gradboard 1.0.0__tar.gz → 1.0.2__tar.gz

{gradboard-1.0.0 → gradboard-1.0.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: gradboard
-Version: 1.0.0
+Version: 1.0.2
 Summary: Easily snowboard down gnarly loss gradients
 License: MIT
 Author: Nicholas Bailey

{gradboard-1.0.0 → gradboard-1.0.2}/gradboard/optimiser.py

@@ -129,7 +129,7 @@ class AdamS(Optimizer):
         return loss
 
 
-def get_optimiser(model, optimiser=AdamW, lr=7e-4, weight_decay=5e-2):
+def get_optimiser(model, optimiser=AdamW, lr=1e-3, weight_decay=1e-2):
     """
     Defaults are from one of the presets from the accompanying repo to Hassani
         et al. (2023) "Escaping the Big Data Paradigm with Compact Transformers",

{gradboard-1.0.0 → gradboard-1.0.2}/gradboard/scheduler.py

@@ -159,20 +159,10 @@ class PASS:
         """
         range_test_results = self._smoothed_range_test(self.range_test_results)
         minimum = min(range_test_results, key=lambda x: x[1])
-        min_loss = minimum[1]
-        points_left_of_min = [p for p in range_test_results if p[0] < minimum[0]]
-        highest_point_left_of_min = max(points_left_of_min, key=lambda x: x[1])
-        loss_difference = highest_point_left_of_min[1] - minimum[1]
-        cool_point_loss = min_loss + 0.8 * loss_difference
-        max_lr_loss = min_loss + 0.2 * loss_difference
         for r in range_test_results:
-            if r[1] < cool_point_loss:
-                self.cool_point = r[0]
-                break
-        for r in range_test_results:
-            if r[1] <= max_lr_loss:
-                self.max_lr = r[0]
-                break
+            if (r[0] < minimum[0]) and (r[1] > minimum[1]):
+                self.max_lr = r[0] / 2
+                self.cool_point = self.max_lr * 0.1
         print("High LR", self.max_lr)
         print("Cool point", self.cool_point)
 

{gradboard-1.0.0 → gradboard-1.0.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "gradboard"
-version = "1.0.0"
+version = "1.0.2"
 description = "Easily snowboard down gnarly loss gradients"
 authors = [
     {name = "Nicholas Bailey"}