Training the Model

Figure \ref{912281} shows the schematic representation of the model architecture for the classification of COVID-19 samples. The model includes two parts: 1) a pre-trained CNN model for feature extraction and 2) a fully-connected network for classification. Under pre-trained models, users can bring their own TensorFlow-based custom models or use system default models for different computer vision tasks (image classification, object detection, and action detection). The default model for classification is GoogLeNet, which is a convolutional neural network with 22 layers. Users have access to a pool of the GoogLeNet base models pre-trained on various data sets (Link). These data sets include different categories of images for action, flower, food, landscape, scene, and vehicle. In this project, we loaded a GoogLeNet base model pre-trained on the ImageNet dataset. Besides the many choices for pre-trained models, users can also change the model hyperparameters in the Advanced settings. These hyperparameters include model features such as max iterations, learning rate, weight decay, and so on. The dataset will be automatically split for internal validation of the model's performance during training; the default split is 80/20.