Figure 2.

Training and Validation Accuracy and Training and Validation Loss At 1 h of Differentiation

Several CNNs were used to train images at 1 h. All networks achieved results close to 100% of accuracy. Insets in all panels shows details of the stable phase. DenseNet architectures showed validation accuracies with a steeper curve, although reaching similar values than ResNet, in particular when simple image augmentation was used. Validation loss was slightly lower when training was done with DenseNet architectures. VGG16, a shallower architecture, could not be trained. All models were run with the same initial weights. NA, no image augmentation; SA, simple image augmentation; CA, complex image augmentation.