torchgfn
torchgfn copied to clipboard
GFNOrg
GFlowNet library
When using a conditional gflownet, we need the LogZ estimate (in algorithms such as TB) to be conditional, i.e., predicted from the state. This means the `logz` parameter must not...
Hello and thank you for developing the torchgfn package! It's been a fantastic resource, and I appreciate all the work that's gone into it. I'm exploring the package and am...
In `detailed_balance.py`, we have: ``` if not self.off_policy: valid_log_pf_actions = transitions.log_probs else: # Evaluate the log PF of the actions sampled off policy. # I suppose the Transitions container should...
A recent PR introduced a bug where train_hypergrid doesn't converge if replay_buffer is set to 100. This should be caught with the tests !
Hello, Thank you for your effort on releasing such great implementation of GFN! I am working on the using GFN to finetune an LLM to be a policy model (which...
Currently the code in `multinode` depends on a private dependency - we should replace it with this: https://github.com/intel/torch-ccl
+ Right now, to do on-policy, or off policy, or "true" off policy sampling efficiently, you need to set "recalculate" flags appropriately. + If you do this incorrectly, you end...
We need clear instructions on common operations, for example like this: https://github.com/GFNOrg/torchgfn/issues/148#issuecomment-1949424048 I'll add a todo list later.
Running the second code cell returns `TypeError: Can't instantiate abstract class Line with abstract methods backward_step, step` This is likely due to the recent changes in the environment API
Idea for v2: having a utility class to compose environments like ```python atoms_env = ... lattice_env = ... spacegroup_env = ... crystal_env = ComposeEnvs(atoms_env, lattice_env, spacegroup_env, shuffle=True) ``` Where the...
