dandip/dsrpytorch

Pytorch implementation of Deep Symbolic Regression

dsrpytorch's Issues

Several bugs for the use of CUDA device

There are several places where the pytorch Tensor is initialize without point out the device. Thus, when I want to run it on CUDA, these cpu Tensors will be incompatible to the CUDA ones. Additionally, in the code the Tensors are usually treated as numpy arrays and use the numpy functions to handle the Tensors. This is nothing wrong when the Tensors are on cpu, but will raise error when the tensors are on CUDA.

Looking for help on how to add constraints

Dear Daniel,
Thank you very much for your contributions to the PyTorch version of DSR. In practice, I found that the constraints on trigonometric nesting (e.g. $sin(x+cos(x))$ is not allowed) and log&exp (e.g. $log(exp(x))$ is not allowed) mentioned in the paper can drastically reduce the search space, which improves the search efficiency. I have tried to add both constraints to this code because of my current scientific work, but have been unsuccessful due to my limited coding skills. So I would like to ask for your help, and I would appreciate it if you could help me to add these two constraints!
Thank you again for your contributions!
Best regards.