mahynski/pgaa-osr-mobilenetv3small-softmax

Summary

This is an OpenSetClassifier Model from PyChemAuth. The closed-set classifier component is a pre-trained (imagenet) convolutional neural network base (mobilenetv3small) with a "CAM" head (global average pooling + 1 dense layer) that was trained on the dataset below. This classified was fixed, then a softmax out-of-distribution (OOD) detector was trained to recognized OOD samples.

This model was first presented in Mahynski, N.A., Sheen, D.A., Paul, R.L. et al. "Encoding PGAA spectra as images for material classification with convolutional neural networks." J Radioanal Nucl Chem (2025).

Also see, the GitHub repository.

Training Dataset

This model was trained on the mahynski/pgaa-sample-gadf-images dataset.
Of the 13 Material classes, 3 were considered "unknown" (Carbon Powder, Phosphate Rock, Zircaloy) during training. The OSR model was still able to use these as "known unknowns" during cross-validation to select optimal OOD hyperparameters (e.g., alpha). Test set performance is illustrated below.

Classifier Training

The closed set classifier was trained using cyclical learning rates, and its performance is summarized below. The validation set used here is the test set in mahynski/pgaa-sample-gadf-images dataset, the same used to compute the confusion matrix above.

OOD Training

The classifier used here corresponds to Model 1 below; different columns indicate its combination with different OOD detectors. All "known" classes (0-9) from the training set are shown in blue, all "unknown" classes (10-12) from both the test and training sets are shown in orange as OOD.