Control Neural Network Training

Training Set

The largest single determinant affecting the network training phase is the composition of the Training Set. The best method for controlling the network training phase is to construct a Training Image Sample Set for your tool. In this way, you can separate images/views into categories that allow you to determine if your tool is generalizing your images/views properly. The use of training set is in common for all tools, but High Detail tools has another data set called "validation set" or "validation data" which is part of the training set, whose amount of data is chosen by users. For High Detail modes, the validation loss (=the loss calculated from the validation data) is calculated for each model during the training phase, and the model who gives the best validation loss in terms of performance and availability is finally selected as the result of training.

Validation Loss (High Detail Mode)

The purpose of validation data is among many neural network models generated from the training data choosing the best model as the final output of training. The training strategy that adopts validation data to achieve this goal is here called "training with validation." Unlike Focused mode tools, High Detail mode tools (Green Classify High Detail Mode and Red Analyze High Detail Mode) provide the training with validation, and you can control the network training with monitoring validation loss. During training at the end of every 1/8 epoch, the neural network calculates the loss value from the validation set you previously configured.

The validation loss stands for the performance of your trained network in terms of accuracy of classification (Green Classify High Detail Mode) or segmentation (Red Analyze High Detail Mode), which means that smaller loss generally means a better network. So it is better to have this value close to 0.

Though, to gain the full-sight regarding how your network truly performs well, you have to test the trained network against some separate data (Test Data) to prevent overfitting.

The validation loss of Green High Detail Mode is calculated per view as the classification is executed on each view. Meanwhile, The validation loss of Red Analyze High Detail Mode is calculated per pixel as the segmentation, which is the binary classification among "Good" or "Defect", is executed on each pixel.

 

Validation Loss

  • Validation Loss of Green Classify High Detail (unit: %)

    • 100 - The average value of the classification precisions of each class

  • Validation Loss of Red Analyze High Detail (unit: %)

    • 100 - The recall of "Defect" class + 40 x (100 - The recall of "Good" class)

 

Note: Red Analyze High Detail Quick and Green Classify High Detail Quick does not require the validation set and thus neither offer the validation loss nor support Loss Inspector because it leaves out of the training with validation scheme.
Tip: You can monitor the change of the validation loss in training for each High Detail mode with Loss Inspector.