what is the difference between tf.nn.convolution and tf.nn.conv2d?

Question

I want to make dilated convolution on a feature. In tensorflow I found tf.nn.convolution and tf.nn.conv2d . But tf.nn.conv2d doesn't seem to support dilated convolution.

So I tried using tf.nn.convolution .

Do the 2 formulations below give the same result?

tf.nn.conv2d(x, w, strides=[1, 1, 2, 2], padding='SAME',data_format='NCHW')

tf.nn.convolution(x, w, strides=[1, 1, 2, 2], padding='SAME',data_format='NCHW')

Answer 1

To use dilated convolution, you need to use the following function:

tf.nn.atrous_conv2d(value, filters, rate, padding, name=None)

Reference: https://www.tensorflow.org/versions/r0.12/api_docs/python/nn/convolution#atrous_conv2d

Answer 2

Functionally, dilations augument in tf.nn.conv2d is the same as dilations_rate in tf.nn.convolution as well as rate in tf.nn.atrous_conv2d .

They all represent the rate by which we upsample the filter values by inserting zeros across the height and width dimensions. The dilation factor for each dimension of input specifying the filter upsampling/input downsampling rate otherwise known as atrous convolution.

The usage differs slightly. Let rate k >= 1 represent the dilation rate,

• in tf.nn.conv2d , the rate k is passed as list of ints [1, k, k,1] for [batch, rate_height, rate_width, channel] .

• in tf.nn.convolution , rate k is passed as a sequence of N ints as [k,k] for [rate_height, rate_width] .

• in tf.nn.atrous_conv2d , rate k is a positive int32 , a single value for both height and width. This library is deprecated and exists only for backwards compatibility.

Hope it helps :)