2D(1, (5,5),....))

Upvotes: -1

Marcin Możejko
Marcin Możejko

Reputation: 40516

So to apply this you shoud create a Lambda layer and max from Backend:

from keras import backend as K

if K.image_data_format() == "channels_first":
    channels_axis = 1
else:
    channels_axis = 3

# To apply MaxOut:

classifier.add(Lambda(lambda x: K.max(x, axis=channels_axis, keepdims=True)))

Upvotes: 2

Jonas Adler
Jonas Adler

Reputation: 10789

You can use keras.backend.max and give it the axis argument in order to take the maximum along the correct axis. Which one depends on your backend.

Upvotes: 0

Related Questions