Trajectory balance: Improved credit assignment in GFlowNets
Authors: Nikolay Malkin, Moksh Jain, Emmanuel Bengio, Chen Sun, Yoshua Bengio
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments on four distinct domains, we empirically demonstrate the benefits of the trajectory balance objective for GFlow Net convergence, diversity of generated samples, and robustness to long action sequences and large action spaces. |
| Researcher Affiliation | Collaboration | Nikolay Malkin Mila, Université de Montréal Montréal, Québec, Canada Moksh Jain Mila, Université de Montréal Montréal, Québec, Canada Emmanuel Bengio Mila, Mc Gill University, Recursion Montréal, Québec, Canada Chen Sun Mila, Université de Montréal Montréal, Québec, Canada Yoshua Bengio Mila, Université de Montréal Montréal, Québec, Canada |
| Pseudocode | Yes | Algorithm 1 Training a GFlow Net with trajectory balance |
| Open Source Code | Yes | Code: https://gist.github.com/malkin1729/9a87ce4f19acdc2c24225782a8b81c15. Code: https://github.com/GFNOrg/gflownet/tree/trajectory_balance. |
| Open Datasets | Yes | We take 6438 known AMP sequences and 9522 non-AMP sequences from the DBAASP database [23]. |
| Dataset Splits | Yes | We then train a classifier on this dataset, using 20% of the data as a validation set. |
| Hardware Specification | No | The paper states: "This research was enabled in part by computational resources provided by Compute Canada." This is too general and does not provide specific hardware details like GPU/CPU models or memory. |
| Software Dependencies | No | The paper mentions using a "Transformer-based architecture" but does not specify any software names with version numbers (e.g., PyTorch 1.9, TensorFlow 2.x, or specific library versions). |
| Experiment Setup | Yes | We train GFlow Nets with the detailed balance (DB) and trajectory balance (TB) objectives with different H, D, and R0. Details are given in B.1. In Appendix B.4, we provide further details regarding the training setup and hyperparameters for all models. |