Unattended Training

Larger, slower training models or models being trained in AutoTrain mode may require extended periods of unattended training. Several steps should be taken when planning long periods of unattended training. These include:

• Make sure the “Time Remaining” field covers the time you wish to train. If not, increase the number of iterations (Options/Iterations...).
• Keep LRC on or set the learn rate (eta) coefficient to a stable value so network divergence is avoided.
• Use standard backpropagation, setting the FAST-Prop coefficient to 0.
• Set the AutoSave rate to a level that allows you to easily return to any point in the training process.
• Disable animated screensaver utilities and CPU suspend features.

The AutoSave rate should be set so that if the run results in overtraining, you can return the network to its optimal training point in a reasonable length of time. For example, if it takes 5 minutes to perform 1000 training iterations, setting the AutoSave Rate to 1000 will guarantee that you can return to any training point within 5 minutes. Another important consideration when using AutoSave during long unattended sessions is disk space. Adequate disk space must be available to store the network model at the requested rate.

Do not use animated screensavers. Animated screensavers drastically reduce the CPU time available to Qnet. Use the monitor’s low power mode (or on/off switch) to protect your monitor and reduce energy use. Other than the monitor, disable any PC power saving feature (disk or CPU suspend features) that would suspend PC operation during training.