Closed hayesall closed 3 years ago
Would you please explain about extra functionalities of nnde with this version of Autograd (not the PyTorch's one) and JAX considering JOSS guidelines?
We updated the paper to address this issue by describing how the code was started before the widespread adoption of TensorFlow 2. We began to convert the software to use TensorFlow 2 several months ago, and early results are promising, so we wanted to document this particular stage in the project prior to further development.
The "Statement of need" (lines 19-22) mentioned that existing deep learning frameworks can speed up computation, but cannot perform analytical derivatives needed for the trial function method implemented by
nnde
.Can this point be clarified or expanded?
For example: PyTorch implements methods for automatic differentiation (autograd) and numerical methods to compute gradients. What is an example where these are insufficient?