torchgfn
torchgfn copied to clipboard
GFNOrg
GFlowNet library
For example - sampling with temperature.
In previous versions of the code, when actions were integers, we had this function that reverts backward trajectories. It's not used as part of the codebase, but I remember using...
Similar to `TrajectoryBasedGFlowNet`, which is the superclass of TB, SubTB, and LogZVar GFlowNets, meaning that the loss function decomposes into a sum of losses over trajectories, we might need to...
Currently, the flow matching loss requires a loop over all possible actions `for action_idx in range(self.env.n_actions - 1):`. This might be impractical if the number of actions blows. We might...
We need to manually write `.rst` files in the `docs` folder to get a more sophisticated documentation system. See https://pytorch-forecasting.readthedocs.io/en/stable/ for an example.
Right now `batch_dim` actually has two, specific dimensions: > would it make more sense for the batch_dim[0] to be named `batch_dim` and batch_dim[1] to be named `trajectory_dim` or something (since...