Kernel size, stride, padding in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Kernel size, stride, padding

Which metric matters for Kernel size, stride, padding and WHY

When working with convolution layers, the key metrics to understand are the output size and receptive field. These depend on kernel size, stride, and padding.

Output size tells us how big the feature map will be after convolution. It affects model speed and memory.

Receptive field is how much of the input each output pixel 'sees'. Larger receptive fields capture more context.

We measure output size using this formula:
Output = floor((Input + 2 * Padding - Kernel Size) / Stride) + 1

Understanding these helps us design models that balance detail and efficiency.

Confusion matrix or equivalent visualization

For kernel size, stride, and padding, a confusion matrix is not used. Instead, we visualize how these parameters affect output size.

Input size: 7
Kernel size: 3
Stride: 2
Padding: 1

Output size = floor((7 + 2*1 - 3) / 2) + 1 = floor((7 + 2 - 3)/2) + 1 = floor(6/2) + 1 = 3 + 1 = 4

Visual:
Input: 7 pixels
Kernel slides over input with stride 2 and padding 1
Output: 4 pixels
    

Precision vs Recall tradeoff analogy for Kernel size, stride, padding

Think of kernel size, stride, and padding like taking photos of a scene:

• Kernel size is the size of your camera lens. Bigger lens sees more detail but takes longer to process.
• Stride is how far you move the camera between shots. Bigger stride means fewer shots, faster but less detail.
• Padding is adding extra space around the scene so you don't miss edges.

Tradeoff:

• Big kernel + small stride + padding = detailed but slow and large output.
• Small kernel + big stride + no padding = fast but may miss details at edges.

Choosing these is like balancing speed and detail in your photos.

What "good" vs "bad" values look like for Kernel size, stride, padding

Good values:

• Kernel size: Usually 3 or 5 for good detail without too much cost.
• Stride: 1 or 2 to keep enough output resolution.
• Padding: Enough to keep output size reasonable and preserve edges.

Bad values:

• Kernel size too large (e.g., 11) can over-smooth and slow down training.
• Stride too large (e.g., 4 or more) causes output to be too small, losing detail.
• No padding when needed causes output to shrink too much and lose edge info.

Common pitfalls with Kernel size, stride, padding

• Not calculating output size correctly leads to shape mismatch errors.
• Using no padding causes output to shrink after each layer, losing info.
• Too large stride skips important details, hurting accuracy.
• Ignoring receptive field size can cause model to miss important context.
• Changing these parameters without adjusting others can break model architecture.

Self-check question

Your convolution layer has input size 28, kernel size 5, stride 1, and no padding. What is the output size? Is this good if you want to keep input size?

Calculate output size:
Output = floor((28 + 2*0 - 5) / 1) + 1 = floor(23) + 1 = 24

Output size is 24, smaller than input 28. So output shrinks.

If you want to keep input size, you should add padding = (kernel size - 1) / 2 = 2.

With padding 2:
Output = floor((28 + 2*2 - 5)/1) + 1 = floor(27) + 1 = 28

Now output size matches input size, which is good for many models.

Key Result

Output size depends on kernel size, stride, and padding; correct calculation ensures proper model design.