gradboard 0.2.0__tar.gz → 1.0.1__tar.gz

{gradboard-0.2.0 → gradboard-1.0.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: gradboard
-Version: 0.2.0
+Version: 1.0.1
 Summary: Easily snowboard down gnarly loss gradients
 License: MIT
 Author: Nicholas Bailey

{gradboard-0.2.0 → gradboard-1.0.1}/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)
 
@@ -130,7 +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-0.2.0 → gradboard-1.0.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "gradboard"
-version = "0.2.0"
+version = "1.0.1"
 description = "Easily snowboard down gnarly loss gradients"
 authors = [
     {name = "Nicholas Bailey"}