gradboard 2.3.3__tar.gz → 2.5.0__tar.gz

{gradboard-2.3.3 → gradboard-2.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: gradboard
-Version: 2.3.3
+Version: 2.5.0
 Summary: Easily snowboard down gnarly loss gradients
 License: MIT
 Author: Nicholas Bailey

{gradboard-2.3.3 → gradboard-2.5.0}/gradboard/optimiser.py

@@ -138,15 +138,7 @@ def get_optimiser(model, optimiser=AdamW, lr=1e-3, weight_decay=1e-2):
     """
     # TODO: print a warning when params are excluded from weight decay IFF wd is set
     weight_decay_exclude = []
-    for keyword in [
-        "nondecay",
-        "bias",
-        "norm",
-        "embedding",
-        "beta",
-        "input_query",
-        "reentrant_query",
-    ]:
+    for keyword in ["nondecay", "bias", "norm", "embedding", "beta"]:
         weight_decay_exclude += [
             p for name, p in model.named_parameters() if keyword in name.lower()
         ]

{gradboard-2.3.3 → gradboard-2.5.0}/gradboard/scheduler.py

@@ -165,13 +165,11 @@ class PASS:
         self.max_lr = None
         self.cool_point = None
         for p in sorted(points_left_of_min, key=lambda x: x[0]):
-            if (self.cool_point is None) and (p[1] < minimum[1] + 0.66 * difference):
-                self.cool_point = p[0]
-            elif (self.max_lr is None) and (p[1] < minimum[1] + 0.33 * difference):
+            if (self.max_lr is None) and (p[1] < minimum[1] + 0.2 * difference):
                 self.max_lr = p[0]
             else:
                 continue
-        self.cool_point = min(self.cool_point, self.max_lr / 4)
+        self.cool_point = self.max_lr / 60
         print("High LR", self.max_lr)
         print("Cool point", self.cool_point)
 

{gradboard-2.3.3 → gradboard-2.5.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "gradboard"
-version = "2.3.3"
+version = "2.5.0"
 description = "Easily snowboard down gnarly loss gradients"
 authors = [
     {name = "Nicholas Bailey"}