Currently, the library supports only PyTorch models built with custom layers, custom loss functions, activation functions, and any neural network function subclassing AutoGrad.