Hands-On Machine Learning for Algorithmic Trading

How to scan the input – strides and padding

The stride defines the step size used for scanning the input, that is, the number of pixels to shift horizontally and vertically. Smaller strides scan more (overlapping) areas, but are computationally more expensive.

The following three options are commonly used when the filter does not fit the input perfectly, and partially crosses the image boundary during the scan:

• Valid convolution: Discard scans where image and filter do not perfectly match
• Same convolution: Zero-pad the input to produce a feature map of equal size
• Full convolution: Zero-pad the input so that each pixel is scanned an equal number of times, including pixels at the border (to avoid oversampling pixels closer to the center)