Can compute parameter derivatives #143
Conversation
|
We really need to fix the CI... |
| void TorchForce::addEnergyParameterDerivative(const string& name) { | ||
| for (int i = 0; i < globalParameters.size(); i++) | ||
| if (name == globalParameters[i].name) { | ||
| energyParameterDerivatives.push_back(i); |
There was a problem hiding this comment.
I have seen this in other parts of OpenMM, what happens if I call this function twice with the same name? Is that handled somewhere before or after this?
| torch::Tensor& forceTensor) { | ||
| torch::Tensor& forceTensor, map<string, torch::Tensor>& derivInputs) { | ||
| vector<torch::Tensor> gradInputs; | ||
| if (!outputsForces) |
There was a problem hiding this comment.
This should be !outputsForces && includeForces right?
|
There should also be a Python side test. Take this one if you'd like: BTW, this is still a problem. I do not think OpenMM-Torch can do anything about it, but perhaps there is a way to detect it and provide a useful message? |
|
Thanks for the comments. It turned out that what I had written didn't work with CUDA graphs. I restructured it to handle the input tensors in a different way and made the test case run both with and without graphs. Can you see if it looks better now? |
|
Is this ok to merge? Getting CI working is its own major project. I'm working on that in another PR. |
Implements #141.