CS231n：VGGNet总内存计算及网络结构细节技术问询

I’ve worked through the CS231n convolutional networks material extensively, so let’s break down VGGNet in a hands-on, case-based way—it’s such a foundational architecture, and its simplicity makes it perfect for understanding core convnet patterns.

Core Building Blocks #

VGGNet’s design is intentionally uniform, relying on just two repeating layer types across all stages:

• CONV Layers: Every convolutional layer uses a 3x3 filter with a stride of 1 and padding of 1. This setup guarantees the spatial dimensions of the input feature map stay identical after convolution (padding=1 offsets the edge loss from the 3x3 filter, and stride=1 means we don’t skip any pixels).
• POOL Layers: All pooling layers are 2x2 max-pooling operations with a stride of 2 and no padding. This halves both the width and height of the feature map every time it’s applied.

Feature Map Size Evolution Walkthrough #

Let’s track how dimensions change using a standard 224x224 RGB input (the default for VGGNet):

1. Input: 224x224x3 (RGB image)
2. First CONV Block: 2-3 stacked 3x3 CONV layers → output remains 224x224x64 (64 is the starting filter count for VGG16)
3. First POOL Layer: 2x2 max-pool → output shrinks to 112x112x64
4. Second CONV Block: 2-3 stacked 3x3 CONV layers (filter count doubles to 128) → stays 112x112x128
5. Second POOL Layer: 2x2 max-pool → 56x56x128
6. Third CONV Block: 3 stacked 3x3 CONV layers (filter count doubles to 256) → stays 56x56x256
7. Third POOL Layer: 2x2 max-pool → 28x28x256
8. Fourth CONV Block: 3 stacked 3x3 CONV layers (filter count doubles to 512) → stays 28x28x512
9. Fourth POOL Layer: 2x2 max-pool → 14x14x512
10. Fifth CONV Block: 3 stacked 3x3 CONV layers → stays 14x14x512
11. Fifth POOL Layer: 2x2 max-pool → 7x7x512

After these conv/pool stages, the 7x7x512 feature map feeds into three fully connected (FC) layers: two FC layers with 4096 units each, followed by a final FC layer with 1000 units (for ImageNet classification).

The real genius of VGGNet is its consistency—once you grasp the "conv stack → halving pool" pattern, you can easily predict how feature maps evolve through the entire network.

内容的提问来源于stack exchange，提问作者Dang Manh Truong