train_dl_classifier_batch — Perform a training step of a deep-learning-based classifier on a batch of
images.
train_dl_classifier_batch(BatchImages : : DLClassifierHandle, BatchLabels : DLClassifierTrainResultHandle)
train_dl_classifier_batch performs a training step of the
deep-learning-based classifier contained in DLClassifierHandle.
The classifier handle DLClassifierHandle has to be read previously
using read_dl_classifier.
In order to apply training steps, classes have to be specified using
set_dl_classifier_param.
Other hyperparameters such as the learning rate and the momentum are also
important for a successful training. They are set using
set_dl_classifier_param.
The training step is done on basis of a single batch of images from the
training dataset, thus the images BatchImages with labels
BatchLabels. The number of images within the batch needs to be
a multiple of the 'batch_size' where the parameter
'batch_size' is limited by the amount of available GPU memory.
In order to process more images in one training step, the classifier
parameter 'batch_size_multiplier' can be set to a value greater
than 1. The number of images being passed to the training operator
needs to be equal to 'batch_size' times
'batch_size_multiplier'. Note that a training step calculated
for a batch and a 'batch_size_multiplier' greater 1 is
an approximation of a training step calculated for the same batch but with
a 'batch_size_multiplier' equal to 1 and an accordingly
greater 'batch_size'. As an example, the loss calculated
with a 'batch_size' of 4 and a 'batch_size_multiplier'
of 2 is usually not equal to the loss calculated with a
'batch_size' of 8 and a 'batch_size_multiplier' of 1,
although the same number of images is used for training in both cases.
However, the approximation generally delivers comparably good results,
so it can be utilized if you wish to train with a larger number of images
than your GPU allows.
In some rare cases the approximation with a 'batch_size' of 1
and an accordingly large 'batch_size_multiplier' does not
show the expected performance which for example can happen when the
pretrained network 'pretrained_dl_classifier_resnet50.hdl' is used.
Setting the 'batch_size' to a value greater than 1 can help to
solve this issue.
Note that the images in BatchImages must fulfill certain conditions
regarding, for example, the image size and gray value range, depending on
the chosen network. Please have a look at read_dl_classifier and
set_dl_classifier_param for more information.
The labels in BatchLabels can be handed over as an array of
strings, or as an array of indices corresponding to the position of the
label within the array of classes (counting from 0) set before via
'classes' with set_dl_classifier_param.
Information about the results of the training step as the value of the loss
are stored in DLClassifierTrainResultHandle and
can be accessed using get_dl_classifier_train_result.
Note that an epoch generally consists of a large number of batches and
that a successful training involves many epochs. Therefore
train_dl_classifier_batch has to be applied several times
with different batches.
For a more detailed explanation, we refer to
Deep Learning / Classification.
During training, a nonlinear optimization algorithm minimizes the value of the loss function. The later one is determined based on the prediction of the neural network on the current batch of images. The algorithm used for optimization is stochastic gradient descent (SGD). It updates the layers' weights of the previous iteration to the new values at iteration as follows:
Here, is the learning rate, for the momentum, and for the classification result of the deep learning-based classifier which depends on the network weights and the input batch . The variable is used to involve the influence of the momentum . The loss function used here is the Multinomial Logistic Loss in combination with a quadratic regularization term ,
Here, is a one-hot encoded target vector that encodes the
label of the -th image of the batch
containing -many images,
and shall be understood to be
a vector such that is applied on each component of
.
The regularization term is a weighted
-norm involving all weights except for biases.
Its influence can be controlled through .
In the above formula, denotes the hyperparameter
'weight_prior' that can be set with
set_dl_classifier_param.
In order to gain more insight, you can retrieve the current value of the
total loss function as well as individual contributions using
get_dl_classifier_train_result.
For an explanation of the concept of deep-learning-based classification see the introduction of chapter Deep Learning / Classification.
The operator train_dl_classifier_batch internally calls functions
that might not be deterministic. Therefore, results from multiple calls of
train_dl_classifier_batch can slightly differ, although the same
input values have been used.
To run this operator, cuDNN and cuBLAS are required.
For further details, please refer to the “Installation Guide”,
paragraph “Requirements for Deep Learning”.
This operator returns a handle. Note that the state of an instance of this handle type may be changed by specific operators even though the handle is used as an input parameter by those operators.
BatchImages (input_object) (multichannel-)image(-array) → object (real)
Images comprising the batch.
DLClassifierHandle (input_control) dl_classifier → (handle)
Handle of the deep-learning-based classifier.
BatchLabels (input_control) string-array → (string / integer)
Corresponding labels for each of the images.
Default value: []
List of values: []
DLClassifierTrainResultHandle (output_control) dl_classifier_train_result → (handle)
Handle of the training results from the deep-learning-based classifier.
If the parameters are valid, the operator
train_dl_classifier_batch returns the value 2 (H_MSG_TRUE). If
necessary, an exception is raised.
read_dl_classifier,
set_dl_classifier_param,
get_dl_classifier_param
get_dl_classifier_train_result,
apply_dl_classifier,
clear_dl_classifier_train_result,
clear_dl_classifier
train_class_mlp,
train_class_svm
Deep Learning Training