train_class_mlp🔗
Short description🔗
train_class_mlp — Train a multilayer perceptron.
Signature🔗
train_class_mlp( class_mlp MLPHandle, integer MaxIterations, real WeightTolerance, real ErrorTolerance, out real Error, out real ErrorLog )
Description🔗
train_class_mlp trains the multilayer perceptron (MLP) given
in MLPHandle. Before the MLP can be trained, all
training samples to be used for the training must be stored in the
MLP using add_sample_class_mlp or
read_samples_class_mlp. If after the training new
additional training samples should be used a new MLP must be created
with create_class_mlp, in which again all training
samples to be used (i.e., the original ones and the additional ones)
must be stored. In these cases, it is useful to save and read the
training data with write_samples_class_mlp and
read_samples_class_mlp, respectively. A second training
with additional training samples is not explicitly forbidden by
train_class_mlp. However, this typically does not lead to
good results because the training of an MLP is a complex nonlinear
optimization problem, and consequently the second training with new
data will very likely lead to the fact that the optimization gets
stuck in a local minimum.
If a rejection class has been specified using
set_rejection_params_class_mlp, before the actual training
the samples for the rejection class are generated.
During the training, the error the MLP achieves on the stored
training samples is minimized by using a nonlinear optimization
algorithm. If the MLP has been regularized with
set_regularization_params_class_mlp, an additional weight
penalty term is taken into account. With this, the MLP weights
described in create_class_mlp are determined. Furthermore,
if an automatic determination of the regularization parameters has
been specified with set_regularization_params_class_mlp,
these parameters are optimized as well. As described at
set_regularization_params_class_mlp, training the MLP with
automatic determination of the regularization parameters requires
significantly more time than training an unregularized MLP or an MLP
with fixed regularization parameters.
create_class_mlp initializes the MLP weights with random
values to make it very likely that the optimization converges to the
global minimum of the error function. Nevertheless, in rare cases
it may happen that the random values determined with
RandSeed in create_class_mlp result in a relatively
large optimum error, i.e., that the optimization gets stuck in a
local minimum. If it can be conjectured that this has happened the
MLP should be created anew with a different value for
RandSeed in order to check whether a significantly smaller
error can be achieved.
The parameters MaxIterations, WeightTolerance, and
ErrorTolerance control the nonlinear optimization
algorithm. Note that if an automatic determination of the
regularization parameters has been specified with
set_regularization_params_class_mlp, these parameters refer
to one training within one step of the evidence procedure.
MaxIterations specifies the maximum number of iterations of
the optimization algorithm. In practice, values between
100 and 200 should be sufficient for most
problems. WeightTolerance specifies a threshold for the
change of the weights per iteration. Here, the absolute value of
the change of the weights between two iterations is summed. Hence,
this value depends on the number of weights as well as the size of
the weights, which in turn depend on the scaling of the training
data. Typically, values between 0.00001 and 1
should be used. ErrorTolerance specifies a threshold for
the change of the error value per iteration. This value depends on
the number of training samples as well as the number of output
variables of the MLP. Also here, values between 0.00001
and 1 should typically be used. The optimization is
terminated if the weight change is smaller than
WeightTolerance and the change of the error value is
smaller than ErrorTolerance. In any case, the optimization
is terminated after at most MaxIterations iterations. It
should be noted that, depending on the size of the MLP and the
number of training samples, the training can take from a few seconds
to several hours.
On output, train_class_mlp returns the error of the MLP with
the optimal weights on the training samples in Error.
Furthermore, ErrorLog contains the error value as a
function of the number of iterations. With this, it is possible to
decide whether a second training of the MLP with the same training
data without creating the MLP anew makes sense. If
ErrorLog is regarded as a function, it should drop off
steeply initially, while leveling out very flatly at the end. If
ErrorLog is still relatively steep at the end, it usually
makes sense to call train_class_mlp again. It should be
noted, however, that this mechanism should not be used to
train the MLP successively with MaxIterations \(=\)
1 (or other small values for MaxIterations)
because this will substantially increase the number of iterations
required to train the MLP. Note that if an automatic determination
of the regularization parameters has been specified with
set_regularization_params_class_mlp, Error and
ErrorLog refer to the last training that was executed in
the evidence procedure. If the error log should be monitored within
the individual iterations of the evidence procedure, the outer
iteration of the evidence procedure must be implemented explicitly,
as described at set_regularization_params_class_mlp.
Execution information🔗
Execution information
-
Multithreading type: reentrant (runs in parallel with non-exclusive operators).
-
Multithreading scope: global (may be called from any thread).
-
Automatically parallelized on internal data level.
This operator modifies the state of the following input parameter:
During execution of this operator, access to the value of this parameter must be synchronized if it is used across multiple threads.
Parameters🔗
MLPHandle (input_control, state is modified) class_mlp → (handle)
MLP handle.
MaxIterations (input_control) integer → (integer)
Maximum number of iterations of the optimization algorithm.
Default: 200
Suggested values: 20, 40, 60, 80, 100, 120, 140, 160, 180, 200, 220, 240, 260, 280, 300
WeightTolerance (input_control) real → (real)
Threshold for the difference of the weights of the MLP between two iterations of the optimization algorithm.
Default: 1.0
Suggested values: 1.0, 0.1, 0.01, 0.001, 0.0001, 0.00001
Restriction: WeightTolerance >= 1.0e-8
ErrorTolerance (input_control) real → (real)
Threshold for the difference of the mean error of the MLP on the training data between two iterations of the optimization algorithm.
Default: 0.01
Suggested values: 1.0, 0.1, 0.01, 0.001, 0.0001, 0.00001
Restriction: ErrorTolerance >= 1.0e-8
Error (output_control) real → (real)
Mean error of the MLP on the training data.
ErrorLog (output_control) real-array → (real)
Mean error of the MLP on the training data as a function of the number of iterations of the optimization algorithm.
Example🔗
(HDevelop)
* Train an MLP
create_class_mlp (NumIn, NumHidden, NumOut, 'softmax', \
'normalization', 1, 42, MLPHandle)
read_samples_class_mlp (MLPHandle, 'samples.mtf')
train_class_mlp (MLPHandle, 100, 1, 0.01, Error, ErrorLog)
write_class_mlp (MLPHandle, 'classifier.mlp')
Result🔗
If the parameters are valid, the operator train_class_mlp
returns the value 2 (H_MSG_TRUE). If necessary, an exception is
raised.
train_class_mlp may return the error 9211 (Matrix is not
positive definite) if Preprocessing \(=\)
'canonical_variates' is used. This typically indicates
that not enough training samples have been stored for each class.
Combinations with other operators🔗
Combinations
Possible predecessors
add_sample_class_mlp, read_samples_class_mlp, set_regularization_params_class_mlp
Possible successors
evaluate_class_mlp, classify_class_mlp, write_class_mlp, create_class_lut_mlp
Alternatives
train_dl_model_batch, read_class_mlp
See also
References🔗
Christopher M. Bishop: “Neural Networks for Pattern Recognition”; Oxford University Press, Oxford; 1995.
Andrew Webb: “Statistical Pattern Recognition”; Arnold, London; 1999.
Module🔗
Foundation