oops

Author: lucidrains

adam_atan2_pytorch/adopt.py

@@ -98,15 +98,15 @@ def step(
                 next_m = grad.div(v.sqrt().clamp(min = eps)) # they claim that a max(value, eps) performs better than adding the epsilon
 
                 if steps > 1:
-                    m.lerp_(next_m, 1. - beta2)
+                    m.lerp_(next_m, 1. - beta1)
 
                 # then update parameters
 
                 p.add_(m, alpha = -lr)
 
                 # update exp grad sq (v)
 
-                v.lerp_(grad_sq, 1. - beta1)
+                v.lerp_(grad_sq, 1. - beta2)
 
                 # increment steps
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "adam-atan2-pytorch"
-version = "0.1.2"
+version = "0.1.4"
 description = "Adam-atan2 for Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }