tvm源码笔记 inception v1/v2/v3/v4

本文的目的是熟悉inception v1网络结构，以便对tvm前端有更深入的了解。

网络结构可以参考TensorFlow实现Inception系列结构这篇文章中的图。

论文在此：
Going Deeper with Convolutions
http://arxiv.org/abs/1409.4842

查看caffe2的inception v1模型文件，只考虑推理softmax2，其中包含的op有：

count    operation
1   type: "AveragePool"
1   type: "Dropout"
1   type: "FC"
1   type: "Softmax"
2   type: "LRN"
9   type: "Concat"
13   type: "MaxPool"
57   type: "Conv"
57   type: "Relu"

其中，LRN就是LocalRespNorm，遗留自AlexNet，后面会被BN层代替。Dropout在最后的FC前面。

keras代码实现：

GoogLeNet in Keras

---

caffe2的inception v2模型文件

论文： Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift, 4.8% test error
http://arxiv.org/abs/1502.03167

v2对v1的改进：

• 加入了BN层
• 把5x5的卷积改成了2个串行的3x3卷积

cout     operation
1   type: "FC"
1   type: "Softmax"
5   type: "MaxPool"
8   type: "AveragePool"
10   type: "Concat"
69   type: "Add"
69   type: "Conv"
69   type: "Mul"
69   type: "Relu"
69   type: "SpatialBN"

因此，从op上看：多了ADD，MUL，SpatialBN，少了LRN，DropOut。

inception v2 tensorflow源码实现

---

inception v3论文：
Rethinking the Inception Architecture for Computer Vision, 3.5% test error
http://arxiv.org/abs/1512.00567

caffe2模型文件，非官方提供，从caffe转换而来：

1   type: "Dropout"
1   type: "FC"
1   type: "Softmax"
5   type: "MaxPool"
9   type: "AveragePool"
11   type: "Concat"
94   type: "Add"
94   type: "Conv"
94   type: "Mul"
94   type: "Relu"
94   type: "SpatialBN"

从op上看，和V2没什么变化。本身的改进也主要在卷积核的大小从nxn变成了1xn和nx1。

---

inception v4论文：
Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning, 3.08% test error
http://arxiv.org/abs/1602.07261

inception v4 和 inception - resnet v2代码实现

https://github.com/titu1994/Inception-v4
https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_resnet_v2.py
https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v4.py

mxnet源码实现
https://github.com/apache/incubator-mxnet/blob/master/example/image-classification/symbols/inception-v4.py

没找到caffe2的prototxt文件，只有caffe模型文件 ：

count   operation
1       type: "constant"
1       type: "xavier"
1   type: "Dropout"
1   type: "InnerProduct"
1   type: "Softmax"
19   type: "Concat"
19   type: "Pooling"
149   type: "BatchNorm"
149   type: "Convolution"
149   type: "ReLU"
149   type: "Scale"

关于 inception系列模型更详细的资料参考：
https://blog.csdn.net/hejin_some/article/details/78636586