Tags give the ability to mark specific points in history as being important
-
0.4.2
Release: v0.4.224f8a69b · ·v0.4.2 Making progress towards v0.4.0 goals. - Functions are moved from projected_gradients to matg_gdm. - Both have identical return types, although matg_gdm does not take desired_gap as argument. - test_scale_invariant ensures that the solver can handle rewards in [0,\infty). This was a really interesting bug in ExtendNE. - API cleanup has also been done: removing unnecessary `num_actions` argumetns when they can be deduced, especially in utils.py. - algorithm.py has also been cleaned up; intermediate functions have been removed/inlined when needed, stray jit calls have been pruned. - some more use of jaxtyping as well. TODO: - Experiments and benchmarks are broken, might need fixing later. - Consider wrapping all algorithm parameters in chex DataClasses. - Expand the test-suite.
-
0.4.0
Release: 0.4.06bfde37f · ·v0.4.0 - compilation and QoL The changelog revolves mostly around efficiency improvements, both for the compilation times as well as execution. - IMPORTANT !!! ProjectedGradients now runs on GPU. This was made possible by a whole suite of improvements w.r.t. to the value function calculation, compilation graph simplification (no more decisions on the hot-path vis-à-vis what to execute), and some use of jax.lax.stop_gradient to better indicate what we need/don't need. - Both versions of ProjectedGradients (w or w/o early-stopping) benefit from this. In particular, the early-stopping also times out at a maximum number of iterations - this was not possible before since I didn't know about jnp.logical_and (JAX and python's `and` do not get along well). TODO: - standardize notation: move away from projected_gradients towards matg_gdm. - Clean up the return-types of both versions - they are identical for now, but move to a dictionary for readability. - testing via pytest. Making the execution identical across both versions is an end-goal. There are minor differences in stopping iterations, but nothing too serious. - documentation and CI/CD.
-