BotNet models were still off, remove weights for bad configs. Add good SE-HaloNet33-TS weights.

Author: rwightman

timm/models/byoanet.py

@@ -36,22 +36,22 @@ def _cfg(url='', **kwargs):
     'botnet26t_256': _cfg(
         url='',
         fixed_input_size=True, input_size=(3, 256, 256), pool_size=(8, 8)),
-    'botnet50t_256': _cfg(
-        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/botnet50t_256-a0e6c3b1.pth',
+    'botnet50ts_256': _cfg(
+        url='',
         fixed_input_size=True, input_size=(3, 256, 256), pool_size=(8, 8)),
     'eca_botnext26ts_256': _cfg(
         url='',
         fixed_input_size=True, input_size=(3, 256, 256), pool_size=(8, 8)),
-    'eca_botnext50ts_256': _cfg(
-        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/eca_botnext26ts_256-fb3bf984.pth',
-        fixed_input_size=True, input_size=(3, 256, 256), pool_size=(8, 8)),
 
     'halonet_h1': _cfg(url='', input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256)),
     'halonet26t': _cfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/halonet26t_256-9b4bf0b3.pth',
         input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256)),
-    'sehalonet33ts': _cfg(url='', input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256)),
-    'halonet50ts': _cfg(url='', input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256)),
+    'sehalonet33ts': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/sehalonet33ts_256-87e053f9.pth',
+        input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256), crop_pct=0.94),
+    'halonet50ts': _cfg(
+        url='', input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256)),
     'eca_halonext26ts': _cfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-attn-weights/eca_halonext26ts_256-1e55880b.pth',
         input_size=(3, 256, 256), pool_size=(8, 8), min_input_size=(3, 256, 256)),
@@ -78,16 +78,17 @@ def _cfg(url='', **kwargs):
         self_attn_layer='bottleneck',
         self_attn_kwargs=dict()
     ),
-    botnet50t=ByoModelCfg(
+    botnet50ts=ByoModelCfg(
         blocks=(
             ByoBlockCfg(type='bottle', d=3, c=256, s=1, gs=0, br=0.25),
-            ByoBlockCfg(type='bottle', d=4, c=512, s=2, gs=0, br=0.25),
-            interleave_blocks(types=('bottle', 'self_attn'), d=2, c=1024, s=2, gs=0, br=0.25),
-            ByoBlockCfg(type='self_attn', d=3, c=2048, s=2, gs=0, br=0.25),
+            interleave_blocks(types=('bottle', 'self_attn'), every=4, d=4, c=512, s=2, gs=0, br=0.25),
+            interleave_blocks(types=('bottle', 'self_attn'), d=6, c=1024, s=2, gs=0, br=0.25),
+            interleave_blocks(types=('bottle', 'self_attn'), d=3, c=2048, s=2, gs=0, br=0.25),
         ),
         stem_chs=64,
         stem_type='tiered',
         stem_pool='maxpool',
+        act_layer='silu',
         fixed_input_size=True,
         self_attn_layer='bottleneck',
         self_attn_kwargs=dict()
@@ -108,22 +109,6 @@ def _cfg(url='', **kwargs):
         self_attn_layer='bottleneck',
         self_attn_kwargs=dict()
     ),
-    eca_botnext50ts=ByoModelCfg(
-        blocks=(
-            ByoBlockCfg(type='bottle', d=3, c=256, s=1, gs=16, br=0.25),
-            ByoBlockCfg(type='bottle', d=4, c=512, s=2, gs=16, br=0.25),
-            interleave_blocks(types=('bottle', 'self_attn'), d=2, c=1024, s=2, gs=16, br=0.25),
-            ByoBlockCfg(type='self_attn', d=3, c=2048, s=2, gs=16, br=0.25),
-        ),
-        stem_chs=64,
-        stem_type='tiered',
-        stem_pool='maxpool',
-        fixed_input_size=True,
-        act_layer='silu',
-        attn_layer='eca',
-        self_attn_layer='bottleneck',
-        self_attn_kwargs=dict()
-    ),
 
     halonet_h1=ByoModelCfg(
         blocks=(
@@ -227,38 +212,31 @@ def _create_byoanet(variant, cfg_variant=None, pretrained=False, **kwargs):
 
 @register_model
 def botnet26t_256(pretrained=False, **kwargs):
-    """ Bottleneck Transformer w/ ResNet26-T backbone. Bottleneck attn in final two stages.
-    FIXME 26t variant was mixed up with 50t arch cfg, retraining and determining why so low
+    """ Bottleneck Transformer w/ ResNet26-T backbone.
+    NOTE: this isn't performing well, may remove
     """
     kwargs.setdefault('img_size', 256)
     return _create_byoanet('botnet26t_256', 'botnet26t', pretrained=pretrained, **kwargs)
 
 
 @register_model
-def botnet50t_256(pretrained=False, **kwargs):
-    """ Bottleneck Transformer w/ ResNet50-T backbone. Bottleneck attn in final two stages.
+def botnet50ts_256(pretrained=False, **kwargs):
+    """ Bottleneck Transformer w/ ResNet50-T backbone, silu act.
+    NOTE: this isn't performing well, may remove
     """
     kwargs.setdefault('img_size', 256)
-    return _create_byoanet('botnet50t_256', 'botnet50t', pretrained=pretrained, **kwargs)
+    return _create_byoanet('botnet50ts_256', 'botnet50ts', pretrained=pretrained, **kwargs)
 
 
 @register_model
 def eca_botnext26ts_256(pretrained=False, **kwargs):
-    """ Bottleneck Transformer w/ ResNet26-T backbone, silu act, Bottleneck attn in final two stages.
-    FIXME 26ts variant was mixed up with 50ts arch cfg, retraining and determining why so low
+    """ Bottleneck Transformer w/ ResNet26-T backbone, silu act.
+    NOTE: this isn't performing well, may remove
     """
     kwargs.setdefault('img_size', 256)
     return _create_byoanet('eca_botnext26ts_256', 'eca_botnext26ts', pretrained=pretrained, **kwargs)
 
 
-@register_model
-def eca_botnext50ts_256(pretrained=False, **kwargs):
-    """ Bottleneck Transformer w/ ResNet26-T backbone, silu act, Bottleneck attn in final two stages.
-    """
-    kwargs.setdefault('img_size', 256)
-    return _create_byoanet('eca_botnext50ts_256', 'eca_botnext50ts', pretrained=pretrained, **kwargs)
-
-
 @register_model
 def halonet_h1(pretrained=False, **kwargs):
     """ HaloNet-H1. Halo attention in all stages as per the paper.