What is fractionally-strided convolution layer?

Answer: Fractionally-strided convolution (deconvolution) layer upsamples by inserting zeros between inputs and applying regular convolution.

To understand fractionally-strided convolution, it’s essential to first grasp the concept of regular (or “valid”) convolution. In a standard convolution operation, a filter (also called a kernel) is applied to the input data, and the result is a weighted sum of the values in the input region covered by the filter. This operation reduces the spatial dimensions of the input.

Fractionally-strided convolution, on the other hand, is designed to perform the opposite operation—upsampling. Instead of downsizing the input, it increases the spatial dimensions. This is achieved by inserting zeros between the elements of the input and then applying a regular convolution operation.

Here’s a step-by-step explanation of how fractionally-strided convolution works:

1. Insert Zeros (Zero Padding): Zeros are inserted between the elements of the input feature map. This effectively expands the spatial dimensions of the input.
2. Apply Regular Convolution: A regular convolution operation is then applied to the zero-padded input. The convolutional filter scans over the input, and for each position, it computes a weighted sum of the values in the filter region, just like in a standard convolution.
3. Resulting Feature Map: The result is a feature map with increased spatial dimensions compared to the input.

The term “fractionally-strided” comes from the fact that the convolutional operation effectively performs a fractional increase in the spatial resolution, as opposed to the integer reduction in resolution that occurs with standard convolution.

Conclusion:

Fractionally-strided convolution layers are commonly used in neural networks for tasks such as image segmentation, image generation, and various other tasks where upsampling is required. They provide a learnable way to increase the resolution of feature maps and capture finer details in the data.