addNet <network> [-intervals <timeIntervals> |
-ticks <ticksPerInterval>] [<net-type>]*
[<num-units> [[+|-]<group-type>]*]*
addNet creates a network with the specified name and makes
it the active network. Multiple networks may be loaded at the same time
but only one may be active and most commands will only operate on the
active network. useNet can be used to list
all networks or change the active network.
If another network with the same name exists, the user will have the option of replacing the other network or aborting. The network is created with a bias layer, called "bias", which contains a single unit.
timeIntervals is an integer specifying the maximum "real" time for which each example will be run. For non-continuous networks, this is the same as the number of events or ticks. For continuous networks, these are in abstract time units. The default is 1.
ticksPerInterval is specific to continuous networks and sets the number of ticks or activation updates per time interval. dt will default to the inverse of this value.
Networks default to standard type. Currently, the valid additional types are SRBPTT (simple recurrent backprop-through-time), CONTINUOUS (fully recurrent), and BOLTZMANN. To get recurrent backprop-through-time (which isn't really continuous), leave the ticksPerInterval 1 and use the CONTINUOUS type.
addNet can also be used like the old autoNet
to create and connect groups in the network with a single command.
If group information is listed after the network type, they will
automatically be added to the network.
Each group is created by specifying the number of units in the group followed by a series of types. Groups will be given names automatically, although these could be altered later using:
lens> setObject group(<i>).name <new-name>
The group types should be specified exactly as they would be when using addGroup or groupType, with the following differences:
If a fixed type is not specified for the first group, it will default to an INPUT group. Likewise, the last group will default to OUTPUT. All other groups will default to no fixed type (hidden), although they could be set to other fixed types by specifying a type.
In order to create an Elman context layer, it is not necessary to define it as a separate group. Giving any group the type ELMAN will cause a context group, having the same number of units, to be created as well. The original group will be Elman-connected (a special one-to-one identity mapping) to the context group and the context group will have a full projection back to the original.
The first input group will be named "input", with subsequent input groups labeled "input2", "input3", etc. Likewise for output and hidden layers. A context group will be given the same number as the corresponding main group. The context of the first hidden layer would be called "context". That of the first output layer would be called "outContext".
To create a new standard network with the name "My Network":
lens> addNet "My Network"
To create a new continuous network with 3 time intervals and 5 ticks per interval:
lens> addNet myNet -i 3 -t 5 CONTINUOUS
The command below creates a simple recurrent network named "myNet" with a maximum of 5 time intervals (each with one tick) per example, 10 inputs, 20 hidden units, 20 context units, and 3 output units. The input layer projects to the hidden layer. The hidden layer has an Elman-projection to the context layer which has a full projection back to the hidden layer. The hidden layer then projects to the output layer. The bias group projects to the hidden and output layers, but not the input and context. The output group is of type SOFT_MAX.
lens> addNet myNet -i 5 10 20 ELMAN 3 SOFT_MAX