mirrors/pytorch-image-models
1
0
Fork 0
mirror of https://github.com/huggingface/pytorch-image-models.git synced 2025-06-03 15:01:08 +08:00
Code Issues Projects Releases Wiki Activity

Fix bug introduced in non layer_decay weight_decay application. Remove debug print, fix arg desc.

Browse Source
This commit is contained in:
Ross Wightman 2022-02-28 17:06:32 -08:00
parent 372ad5fa0d
commit 0557c8257d
3 changed files with 2 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
timm/models/helpers.py
View File

@ -660,7 +660,6 @@ def group_with_matcher(
for k in sorted(filter(lambda x: x is not None, grouping.keys())):
if lid < 0 or k[-1] != MATCH_PREV_GROUP[0]:
lid += 1
print(lid, k, grouping[k])
layer_id_to_param[lid].extend(grouping[k])
if reverse:

2
timm/optim/optim_factory.py
View File

@ -44,7 +44,7 @@ def param_groups_weight_decay(
if not param.requires_grad:
continue
if param.ndim or name.endswith(".bias") or name in no_weight_decay_list:
if param.ndim <= 1 or name.endswith(".bias") or name in no_weight_decay_list:
no_decay.append(param)
else:
decay.append(param)

2
train.py
View File

@ -140,7 +140,7 @@ parser.add_argument('--clip-grad', type=float, default=None, metavar='NORM',
parser.add_argument('--clip-mode', type=str, default='norm',
help='Gradient clipping mode. One of ("norm", "value", "agc")')
parser.add_argument('--layer-decay', type=float, default=None,
help='weight decay (default: None)')
help='layer-wise learning rate decay (default: None)')
# Learning rate schedule parameters
parser.add_argument('--sched', default='cosine', type=str, metavar='SCHEDULER',