Clever Geek Handbook

Theano conv2d and max_pool_2d

When implementing a Convolutional Neural Network (CNN) in theno, everyone encounters two operator variants conv2d

:

And the max-pooling implementation:

My questions:

  • What is the difference between the two implementations conv2d

    ?

  • What is the difference between using the subsample

    for argument conv2d

    and the max_pool_2d

    subsampling application after conv2d

    ?
    This is the difference between:

    conv2d( ..., subsample=(2,2) )

    and

    a = conv2d( ..., subsample = (1,1) )
max_pool_2d( a, ds = (2,2) )
+3


source to share


1 answer


In response to your first question, here is the section of Theano doc that refers to it :

There are two similar implementations for conv2d:

signal.conv2d and nnet.conv2d.

The former implements traditional 2D convolution, while the latter implements the convolutional layers present in convolutional neural networks (where 3D filters and pooling across multiple input channels).

Under the hood, they both call the same function, so the only difference is the user interface.

As for your second question, the result is different. Equivalent call:



conv2(..., subsample=(2,2))

:

conv2d(...,subsample=(1,1))[:,:,::2,::2]

In other words, it conv2d

does not accept a max over the entire join area, but rather an element in the index of [0,0]

the join area.

+6


source






More articles:


All Articles