tvm源码笔记 vggnet16/vggnet19/resnet

caffe模型文件 vggnet 16

1   type: "Input"
1   type: "Softmax"
2   type: "Dropout"
3   type: "InnerProduct"
5   type: "Pooling"
13   type: "Convolution"
15   type: "ReLU"

caffe模型文件vggnet19

1   type: "Input"
1   type: "Softmax"
2   type: "Dropout"
3   type: "InnerProduct"
5   type: "Pooling"
16   type: "Convolution"
18   type: "ReLU"

caffe2模型文件vggnet19

1   type: "Softmax"
2   type: "Dropout"
3   type: "FC"
5   type: "MaxPool"
16   type: "Conv"
18   type: "Relu"

很明显vgg19比16多了3个卷积和relu，caffe2的FC就是caffe的InnerProduct，MaxPool就是Pooling。

---

caffe2模型文件的resnet50

1   type: "AveragePool"
1   type: "FC"
1   type: "MaxPool"
1   type: "Softmax"
16   type: "Sum"
49   type: "Relu"
53   type: "Conv"
53   type: "SpatialBN"

比vgg多了BN层和sum算子。

---

caffe2的模型文件densenet121

1   type: "MaxPool"
4   type: "AveragePool"
58   type: "Concat"
121   type: "Add"
121   type: "Conv"
121   type: "Mul"
121   type: "Relu"
121   type: "SpatialBN"

---

caffe2的模型文件mobilenet_v2

1   type: "AveragePool"
1   type: "FC"
1   type: "Softmax"
10   type: "Sum"
36   type: "Relu"
53   type: "Conv"

---

caffe模型文件squeezenet

1       type: "gaussian"
1   type: "Data"
1   type: "Dropout"
1   type: "MemoryData"
1   type: "Softmax"
1   type: "SoftmaxWithLoss"
4   type: "Pooling"
8   type: "Concat"
25       type: "xavier"
26   type: "Convolution"
26   type: "ReLU"

其中，gaussian和xavier都是卷积层的weight_filler。

---

caffe的模型文件shufflenet

1       type: "constant"
2   type: "Pooling"
13   type: "Slice"
16   type: "Concat"
16   type: "ShuffleChannel"
19   type: "ConvolutionDepthwise"
37   type: "ReLU"
38   type: "Convolution"
56   type: "BatchNorm"
56   type: "Scale"
57       type: "msra"

其实，看了这么多，视觉的模型主要就几个算子构成。