Is it possible to generate an 1D dimensional output of a 2D convolutional layer in Keras?

Question

I'm trying to apply convolutional neural networks for dealing with a 2D input, which is a 2X300 matrix. It is basically a matrix with 2 lines, where each line is a vector of 300 positions.

I would like to apply a kernel of size 2X1 (two lines and one column). The idea is to apply the kernel to each position i of the two vectors. Intuitively, I think that this convolution operation would generate an output of size 1X300. That is, I think that the output will be an unidimensional vector with 300 columns. Am I right?

I would like to include a convolutional layer like this:

layers.Conv2D(10, kernel_size=(2, 1), activation="relu",name="conv1")

That makes sense? Will this layer generate a one-dimensional vector of 300 positions?

Answer 1

No, an input to a 2D convolutional layer needs to have 3 dimensions (typically width,height,channel). What you are looking for is a 1D convolutional layer, which operates on a sequence of data (typically timestep,channel). In your case, the channel is the first dimension, ie, the dimension where all inputs are taken into account for each step, and your second dimension is the steps (over each of which the convolutional kernel is shifted). This needs to be specified with the data_format keyword as 'channels_last' would be default for tensorflow/keras.

model.add(Conv1D(1, kernel_size=(1,), data_format='channels_first', activation = 'relu', input_shape=(2, 300)))

This will give you an output of shape (1,300) . The first parameter is the number of output channels (1 in your case). The kernel size is 1, as you do not take into account samples from other steps for the convolution, only the two channels (1D convolution is always performed across all channels). If you choose a larger kernel size, padding='same' might be required as an additional parameter to get an output of the same number of steps.