Performing 1d convolution using 2d kernel in keras

Question

I am currently working on a CNN network, in which i want to apply a 2d kernel on a image, but it only has to perform 1d convolution, meaning that it only has to move along one axis (x-axis in this case).

The shape of the kernel is same as the y-axis of the image. The number of filters applied is not a concern at the moment.

An example: Given a image of size (6,3,3) = (rows, cols, color_channel)

How should i perform a 1d convolution given a 2d filter?

Tried what was suggested by @Marcin Możejko

dim_x = 3
dim_y = 6
color_channels = 3
#model.add(ZeroPadding2D((6,4),input_shape=(6,3,3)))
model.add(Conv2D(filters = 32,kernel_size=(dim_y,1) , activation='linear' , input_shape = (6,3,3)))
print model.output_shape
model.add(Reshape((dim_x,color_channels)))

Error:

The total size of the new array must be unchanged

Answer 1

Assuming that your image shape=(dim_x, dim_y, img_channels) you can obtain a 1D convolution by setting:

conv1d_on_image = Convolution2D(output_channels, 1, dim_y, border_mode='valid')(input)

Remember that the output from this layer would have shape (dim_x, 1, output_channels). If you want your input to be sequential you may use the Reshape layer by setting:

conv1d_on_image = Reshape((dim_x, output_channels))(conv1d_on_image)

This would produce output with shape (dim_x, output_channels).

An interesting fact is that this is exactly the way how Conv1D works in Keras with tf backend.