gflownet
gflownet copied to clipboard
alexhernandezgarcia
Generative Flow Networks - GFlowNet
Several updates from the conformer project: * improving memory efficiency by using torch.no_grad and .detach() torch tensors where the computation graph is not needed * adding nested sampling for sampling...
Those are changes that have already been submitted in a PR for a specific project and have already been reviewed in that context. This is a PR to make those...
Scripts for plotting figures in the paper. Links to the updated plots are on Notion Command to run plotting script for Fig 3, 11-14: `PYTHONPATH=/home/mila/a/alexandra.volokhova/projects/gflownet python plots_iclm24.py --pkl_path /home/mila/a/alexandra.volokhova/projects/gflownet/external/starling_fe/samples/gfn_iter50k_samples.pkl --random_pkl_path...
There're several methods in the environments which work with states (e.g. `state2policy`, `get_parents`) and either take state as an input or use internal state of the enrolment `self.state`. This was...
**Observation:** Currently there're no documented conventions on how different methods in different modules (environments, gflownet agent, policies, proxies) interact with each other and it makes it difficult to introduce changes...
Hey Alex and others, First of all thank you for your generous work in the field and for spreading the ideas with open source code for the rest of us!...
Currently, it depends on the environment's `top_k_metrics_and_plots()`, which is obsolete since it uses the proxy directly, as thought it was still part of the environment.
This was discussed in the context of the Evaluator PR https://github.com/alexhernandezgarcia/gflownet/pull/287.
Currently, the codebase uses helper methods (`tfloat`, `tlong`, `tint`, `tbool`) to convert numbers / lists / arrays into tensors with the corresponding dtype and send them to the right device....
I believe that the complex but powerful functionality implemented in the [Batch class](https://github.com/alexhernandezgarcia/gflownet/blob/main/gflownet/utils/batch.py) could be leveraged to improve the [Buffer](https://github.com/alexhernandezgarcia/gflownet/blob/main/gflownet/utils/buffer.py). In particular, I think the Buffer could be re-implemented as...