Clothing type classifier - machine learning module for WearMeUp.Club

In the experiments, 12 machine learning classifiers based on convolutional neural networks (CNNs) were trained.

Many well-known (state of the art, SOTA) CNN network structures such as ResNet50, Vgg16, Inception v3, EfficientNet were tested during the training.

In addition, dedicated structures were programmed from scratch. The models were tested with a variable size of the first layer (input layer) of the neural network, and more precisely the sizes are as follows: 96×96, 160×160, 224×224.

The training process took place on an adapted machine containing a graphics processing unit (GPU), in order to speed up the training process.

A report was prepared for each trained model, which includes documentation of the trained model. The documentation includes the full structure of the model, the layers, their dimensions and the number of parameters. The report includes accuracy results obtained on the test set for the following measures: accuracy, balanced accuracy, precision, recall, measure f1.

A total of 221 training and testing processes were carried out, and the aforementioned report was prepared for each of them.

In addition, at the analysis stage, the full structure including weights was saved in .keras format for the models that scored highest on the balanced accuracy measure (the data were moderately unbalanced).