Inception-ResNet v1
在 Inception 模块中引入 Residual 连接，提高学习速度和网络性能。
总体结构
graph TD;
image["Image"];
stem["Stem"];
inception_a["5 x Inception-ResNet-A"];
reduction_a["Reduction-A"];
inception_b["10 x Inception-ResNet-B"];
reduction_b["Reduction-B"];
inception_c["5 x Inception-ResNet-C"];
pool["AvgPool k=8"];
dropout["Dropout 0.2"];
classifier["softmax classifier"];
image --> |"3*299*299"| stem;
stem --> |"256*35*35"| inception_a;
inception_a --> |"256*35*35"| reduction_a;
reduction_a --> |"896*17*17"| inception_b;
inception_b --> |"896*17*17"| reduction_b;
reduction_b --> |"1792*8*8"| inception_c;
inception_c --> |"1792*8*8"| pool;
pool --> |"1792"| dropout;
dropout --> |"1792"| classifier;
内部模块
- Stem
graph TD;
image["Image"];
conv1["Conv k=3 s=2 ReLU"];
conv2["Conv k=3 ReLU"];
conv3["Conv k=3 p=1 ReLU"];
conv4["Conv k=3 s=2 ReLU"];
conv5["Conv k=1 ReLU"];
conv6["Conv k=3 ReLU"];
conv7["Conv k=3 s=2 ReLU"];
bn["BN ReLU"];
output["Output"];
image --> |"3*299*299"| conv1;
conv1 --> |"32*149*149"| conv2;
conv2 --> |"32*147*147"| conv3;
conv3 --> |"64*147*147"| conv4;
conv4 --> |"64*73*73"| conv5;
conv5 --> |"80*73*73"| conv6;
conv6 --> |"192*71*71"| conv7;
conv7 --> |"256*35*35"| bn;
bn --> |"256*35*35"| output;
- Inception-ResNet-A
graph TD;
input["Input"];
conv_1x1["Conv k=1 ReLU"];
conv_3x3_a_reduce["Conv k=1 ReLU"];
conv_3x3_a["Conv k=3 p=1 ReLU"];
conv_3x3_b_reduce["Conv k=1 ReLU"];
conv_3x3_b1["Conv k=3 p=1 ReLU"];
conv_3x3_b2["Conv k=3 p=1 ReLU"];
concat["DepthConcat"];
concat_proj["Conv k=1"];
sum["Sum"];
bn["BN ReLU"];
output["Output"];
input --> |"256*35*35"| conv_1x1;
conv_1x1 --> |"32*35*35"| concat;
input --> |"256*35*35"| conv_3x3_a_reduce;
conv_3x3_a_reduce --> |"32*35*35"| conv_3x3_a;
conv_3x3_a --> |"32*35*35"| concat;
input --> |"256*35*35"| conv_3x3_b_reduce;
conv_3x3_b_reduce --> |"32*35*35"| conv_3x3_b1;
conv_3x3_b1 --> |"32*35*35"| conv_3x3_b2;
conv_3x3_b2 --> |"32*35*35"| concat;
concat --> |"96*35*35"| concat_proj;
concat_proj --> |"256*35*35"| sum;
input --> |"256*35*35"| sum;
sum --> |"256*35*35"| bn;
bn --> |"256*35*35"| output;
- Inception-ResNet-B
graph TD;
input["Input"];
conv_1x1["Conv k=1 ReLU"];
conv_reduce["Conv k=1 ReLU"];
conv_1x7["Conv k=[1,7] p=[0,3] ReLU"];
conv_7x1["Conv k=[7,1] p=[3,0] ReLU"];
concat["DepthConcat"];
concat_proj["Conv k=1"];
sum["Sum"];
bn["BN ReLU"];
output["Output"];
input --> |"896*17*17"| conv_1x1;
conv_1x1 --> |"128*17*17"| concat;
input --> |"896*17*17"| conv_reduce;
conv_reduce --> |"128*17*17"| conv_1x7;
conv_1x7 --> |"128*17*17"| conv_7x1;
conv_7x1 --> |"128*17*17"| concat;
concat --> |"256*17*17"| concat_proj;
concat_proj --> |"896*17*17"| sum;
