Skip to content

Paper Edition

Latest
Compare
Choose a tag to compare
@kmheckel kmheckel released this 23 Feb 18:29
· 15 commits to main since this release

Major Revamp:

  1. Replaced spyx.axn.Axon with spyx.axn.custom
  2. All preimplemented surrogate gradients are now standalone activation functions, no need to wrap them with another function
  3. spyx.axn.ActivityRegularization was moved to spyx.nn because it is a hk.Module and therefore a layer.
  4. Loss and accuracy functions converted to higher order funcs which return the version used in training loops. Time axis arg added.
  5. Notebooks in the docs were updated to reflect syntax changes.
  6. time constants are now constrained via jnp.clip which is cleaner and more efficient.
  7. Fixed bug in data shuffling and optimized it to remove an unnecessary permutation call.