gradboard 0.1.21__tar.gz → 1.0.0__tar.gz

{gradboard-0.1.21 → gradboard-1.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: gradboard
-Version: 0.1.21
+Version: 1.0.0
 Summary: Easily snowboard down gnarly loss gradients
 License: MIT
 Author: Nicholas Bailey

{gradboard-0.1.21 → gradboard-1.0.0}/gradboard/optimiser.py

@@ -25,9 +25,7 @@ class AdamS(Optimizer):
         weight_decay (float, optional): weight decay coefficient (default: 1e-4)
     """
 
-    def __init__(
-        self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=1e-4
-    ):
+    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.1):
         if not 0.0 <= lr:
             raise ValueError("Invalid learning rate: {}".format(lr))
         if not 0.0 <= eps:
@@ -38,6 +36,7 @@ class AdamS(Optimizer):
             raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
         if not 0.0 <= weight_decay:
             raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        weight_decay *= lr
         defaults = {"lr": lr, "betas": betas, "eps": eps, "weight_decay": weight_decay}
         super().__init__(params, defaults)
 

{gradboard-0.1.21 → gradboard-1.0.0}/gradboard/scheduler.py

@@ -7,7 +7,6 @@ import copy
 import math
 
 from scipy.ndimage import gaussian_filter1d
-import scipy.constants as sc
 
 from torch.amp import GradScaler
 
@@ -164,15 +163,15 @@ class PASS:
         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 + 0.75 * loss_difference
-        max_lr_loss = min_loss + 0.25 * loss_difference
+        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] * sc.golden
+                self.cool_point = r[0]
                 break
         for r in range_test_results:
             if r[1] <= max_lr_loss:
-                self.max_lr = r[0] / sc.golden
+                self.max_lr = r[0]
                 break
         print("High LR", self.max_lr)
         print("Cool point", self.cool_point)

{gradboard-0.1.21 → gradboard-1.0.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "gradboard"
-version = "0.1.21"
+version = "1.0.0"
 description = "Easily snowboard down gnarly loss gradients"
 authors = [
     {name = "Nicholas Bailey"}