SUDL

A light deep learning tools box by c++

Contains

Network Architecture

1. Convolutional Neural Network
2. Normal Neural Network
3. Reccurent Neural Network with three mainstream varieties(LSTM, LSTM-peelhole, GRU)(deep architecture supported)
4. bi-directional LSTM(peephole) & GRU & RNN (deep architecture supported)

Nonlinearities

1. ReLU
2. Sigmoid
3. tanh

TODO

1. GPU supported (No Gpu for testing :( )
2. network architecture configurable by proto Done (protoc is needed to be installed first)

Compile

sh build.sh(cmake is needed)

Usage

net architecture is built by proto file that you defined, just like what the examples do.

rnn.prototxt

name: "test"
layer {
    name: "DataFeedLayer"
    type: "DataFeedLayer"
    top: "input_data"
}

layer {
    name: "WordEmbeddingLayer"
    type: "WordEmbeddingLayer"
    top: "emb1"
    bottoms: "input_data"
    fc_param {
        output_dim: 14
        input_dim: 0
    }
}

layer {
    name: "LstmCell"
    type: "LstmCell"
    top: "lstm1"
    bottoms: "emb1"

    rnn_cell_param {
        input_dim: 14
        output_dim: 16
        use_peephole: false
    }
}

layer {
    name: "LstmCell1"
    type: "LstmCell"
    top: "lstm2"
    bottoms: "lstm1"
    rnn_cell_param {
        input_dim: 16
        output_dim: 16
        use_peephole: true
    }
}

layer {
    name: "SeqFullConnSoftmaxLayer"
    type: "SeqFullConnSoftmaxLayer"
    top: "seqsoftmax1"
    bottoms: "lstm2"
    fc_param {
        input_dim: 16
        output_dim: 4
    }

}
layer {
    name: "SeqCrossEntropyLossLayer"
    type: "SeqCrossEntropyLossLayer"
    top: "loss"
    bottoms: "seqsoftmax1"
}

cnn.prototxt

name: "cnn" 
layer {
    name: "DataFeedLayer"
    type: "DataFeedLayer"
    top: "input_data"
}

layer {
    name: "ConvLayer1"
    type: "ConvLayer"
    bottoms: "input_data"
    top: "conv1"

    conv_param {
        input_dim: 1
        output_dim: 2
        kernel_x_dim: 11
        kernel_y_dim: 11
        feature_x_dim: 18
        feature_y_dim: 18
    }

}

layer {
    name: "ReluLayer1"
    type: "ReluLayer"
    bottoms: "conv1"
    top: "relu1"
}

layer {
    name: "PoolingLayer1"
    type: "PoolingLayer"
    bottoms: "relu1"
    top: "pool1"

    pool_param {
        input_dim: 2
        output_dim: 2
        pooling_x_dim: 2
        pooling_y_dim: 2
        feature_x_dim: 9
        feature_y_dim: 9
    }

}

layer {
    name: "ConvLayer2"
    type: "ConvLayer"
    bottoms: "pool1"
    top: "conv2"

    conv_param {
        input_dim: 2
        output_dim: 2
        kernel_x_dim: 4
        kernel_y_dim: 4
        feature_x_dim: 6
        feature_y_dim: 6
    }
}

layer {
    name: "ReluLayer2"
    type: "ReluLayer"
    bottoms: "conv2"
    top: "relu2"
}


layer {
    name: "FlattenLayer1"
    type: "FlattenLayer"
    bottoms: "relu2"
    top: "flat1"
}

layer {
    name: "FullConnLayer"
    type: "FullConnLayer"
    bottoms: "flat1"
    top: "full1"

    fc_param {
        input_dim: 72
        output_dim: 32
    }
}

layer {
    name: "SigmoidLayer"
    type: "SigmoidLayer"
    bottoms: "full1"
    top: "sigmoid1"

}


layer {
    name: "FullConnSoftmaxLayer"
    type: "FullConnSoftmaxLayer"
    bottoms: "sigmoid1"
    top: "full2"

    fc_param {
        input_dim: 32
        output_dim: 10
    }
}

layer {
    name: "loss"
    type: "CrossEntropyLossLayer"
    bottoms: "full2"
    top: "cross"
}