Crystal environment with continuous lattice parameters #221
Conversation
…ontinuous hypercube.
…ns, logprobs of increments and log of the diagonal of the Jacobian.
…d the comparisons with the source by taking into account the effective dimensions only.
…y of the Cube about ignored dimensions.
…gflownet into cont-lattice-params
…merge both branches.
…rt of the distr params dictionaries.
[👶 PR] Batched eval sampling
| the underlying environments, which should lead to every padded action being | ||
| unique. | ||
| """ | ||
| return -(self.value + 2) |
There was a problem hiding this comment.
I'm not sure about having this method here. It makes Stage (who's purpose is simply to represent the different possible stages) dependant on the internal logic of the crystal class, creating unnecessary coupling
There was a problem hiding this comment.
(same for the from_pad() method)
| """ | ||
| if self.value - 1 < 0: | ||
| return Stage.DONE | ||
| return Stage(self.value - 1) |
There was a problem hiding this comment.
Not sure that these methods next() and prev() should exist here given that we want eventually the option to sample space groups first. The logic should probably be at the level of the crystal/ccrystal classes.
Btw, this is how I did it in #199 to allow the two different stage orderings to coexist. I we change this implementation to methods in the crystal class, we could do the change in a way compatible with that PR.
Ccrystal fix tests
…eing loaded as expected
…ing loaded as expected
Fix: get_logprobs of base env (reason why TB was not training well)
At least according to the tests the tricky parts seem alright. Missing things are:
Builds upon (at least) the following PRs: