论文笔记 Very Deep Convolutional Networks for Large-Scale Visual Recognition - ICLR 2014

VGG Very Deep Convolutional Networks for Large-Scale Visual Recognition

Karen Simonyan and Andrew Zisserman ICLR, 2014 (PDF) (Citations 73354)

Contribution

• 通过堆叠多个3x3的卷积核来替代大尺度卷积核（减少所需参数，两个3x3的卷积核和一个5x5的卷积核具有相同的感受野，三个3x3的卷积核和一个7x7的卷积核具有相同的感受野）。
• AlexNet提出的LRN实际用处不大（可以使用BN）。

Details

• ILSVRC’14 2nd in classification, 1st in localization
• Use VGG16 or VGG19 (VGG19 only slightly better, more memory)
• Use ensembles for best results
• FC7 features generalize well to other tasks

参数计算（VGG16, not counting biases）

Layer	input size	memory	params
INPUT	[224×224×3]	224×224×3=150K	0
CONV3-64	[224×224×64]	224×224×64=3.2M	(3×3×3)×64=1,728
CONV3-64	[224×224×64]	224×224×64=3.2M	(3×3×64)×64=36,864
POOL2	[112×112×64]	112×112×64=800K	0
CONV3-128	[112×112×128]	112×112×128=1.6M	(3×3×64)×128=73,728
CONV3-128	[112×112×128]	112×112×128=1.6M	(3×3×128)×128=147,456
POOL2	[56×56×128]	56×56×128=400K	0
CONV3-256	[56×56×256]	56×56×256=800K	(3×3×128)×256=294,912
CONV3-256	[56×56×256]	56×56×256=800K	(3×3×256)×256=589,824
CONV3-256	[56×56×256]	56×56×256=800K	(3×3×256)×256=589,824
POOL2	[28×28×256]	28×28×256=200K	0
CONV3-512	[28×28×512]	28×28×512=400K	(3×3×256)×512=1,179,648
CONV3-512	[28×28×512]	28×28×512=400K	(3×3×512)×512=2,359,296
CONV3-512	[28×28×512]	28×28×512=400K	(3×3×512)×512=2,359,296
POOL2	[14×14×512]	14×14×512=100K	0
CONV3-512	[14×14×512]	14×14×512=100K	(3×3×512)×512=2,359,296
CONV3-512	[14×14×512]	14×14×512=100K	(3×3×512)×512=2,359,296
CONV3-512	[14×14×512]	14×14×512=100K	(3×3×512)×512=2,359,296
POOL2	[7×7×512]	7×7×512=25K	0
D	[1×1×4096]	4096	7×7×512×4096=102,760,448
FC	[1×1×4096]	4096	4096×4096 = 16,777,216
FC	[1×1×1000]	1000	4096×1000 = 4,096,000

TOTAL memory: 24M × 4 bytes ≈ 96MB / image (for a forward pass)

TOTAL params: 138M parameters

Notes:

• Most memory is in early CONV
• Most params are in late FC

References

• cs231n