I spent the weekend desiging a project to show how complex PyTorch computational graphs can be turned into collective variables inside Plumed. This was done in two parts.
This Jupyter notebook encodes Metadynamics into PyTorch using a custom loss function depenedent on the history. The forces then become the negative of the derivatives which are automatically obtained via back propagation. This was performed on the Muller potential.
This Jupyter Notebook transfers a 3-layer Image Net PyTorch Classifier into Plumed, encodes the images as molecular trajectories, and use Plumed to predict the un-normalized image scores. The plumed input file image_plumed.dat has the actual plumed neural network script.
All the code + details are available on my github.