Skip to content

refactor: use clip for gradient clipping #53

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
oliverdutton wants to merge 1 commit into
base: main
Choose a base branch
from
Open

refactor: use clip for gradient clipping #53

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 3 additions & 9 deletions bgflow/nn/flow/transformer/jax.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@

try:
import jax
import jax.numpy as jnp
from jax import lax, numpy as jnp
except ImportError:
jax = None
jnp = None
Expand All @@ -28,16 +28,10 @@ def affine_transform(x, a, b):
def smooth_ramp(x, logalpha, power=1, eps=1e-9):
"""Smooth ramp."""
assert power > 0
assert isinstance(power, int)
assert eps > 0
alpha = jnp.exp(logalpha)
# double `where` trick to avoid NaN in backward pass
z = jnp.where(x > eps, x, jnp.ones_like(x) * eps)
normalizer = jnp.exp(-alpha * 1.)
return jnp.where(
x > eps,
jnp.exp(-alpha * jnp.power(z, -power)) / normalizer,
jnp.zeros_like(z))
x = jnp.clip(x, a_min=eps)
return jnp.exp(alpha * (1-lax.integer_pow(x, -power)))


def monomial_ramp(x, order=2):
Expand Down