gradboard 2.2.0__py3-none-any.whl → 2.3.0__py3-none-any.whl

gradboard/optimiser.py

@@ -142,9 +142,7 @@ def get_optimiser(model, optimiser=AdamW, lr=1e-3, weight_decay=1e-2):
         "norm",
         "embedding",
         "swiglu_beta",
-        "sigma",
-        "scale",
-        "input_query",
+        "sigma_reparam" "input_query",
         "reentrant_query",
     ]:
         weight_decay_exclude += [

gradboard/scheduler.py

@@ -159,23 +159,16 @@ class PASS:
         """
         range_test_results = self._smoothed_range_test(self.range_test_results)
         minimum = min(range_test_results, key=lambda x: x[1])
-        max_left_of_min = (0, 0)
-        for r in range_test_results:
-            if r[0] > minimum[0]:
-                break
-            elif r[1] > max_left_of_min[1]:
-                max_left_of_min = r
-            else:
-                continue
-        difference = max_left_of_min[1] - minimum[1]
         points_left_of_min = [r for r in range_test_results if r[0] < minimum[0]]
+        max_left_of_min = max(points_left_of_min, key=lambda x: x[1])
+        difference = max_left_of_min[1] - minimum[1]
         self.max_lr = None
         self.cool_point = None
-        for p in points_left_of_min:
-            if (self.cool_point is None) and (p[1] > minimum[1] + 0.66 * difference):
-                self.cool_point = r[0]
-            elif (self.max_lr is None) and (p[1] > minimum[1] + 0.33 * difference):
-                self.max_lr = r[0]
+        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):
+                self.max_lr = p[0]
             else:
                 continue
         self.cool_point = min(self.cool_point, self.max_lr / 4)

{gradboard-2.2.0.dist-info → gradboard-2.3.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: gradboard
-Version: 2.2.0
+Version: 2.3.0
 Summary: Easily snowboard down gnarly loss gradients
 License: MIT
 Author: Nicholas Bailey

gradboard-2.3.0.dist-info/RECORD

@@ -0,0 +1,8 @@
+gradboard/__init__.py,sha256=57AkHusYwLCsusiVnajH5pMFKioRCj-3IjF9qpdOzE0,69
+gradboard/cycles.py,sha256=iGEW3Rlp-JNyQZLfpaDXxUCAcEV01ANjLF-Fnhug-qA,10120
+gradboard/optimiser.py,sha256=b29ir7lvCWZuC6yctQ1unRUeIMHednn0aWF9NkWH5fQ,6183
+gradboard/scheduler.py,sha256=TlfdIGx5ufb5dXrhWbDkpePfqLMoHJSGZbdauPgd2FU,6770
+gradboard-2.3.0.dist-info/LICENSE,sha256=0BAzJE5BqQ7Iixp_AFdB2W1uO-HCRX-Qfun8PHt6yVM,1073
+gradboard-2.3.0.dist-info/METADATA,sha256=bPw6qxnTxFFGSbtd6e6CsebFYjpFZH8gy1AYv55-8Vc,2173
+gradboard-2.3.0.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
+gradboard-2.3.0.dist-info/RECORD,,

gradboard-2.2.0.dist-info/RECORD

@@ -1,8 +0,0 @@
-gradboard/__init__.py,sha256=57AkHusYwLCsusiVnajH5pMFKioRCj-3IjF9qpdOzE0,69
-gradboard/cycles.py,sha256=iGEW3Rlp-JNyQZLfpaDXxUCAcEV01ANjLF-Fnhug-qA,10120
-gradboard/optimiser.py,sha256=h5nntenAhD8E9-Ayt17r7AK42S7O-XdhsrxedDd_4co,6201
-gradboard/scheduler.py,sha256=kRcJGmxZ8lYghz-zgyySXXuQnQUpObhRQ_0wknmQlEs,6921
-gradboard-2.2.0.dist-info/LICENSE,sha256=0BAzJE5BqQ7Iixp_AFdB2W1uO-HCRX-Qfun8PHt6yVM,1073
-gradboard-2.2.0.dist-info/METADATA,sha256=jC6UQTUHCyyeE8SxbekSafXT_dAXM-0deE0p1rDjTb0,2173
-gradboard-2.2.0.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
-gradboard-2.2.0.dist-info/RECORD,,