FIRNet Usage
firnet -f filename [switches...]
Switches:
-l layers # of layers in the FIR network: 2 (default) or 3
-n n1 [n2] # of nodes in each hidden layer
-t n1 n2 [n3] # of taps in each layer
-p start end Range of time series values to predict
-e epochs # of epochs to train for
-c Force training (if disabled by loading state)
-r rate Learning rate
-m alpha Use momentum with parameter alpha in [0,1]
-d Linearly decrease learning rate
Notes:
- Time series files are formatted as n-dimensional real vectors separated by linefeeds.
- For an input file named series.dat, there are two corresponding files: series.net, the neural network state, and series.out, the predicted output time series.
- If a .net file does not exist for an input file, FIRNet
will train a network on the input series, save the network state, and
then continue on to the prediction phase.
- If a .net file does exist (and the -c switch is not invoked), training will be skipped and the network state in the .net file will be used for prediction.
- FIRNet only implements the two special case networks having 2 and 3
layers. (Note: the usage of "layer" here differs from standard neural
network parlance -- it refers to the set of synapses in between nodes.
So, a 2 layer FIR net really has 3 distinct sets of nodes.)
- The time tap parameters (set with the -t switch) refer to the size of the FIR filters on each layer's synapses.
The FIRNet Page