Re-order decay/no-decay groups to match old order and pass existing test. Change end lr decay test to min scale.

Author: rwightman

tests/test_optim.py

@@ -552,7 +552,7 @@ def test_lookahead_radam(optimizer):
     )
 
 
-def test_param_groups_layer_decay_with_end_decay():
+def test_param_groups_layer_decay_with_min():
     model = torch.nn.Sequential(
         torch.nn.Linear(10, 5),
         torch.nn.ReLU(),
@@ -563,12 +563,12 @@ def test_param_groups_layer_decay_with_end_decay():
         model,
         weight_decay=0.05,
         layer_decay=0.75,
-        end_layer_decay=0.5,
+        min_scale=0.5,
         verbose=True
     )
 
     assert len(param_groups) > 0
-    # Verify layer scaling is applied with end decay
+    # Verify layer scaling is applied with a min scale
     for group in param_groups:
         assert 'lr_scale' in group
         assert group['lr_scale'] <= 1.0

timm/optim/_param_groups.py

@@ -49,14 +49,14 @@ def param_groups_weight_decay(
                 decay.append(param)
 
     groups = []
-    if decay:
-        groups.append({'params': decay, 'weight_decay': weight_decay})
-    if decay_simple:
-        groups.append({'params': decay_simple, 'weight_decay': weight_decay, 'simple': True})
     if no_decay:
         groups.append({'params': no_decay, 'weight_decay': 0.})
+    if decay:
+        groups.append({'params': decay, 'weight_decay': weight_decay})
     if no_decay_simple:
         groups.append({'params': no_decay_simple, 'weight_decay': 0., 'simple': True})
+    if decay_simple:
+        groups.append({'params': decay_simple, 'weight_decay': weight_decay, 'simple': True})
 
     return groups