Emmanuel Bengio
Emmanuel Bengio
By default z.lua [writes into `~/.zlua`]( https://github.com/skywind3000/z.lua/blob/master/z.lua#L107). This pollutes the home directory. In Unix, writing in a [specified config folder](https://specifications.freedesktop.org/basedir-spec/basedir-spec-latest.html) is preferable. Traditionally `~/.config/` serves this purpose. In Windows `%LOCALAPPDATA%`...
This PR implements a better way of sharing torch tensors between process by creating (large enough) shared tensors that are created once are used as a transfer mechanism. Doing this...
This PR makes it possible to have a separate set of parameters for P_B by setting `cfg.model.do_separate_p_b`. It also introduces `cfg.algo.min_len`, which defaults to `0` and prevents and agent from...
Current multiprocessing/threading routines are not explicitly stopped, they just rely on the objects they belong to to be garbage collected to stop. This sometimes causes aesthetically displeasing logs where all...
I've been using this environment to "debug" gflownets. PR not ready yet but this branch may be useful for others.
Masks are currently used somewhat ad-hoc here and there, in particular: - `GraphSampler` explicitly uses them, but it should be abstracted away - `GraphTransformerGFN` and derivatives end up applying masks...
We currently pass around conditional information (`cond_info`) as dicts where side information is stored under certain keys, and the model eventually receives just the `encoding`. It would be relevant to...
In the original (Bengio et al. 2021) fragment environment, some fragments had stem atoms that could be used multiple times (e.g. carbon atoms with more than one open valence electron)....
It's not clear that we are doing multiprocessing in the best way possible. - we have been seeing problems with too many shared tensors being created, the current patch to...