Tropical-Convolutional-Neural-Networks

TCNN-An Alternative Practice of Tropical Convolution to Traditional Convolutional Neural Networks

Paper：https://ui.adsabs.harvard.edu/abs/2021arXiv210302096F/abstract

运行环境

python 3.6.5 pytorch 1.7.0

参数设置

Operation parameter reference the function "options_func()", and note that in "load_data()" needs to add the data set you need and its storage location.

运行参数参考 Options: -h, --help show this help message and exit

--net=NET Choosing which net structure to use [1].

--dataset=DATASET Choosing dataset [mnist ashion-mnist andomxt\CIFAR10\STL10].

--channels=CHANNELS Input channels of train data.

--dim=DIM Dimension of train data. mnist-784/STL10-27468/CIFAR10-3072

--out=OUT Dimension of out data.

--k1=K1 The count of hidden nodes.

--k2=K2 The count of hidden nodes.

--k3=K3 The count of linner layers param.

--kernel=KERNEL Kernel size.

--stride=STRIDE stride.

--lr=LR The learning rate of training.

--loss=LOSS Loss Function.

--epoch=EPOCH Training epoch.

--print_freq=PRINT_FREQ The frequency of printing.

--save_loss_freq=SAVE_LOSS_FREQ [The frequency of saving loss.

--bs=BS batch_size

--momentum=MOMENTUM Momentum of optimizer

--cuda=CUDA use cuda or not

--fitepoch=FITEPOCH the min epoch after using early stopping

--patience=PATIENCE the patience with using early stopping

--image_height=IMAGE_HEIGHT [length of image

--image_width=IMAGE_WIDTH [width of image

并注意在load_data()中需要添加你需要的数据集及其存放的位置。

运行示例

python .\TCNN.py --net net0

代码包括

• 6 tropical convolution layers

MinPlus-Sum_Conv Layer (MinP-S)

MaxPlus-Sum-Conv Layer (MaxP-S)

MinPlus-Max-Conv Layer (MinP-Max)

MaxPlus-Min-Conv Layer (MaxP-Min)

MinPlus-Min-Conv Layer (MinP-Min)

MaxPlus-Max-Conv Layer (MaxP-Max)

• 6 network structures

See the paper for details.