Configuration File

The configuration file CNNTestApp_conf.cnn is a text file included in \CNN_APP\Release\ that allows the user to select the CDF files used for training (or testing) and to specify the CNN training and configuration parameters.

CDF file selection:

• item_no is the ID of the CDF training and testing files.

• dir is the directory of the CDF files. Please make sure that it includes no spaces.

• title is the title of the CDF training and testing files. Note that the application requires naming the CDF files as title_ID_train.CDF and title_ID_test.CDF.

• id is the ID of the CDF training and testing files.

• btrain determines whether the application will be used for CNN training or testing. Use btrain t to tell the application to use the data in title_ID_train.CDF to train a CNN. If any value other than t is used for this parameter, the application will run CNN testing using the data in title_ID_test.CDF.

CNN Training Parameters:

• no_iter is the no of iterations.

• no_runs is the total number of runs.

• update 0: batch, 1: online, N>1: update sensitivities every N iteration.

• tr_rate is the train rate.

• lp is the BP eps.

• decay_lp is the lp decay factor.

• stop_ce is the target train CE to check stopping criteria.

• delta_mse is the train stop criteria over delta MSE.

• flags: 4 output flags for MATLAB, CNN, AS and LOG files. Use flags 1 1 1 1 to activate the 4 outputs.

CNN Configuration Parameters:

• ch_bias: please keep this as 0.

• in_size: please keep this as 1.

• fs is the filter size.

• ssx is the subsampling factor.

• ssy: use any number here. It has no effect on the results.

• fn: specifies the activation functions: _tanh = 0, _sigm = 1, _linear = 2, _binary = 3, _relu = 4, _lincut = 5.

• ctype: please keep this as 0.

• ptype is the subsampling mode: max. pool = 0, avg. pool=1

• cnn_nol is the number of CNN layers (including the input layer).

• mlp_nol is the number of MLP layers (including the output layer).

• width is the number of samples in the input frame (i.e. array).

• heigth use any number here. It has no effect on the results.

• min_non indicates the number of neuron in each hidden CNN and MLP layer. For example, if your CNN has 2 hidden CNN layers each with 60 neurons and 3 hidden MLP layers each with 10 neurons, you need to write: min_non 0 60 60 10 10 10 0.

• max_non: please use the same value specified for min_non.

For more information regarding the 1D CNN parameters and structure, please refer to the following papers:

Real-Time Patient-Specific ECG Classification by 1-D Convolutional Neural Networks.

Real-time vibration-based structural damage detection using one-dimensional convolutional neural networks.