Mastering Machine Learning Algorithms

Transpose convolution

A transpose convolution (sometimes wrongly called deconvolution, even if the mathematical definition is different) is not very different from a standard convolution, but its goal is to rebuild a structure with the same features as the input sample. Let's suppose that the output of a convolutional network is the feature map X ∈ ℜ^{w' × h' × p} and we need to build an output element Y ∈ ℜ^{w × h × 3} (assuming the w and h are the original dimensions). We can achieve this result by applying a transpose convolution with appropriate strides and padding to X. For example, let's suppose that X ∈ ℜ^{128 × 128 × 256} and our output must be 512 × 512 × 3. The last transpose convolution must learn three filters with strides set to four ...