卷积神经网络层尺寸疑问：特征图维度与数量变化困惑

Hey there! Let's unpack your confusion about the CNN diagram step by step—this is such a common question when diving into backprop and CNN architectures, so you’re totally on the right track by questioning these details.

一、关于特征图尺寸从32×32到18×18的疑惑 #

Your calculation of $(32-5+1) \times (32-5+1) = 28 \times 28$ is 100% correct for a standard 5×5 convolution with no padding (padding=0) and stride=1. So why does the diagram show 18×18? Here are the most likely explanations:

• Larger convolution kernel: If the diagram uses a 15×15 kernel (instead of 5×5) with stride=1 and no padding, the output size becomes $(32-15+1) = 18$, which matches exactly. While 15×15 kernels are less common than small ones like 3×3 or 5×5, they’re sometimes used for tasks needing global context.
• Diagram simplification/typo: Sometimes educational diagrams fudge numbers a bit to make a core point (like demonstrating channel growth without getting bogged down in exact size math). The key takeaway here is that output size always depends on four variables: input size, kernel size, stride, and padding. You can calculate it reliably with this formula:

  Output Size = floor((Input Size - Kernel Size + 2*Padding)/Stride) + 1
  

二、关于特征图数量的疑问 #

This one clicks once you map it to CNN fundamentals:

• The first layer’s 3 feature maps are almost certainly the input channels (like a standard RGB image, which has 3 color channels: red, green, blue). That’s why it matches the input size of 32×32—this is just the raw input data.
• The second layer’s 32 feature maps come from using 32 different convolution kernels on the input. Each kernel learns to detect a specific feature (edges, textures, etc.) across all 3 input channels, and each kernel outputs one single-channel feature map. Stacking 32 of these gives you the 32-channel output. Choosing 32 is a standard hyperparameter choice—you could pick 16, 64, etc., but 32 is a common starting point for balancing model capacity and compute cost.

Hope this clears things up! Feel free to follow up if you want to dig deeper into any of these details.

内容的提问来源于stack exchange，提问作者koryakinp