Model Training Panel

Lists all of the selected training sets, which include inputs, outputs, and masks. Options in the Training Data list include the following:

Show Training Data Statistics… Click the Show Statistics button to open the Training Data Statistics dialog, shown below. The dialog, which is available for segmentation models only, lists the total number of voxels in the training set(s) added to the model, the percentage of labeled training voxels, and the percentage of labeled voxels associated with each class. Cumulative statistics for all training data is also presented. You should note that 'Class 1' is always associated with the background, and that all statistics are computed within selected masks.

Note This option is NOT available for regression models.

Add New Training Dataset… Click the Add New button to add a new item to the Training Data list. The input(s) and output for the item can be selected in the Input drop-down menu, as shown below.

Remove New Training Dataset… Click the Remove button to remove the selected item from the Training Data list.

Lets you select a target output for training. You should note that outputs are dependent on the type of model that will be trained and must be same size and shape as the input data for training semantic segmentation and denoising models. Outputs for super-resolution can be a factor — 2, 4, or 8 times — of the input X-Y dimension.

If you are training a model for continuous output, for example with an autoencoder for denoising or super-resolution, you have to select an image channel as the output. If you are training a model for semantic segmentation, you will need to select a multi-ROI with the same number of classes as the model's 'class count' (see Model Generator).

If selected, data augmentation will be applied during training (see Data Augmentation Settings). Data augmentation is a technique that can be used to artificially expand the size of a training dataset by creating modified versions of the images in the dataset

Note The current data augmentation settings are applied to all new models by default.

Lets you choose how many times each data patch is augmented during a single training epoch. You should note that at a setting of '1', the amount of training data will be doubled, at a setting of '2', the training data will be tripled, and so on.

Flips data patches vertically by reversing the rows of pixels.

Randomly shears data patches clockwise by a given number of degrees from 0 to the set maximum.

Maximum degrees… Lets you set the maximum number of degrees to shear images. The maximum number of degrees that the data can be sheared is 45.

Randomly darkens images within a specified range, for example between 0.2 or 20% and 1.0 (no change). In this case, the intent is to allow a model to generalize across images trained on different illumination levels.

Randomly adds an elastic transformation, within a specified range, to the image.

Note Computing elastic transformations is computationally expensive and selecting this option will likely increase training times significantly.

Lets you automatically split the model inputs into training and validation sets.

Percentage of training data to be used for validation… Lets you choose the percentage of the training data that will be used for validation. In this case, the training set will be used for neural network training and the validation set will be used only for accuracy evaluation, but not for training.

During training, training data is split into smaller 2D data patches, which is defined by the 'Input (Patch) Size' parameter.

For example, if you choose an Input (Patch) Size of 64, the Deep Learning Tool will cut the dataset into sub-sections of 64´64 pixels. These subsections will then be used as the training dataset. By subdividing images, each pass or 'epoch' should be faster and use less memory.

A single pass over all the data patches is called epoch, and the number of epochs is controlled by the 'Epochs Number' parameter.

Loss functions, which are selectable in the drop-down menu, measure the error between the neural network's prediction and reality. The error is then used to update the model parameters (go to www.tensorflow.org/api_docs/python/tf/keras/losses for additional information about the loss functions available in Dragonfly's Deep Learning Tool).

You should note that not all the loss functions will work well with all models and the available selections are automatically filtered according to the model type — Regression (for super-resolution and denoising) and Semantic Segmentation (for binary and multi-class segmentations).

Regressive loss functions… Are used in cases of regressive problems, that is when the target variable is continuous. One of the most widely used regressive loss functions is Mean Squared Error. Other loss functions you might consider are Cosine Similarity, Huber, Mean Absolute Error, Poisson, and others listed in the drop-down menu (see Loss Functions for Regression Models).

Semantic segmentation loss functions… Are used in cases of segmentation problems, that is when the target output is a multi-ROI. When training a multi-class segmentation model, 'CategoricalCrossentropy' is generally a good choice as a classification for each pixel must be made. See Loss Functions for Semantic Segmentation Models for additional information about the available loss functions.

Displays the estimated memory ratio, which is calculated as the ratio of your system's capability and the estimated memory needed to train the model at the current settings. You should note that the total memory requirements to train a model depends on the implementation and selected optimizer. In some cases, the size of the network may be bound by your system's available memory.

Green … The estimated memory requirements are within your system's capabilities.

Yellow … The estimated memory requirements are approaching your system's capabilities.

Red … The estimated memory requirements exceed your system's capabilities. You should consider adjusting the model training parameters or selecting a shallower model.

Note Memory is one of the biggest challenges in training deep neural networks. Memory is required to store input data, weight parameters and activations as an input propagates through the network. In training, activations from a forward pass must be retained until they can be used to calculate the error gradients in the backwards pass. As an example, the 50-layer ResNet network has about 26 million weight parameters and computes close to 16 million activations in the forward pass. If you use a 32-bit floating-point value to store each weight and activation this would give a total storage requirement of 168 MB. By using a lower precision value to store these weights and activations you could halve or even quarter this storage requirement.

Note Refer to imatge-upc.github.io/telecombcn-2016-dlcv/slides/D2L1-memory.pdf for information about calculating memory requirements.