Generally, a layer in a time delay neural network consists of feature units and their delay units. SNNS2C generates code only containing the feature units. The delay units are only additional activations in the feature unit. This is possible because every delay unit has the same link weights to it's corresponding source units as its feature unit.
So the input layer consists only of its prototype units, too. Therefore it's not possible to present the whole input pattern to the network. This is not necessary because it can be presented step by step to the inputs. This is useful for a real-time application, with the newest feature units as inputs. To mark a new sequence the init-flag (parameter of the function) can be set to 1. After this, the delays are filled when the init flag is set to 0 again. To avoid meaningless outputs the function returns NOT_VALID until the delays are filled again.
There is a new variable in the record of the header file for TDNNs. It is called ``MinDelay'' and is the minimum number of time steps which are needed to get a valid output after the init-flag was set.