trainf_ocr_net ( : : NetId, FileName, Iteration, Reduction, Rate : Error )

Training an OCR network with the help of a training file.

The operator trainf_ocr_net trains the net NetId directly via the indicated training files. Any number of files can be indicated. The parameter Rate determines the learning rate. The smaller the value, the more slowly but also the more stably the network converges. If the training error oscillates (up and down) or increases, the rate must be reduced. A oscillating of the error may also hint to the fact that there are characters which cannot be distinguished by the network (for example 0 (zero) and the character O). An increase in number makes it necessary to reduce the learning rate. The operator trainf_ocr_net may also be called iteratively whereby the learning rate is reduced all the time. Hereby the training speed on the whole may be accelerated. The parameter Reduction indicates the criterion for termination: if the initial error times Reduction is smaller than the current error, the training will be interrupted before all iterations have been finished. The parameter Error informs about the convergence of the network after each training step. After the training, the error should lie below 0.1.


Attention

The names of the characters in the file must fit the network.


Parameters

NetId (input_control)
integer -> integer
ID of the desired OCR-network.
Default value: 'integer'

FileName (input_control)
filename(-array) -> string
Name(s) of the training file(s).
Default value: ''train_ocr''

Iteration (input_control)
integer -> integer
Maximum number of iterations.
Default value: 10
Suggested values: 1, 2, 3, 4, 5, 6, 7, 10, 15, 20, 50, 100
Range of values: 1 <= Iteration <= 100000
Minimum increment: 1
Recommended increment: 1

Reduction (input_control)
real -> real
Factor to minimize the error (terminating condition).
Default value: 0.1
Suggested values: 1.5, 2, 3, 4, 5, 6, 7, 10

Rate (input_control)
real -> real
Learning rate for back propagation.
Default value: 0.0001
Suggested values: 0.01, 0.005, 0.001, 0.0005, 0.0001, 0.00005, 0.00001
Range of values: 0.0000001 <= Rate <= 0.2
Minimum increment: 0.000001
Recommended increment: 0.00001

Error (output_control)
real-array -> real
Network error after each iteration.


Result

If the file name is correct and the data fit the network, the operator trainf_ocr_net returns the value TRUE. Otherwise an exception will be raised.


Possible Predecessors

create_ocr_net, read_ocr_net


Possible Successors

traind_ocr_net, info_ocr_net, write_ocr_net, do_ocr_b, do_ocr_m


Alternatives

trainf_ocr_net



Copyright © 1996-1997 MVTec Software GmbH