Change Log

torchdyn 0.1.0 (April 26, 2020)

  • First commit

torchdyn 0.1.1 (April 30, 2020)

  • Added new tutorial on integral adjoint training for trajectory tracking

torchdyn 0.2.0 (July 3, 2020)

  • Introduced new CNF nn.Module for continuous normalizing flows. CNFs disentangle Jacobian trace computation from data-dynamics, allowing for convenient extension to other variants.

  • Introduced Stable, HNN, LNN energy models. These wrap the func and handle the additional autograd calls, as well as dimension bookkeeping and concatenation.

  • Added several new tutorial notebooks

  • New static datasets: gaussians, gaussians_spiral, diffeqml.

  • Improved Adjoint to handle both terminal and integral loss functions simultaneously

  • Restructured overall API, including NeuralDE

  • controlled not a setting anymore: introduction of DataControl module

  • order, solver, atol, rtol are now arguments of NeuralDE

  • DEFunc is now implicitly called inside the NeuralDE class.

  • Slimmed down NeuralDE management of correct ODE solving call.

  • New test suite for adjoint, normalizing flows and NeuralDE`.