input --> |"896*17*17"| sum;
sum --> |"896*17*17"| bn;
bn --> |"896*17*17"| output;
- Inception-ResNet-C
graph TD;
input["Input"];
conv_1x1["Conv k=1 ReLU"];
conv_reduce["Conv k=1 ReLU"];
conv_1x3["Conv k=[1,3] p=[0,1] ReLU"];
conv_3x1["Conv k=[3,1] p=[1,0] ReLU"];
concat["DepthConcat"];
concat_proj["Conv k=1"];
sum["Sum"];
bn["BN ReLU"];
output["Output"];
input --> |"1792*8*8"| conv_1x1;
conv_1x1 --> |"192*8*8"| concat;
input --> |"1792*8*8"| conv_reduce;
conv_reduce --> |"192*8*8"| conv_1x3;
conv_1x3 --> |"192*8*8"| conv_3x1;
conv_3x1 --> |"192*8*8"| concat;
concat --> |"384*8*8"| concat_proj;
concat_proj --> |"1792*8*8"| sum;
input --> |"1792*8*8"| sum;
sum --> |"1792*8*8"| bn;
bn --> |"1792*8*8"| output;
- Reduction A
graph TD;
input["Input"];
conv_3x3_a["Conv k=3 s=2 ReLU"];
conv_3x3_b_reduce["Conv k=1 ReLU"];
conv_3x3_b1["Conv k=3 p=1 ReLU"];
conv_3x3_b2["Conv k=3 s=2 ReLU"];
pool["MaxPool k=3 s=2"];
concat["DepthConcat"];
bn["BN ReLU"];
output["Output"];
input --> |"256*35*35"| conv_3x3_a;
input --> |"256*35*35"| conv_3x3_b_reduce;
input --> |"256*35*35"| pool;
pool --> |"256*17*17"| concat;
conv_3x3_a --> |"384*17*17"| concat;
conv_3x3_b_reduce --> |"192*35*35"| conv_3x3_b1;
conv_3x3_b1 --> |"192*35*35"| conv_3x3_b2;
conv_3x3_b2 --> |"256*17*17"| concat;
concat --> |"896*17*17"| bn;
bn --> |"896*17*17"| output;
- Reduction B
graph TD;
input["Input"];
conv_3x3_a_reduce["Conv k=1 ReLU"]
conv_3x3_a["Conv k=3 s=2 ReLU"];
conv_3x3_b_reduce["Conv k=1 ReLU"];
conv_3x3_b1["Conv k=3 p=1 ReLU"];
conv_3x3_b2["Conv k=3 s=2 ReLU"];
conv_3x3_c_reduce["Conv k=1 ReLU"];
conv_3x3_c["Conv k=3 s=2 ReLU"];
pool["MaxPool k=3 s=2"];
concat["DepthConcat"];
bn["BN ReLU"];
output["Output"];
input --> |"896*17*17"| conv_3x3_a_reduce;
conv_3x3_a_reduce --> |"256*17*17"| conv_3x3_a;
conv_3x3_a --> |"384*8*8"| concat;
input --> |"896*17*17"| conv_3x3_b_reduce;
conv_3x3_b_reduce --> |"256*17*17"| conv_3x3_b1;
conv_3x3_b1 --> |"256*17*17"| conv_3x3_b2;
conv_3x3_b2 --> |"256*8*8"| concat;
input --> |"896*17*17"| conv_3x3_c_reduce;
conv_3x3_c_reduce --> |"256*17*17"| conv_3x3_c;
conv_3x3_c --> |"256*8*8"| concat;
input --> |"896*17*17"| pool;
pool --> |"896*8*8"| concat;
concat --> |"1792*8*8"| bn;
bn --> |"1792*8*8"| output;
实验结果
- ILSVRC 2012 上 144 切割的实验结果
| Network | Crops | Top-1 Error | Top-5 Error |
|---|---|---|---|
| ResNet-151 | dense | 19.4% | 4.5% |
| Inception-v3 | 144 | 18.9% | 4.3% |
| Inception-v4 | 144 | 17.7% | 3.8% |
| Inception-ResNet-v1 | 144 | 18.8% | 4.3% |
参考文献
- Szegedy, C., Io e, S., Vanhoucke, V.: Inception-v4, inception-resnet and the impact of residual connections on learning. arXiv:1602.07261 (2016)