torchgfn
torchgfn copied to clipboard
GFNOrg
GFlowNet library
Worth considering if we should apply to a foundation like this https://numfocus.org/project/sciml Which would also facilitate industry contributions to the codebase.
Hello everyone, Thank you for making this repo. My colleagues and I are applying GFlowNets on a hypergrid environment using our own reward distribution. However upon training, we get the...
https://github.com/GFNOrg/torchgfn/blob/d9ca55815c43e129fad95fdcb408280c90b19963/src/gfn/gflownet/flow_matching.py#L98 There should be a debug flag you can use to inspect the logits going into the loss and other issues like this
https://github.com/GFNOrg/torchgfn/blob/d9ca55815c43e129fad95fdcb408280c90b19963/src/gfn/gflownet/flow_matching.py#L42 for example, to be able to sample with some epsilon.
```Reading through the code for this file, it seems like a lot of the explicit typing of parameters on any given function is directly related to the return type of...
Such a policy would not be useful for applications, but maybe good for either verifying that A) your trained GFN is learning (relative to this baseline) or B) reporting a...
https://arxiv.org/abs/2202.08450 Ideally the oracle would be a readily available huggingface model.
Thank you for this great library. In the tutorial notebook, for computing PB of trajectory, we only consider timestep range [trajectory_length, 2). Why do we stop at time step 2?...
Here's an example of how I envision conditioning working. We want conditioning information to be passed to the gflownet from the training loop, in the form of a tensor which...