static Symbol get_ocrnet(int batchSize)
{
using (NameScop scop = new NameScop())
{
var data = Symbol.Variable("data");
var label = Symbol.Variable("softmax_label");
var conv1 = Symbol.Convolution(data, kernel: new Shape(5, 5), numFilter: 32);
var pool1 = Symbol.Pooling(conv1, poolType: Symbol.PoolingPoolType.Max, kernel: new Shape(2, 2),
stride: new Shape(1, 1));
var relu1 = Symbol.Activation(data: pool1, actType: Symbol.ActivationActType.Relu);
var conv2 = Symbol.Convolution(relu1, kernel: new Shape(5, 5), numFilter: 32);
var pool2 = Symbol.Pooling(data: conv2, poolType: Symbol.PoolingPoolType.Avg,
kernel: new Shape(2, 2), stride: new Shape(1, 1));
var relu2 = Symbol.Activation(data: pool2, actType: Symbol.ActivationActType.Relu);
var conv3 = Symbol.Convolution(data: relu2, kernel: new Shape(3, 3), numFilter: 32);
var pool3 = Symbol.Pooling(data: conv3, poolType: Symbol.PoolingPoolType.Avg,
kernel: new Shape(2, 2), stride: new Shape(1, 1));
var relu3 = Symbol.Activation(data: pool3, actType: Symbol.ActivationActType.Relu);
var flatten = Symbol.Flatten(data: relu3);
var fc1 = Symbol.FullyConnected(data: flatten, numHidden: 512);
var fc21 = Symbol.FullyConnected(data: fc1, numHidden: 10);
var fc22 = Symbol.FullyConnected(data: fc1, numHidden: 10);
var fc23 = Symbol.FullyConnected(data: fc1, numHidden: 10);
var fc24 = Symbol.FullyConnected(data: fc1, numHidden: 10);
var fc2 = Symbol.Concat(new Symbol[] { fc21, fc22, fc23, fc24 }, 4, dim: 0);
label = Symbol.Transpose(data = label);
label = Symbol.Reshape(data = label, shape: new Shape((uint)(batchSize * 4)));
return(Symbol.SoftmaxOutput("softmax", fc2, label));
}
}
