public FullyConnectedLayer(LayerDefinition def) : base()
{
// required
this.out_depth = def.num_neurons;
// optional
this.l1_decay_mul = def.l1_decay_mul != double.MinValue ? def.l1_decay_mul : 0.0;
this.l2_decay_mul = def.l2_decay_mul != double.MinValue ? def.l2_decay_mul : 1.0;
// computed
this.num_inputs = def.in_sx * def.in_sy * def.in_depth;
this.out_sx = 1;
this.out_sy = 1;
this.type = "fc";
// initializations
var bias = def.bias_pref != double.MinValue ? def.bias_pref : 0.0;
this.filters = new List <Volume>();
for (var i = 0; i < this.out_depth; i++)
{
this.filters.Add(new Volume(1, 1, this.num_inputs));
}
this.biases = new Volume(1, 1, this.out_depth, bias);
}
public ConvLayer(LayerDefinition def) : base()
{
// required
this.out_depth = def.n_filters;
this.sx = def.sx; // filter size. Should be odd if possible, it's cleaner.
this.in_depth = def.in_depth;
this.in_sx = def.in_sx;
this.in_sy = def.in_sy;
// optional
this.sy = def.sy != int.MinValue ? def.sy : this.sx;
this.stride = def.stride != int.MinValue ? def.stride : 1; // stride at which we apply filters to input volume
this.pad = def.pad != int.MinValue ? def.pad : 0; // amount of 0 padding to add around borders of input volume
this.l1_decay_mul = def.l1_decay_mul != double.MinValue ? def.l1_decay_mul : 0.0;
this.l2_decay_mul = def.l2_decay_mul != double.MinValue ? def.l2_decay_mul : 1.0;
// computed
// note we are doing floor, so if the strided convolution of the filter doesnt fit into the input
// volume exactly, the output volume will be trimmed and not contain the (incomplete) computed
// final application.
this.out_sx = (int)Math.Floor((double)(def.in_sx + this.pad * 2 - this.sx) / this.stride + 1);
this.out_sy = (int)Math.Floor((double)(def.in_sy + this.pad * 2 - this.sy) / this.stride + 1);
this.type = "conv";
// initializations
var bias = def.bias_pref != double.MinValue ? def.bias_pref : 0.0;
this.filters = new List <Volume>();
for (var i = 0; i < this.out_depth; i++)
{
this.filters.Add(new Volume(this.sx, this.sy, this.in_depth));
}
this.biases = new Volume(1, 1, this.out_depth, bias);
}
public ReLULayer(LayerDefinition def) : base()
{
// computed
this.out_sx = def.in_sx;
this.out_sy = def.in_sy;
this.out_depth = def.in_depth;
this.type = "relu";
}
public SVMLayer(LayerDefinition def) : base()
{
// computed
this.num_inputs = def.in_sx * def.in_sy * def.in_depth;
this.out_depth = this.num_inputs;
this.out_sx = 1;
this.out_sy = 1;
this.type = "svm";
}
public RegressionLayer(LayerDefinition def) : base()
{
// computed
this.num_inputs = def.in_sx * def.in_sy * def.in_depth;
this.out_depth = this.num_inputs;
this.out_sx = 1;
this.out_sy = 1;
this.type = "regression";
}
public DropoutLayer(LayerDefinition def) : base()
{
// computed
this.out_sx = def.in_sx;
this.out_sy = def.in_sy;
this.out_depth = def.in_depth;
this.type = "dropout";
this.drop_prob = def.drop_prob != double.NaN ? def.drop_prob : 0.5;
this.dropped = new bool[this.out_sx * this.out_sy * this.out_depth];
}
public InputLayer(LayerDefinition def) : base()
{
// required: depth
this.out_depth = def.out_depth;
// optional: default these dimensions to 1
this.out_sx = def.out_sx;
this.out_sy = def.out_sy;
// computed
this.type = "input";
}
