How to find the sparsity of fully connected layer representing convolutional layer?

27 Views Asked by At

We can represent the convolutional layer exactly as a fully connected layer. Suppose we have a 2-D convolutional layer, with input size = 3x32x32, number of filters = 16, filter size = 3x3, stride = 1, padding = 0. Suppose all filter entries are non-zero. What's the sparsity (proportion of zero entries to the total number of entries) of the converted fully connected layer weight matrix?