[WIP] Add ResNet-RS models (huggingface#554)

* Add ResNet-RS models

* Only include resnet-rs changes

* remove whitespace diff

* EOF newline

* Update time

* increase time

* Add first conv

* Try running only resnetv2_101x1_bitm on Linux runner

* Add to exclude filter

* Run test_model_forward_features for all

* Add to exclude ftrs

* back to defaults

* only run test_forward_features

* run all tests

* Run all tests

* Add bigger resnetrs to model filters to fix Github CLI

* Remove resnetv2_101x1_bitm from exclude feat features

* Remove hardcoded values

* Make sure reduction ratio in resnetrs is 0.25

* There is no bias in replaced maxpool so remove it

Author: amaarora

tests/test_models.py

@@ -22,8 +22,9 @@
 if 'GITHUB_ACTIONS' in os.environ:  # and 'Linux' in platform.system():
     # GitHub Linux runner is slower and hits memory limits sooner than MacOS, exclude bigger models
     EXCLUDE_FILTERS = [
-        '*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm', '*101x3_bitm',
-        '*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*'] + NON_STD_FILTERS
+        '*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm',
+        '*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*', 
+        '*resnetrs200*', '*resnetrs270*', '*resnetrs350*', '*resnetrs420*'] + NON_STD_FILTERS
 else:
     EXCLUDE_FILTERS = NON_STD_FILTERS
 

timm/models/resnet.py

@@ -236,7 +236,23 @@ def _cfg(url='', **kwargs):
         interpolation='bicubic'),
     'resnetblur50': _cfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnetblur50-84f4748f.pth',
-        interpolation='bicubic')
+        interpolation='bicubic'),
+
+    # ResNet-RS models
+    'resnetrs50': _cfg(
+        interpolation='bicubic', first_conv='conv1.0'),
+    'resnetrs101': _cfg(
+        interpolation='bicubic', first_conv='conv1.0'),
+    'resnetrs152': _cfg(
+        interpolation='bicubic', first_conv='conv1.0'),
+    'resnetrs200': _cfg(
+        interpolation='bicubic', first_conv='conv1.0'),
+    'resnetrs270': _cfg(
+        interpolation='bicubic', first_conv='conv1.0'),
+    'resnetrs350': _cfg(
+        interpolation='bicubic', first_conv='conv1.0'),
+    'resnetrs420': _cfg(
+        interpolation='bicubic', first_conv='conv1.0'),
 }
 
 
@@ -318,7 +334,7 @@ class Bottleneck(nn.Module):
 
     def __init__(self, inplanes, planes, stride=1, downsample=None, cardinality=1, base_width=64,
                  reduce_first=1, dilation=1, first_dilation=None, act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d,
-                 attn_layer=None, aa_layer=None, drop_block=None, drop_path=None):
+                 attn_layer=None, aa_layer=None, drop_block=None, drop_path=None, **kwargs):
         super(Bottleneck, self).__init__()
 
         width = int(math.floor(planes * (base_width / 64)) * cardinality)
@@ -341,7 +357,7 @@ def __init__(self, inplanes, planes, stride=1, downsample=None, cardinality=1, b
         self.conv3 = nn.Conv2d(width, outplanes, kernel_size=1, bias=False)
         self.bn3 = norm_layer(outplanes)
 
-        self.se = create_attn(attn_layer, outplanes)
+        self.se = create_attn(attn_layer, outplanes, **kwargs)
 
         self.act3 = act_layer(inplace=True)
         self.downsample = downsample
@@ -545,11 +561,12 @@ def __init__(self, block, layers, num_classes=1000, in_chans=3,
                  cardinality=1, base_width=64, stem_width=64, stem_type='',
                  output_stride=32, block_reduce_first=1, down_kernel_size=1, avg_down=False,
                  act_layer=nn.ReLU, norm_layer=nn.BatchNorm2d, aa_layer=None, drop_rate=0.0, drop_path_rate=0.,
-                 drop_block_rate=0., global_pool='avg', zero_init_last_bn=True, block_args=None):
+                 drop_block_rate=0., global_pool='avg', zero_init_last_bn=True, block_args=None, replace_stem_max_pool=False):
         block_args = block_args or dict()
         assert output_stride in (8, 16, 32)
         self.num_classes = num_classes
         self.drop_rate = drop_rate
+        self.replace_stem_max_pool = replace_stem_max_pool
         super(ResNet, self).__init__()
 
         # Stem
@@ -574,12 +591,19 @@ def __init__(self, block, layers, num_classes=1000, in_chans=3,
         self.feature_info = [dict(num_chs=inplanes, reduction=2, module='act1')]
 
         # Stem Pooling
-        if aa_layer is not None:
-            self.maxpool = nn.Sequential(*[
-                nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
-                aa_layer(channels=inplanes, stride=2)])
+        if not self.replace_stem_max_pool:
+            if aa_layer is not None:
+                self.maxpool = nn.Sequential(*[
+                    nn.MaxPool2d(kernel_size=3, stride=1, padding=1),
+                    aa_layer(channels=inplanes, stride=2)])
+            else:
+                self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
         else:
-            self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+            self.maxpool = nn.Sequential(*[
+                nn.Conv2d(inplanes, inplanes, 3, stride=2, padding=1, bias=False), 
+                norm_layer(inplanes), 
+                act_layer(inplace=True)
+            ])
 
         # Feature Blocks
         channels = [64, 128, 256, 512]
@@ -1065,6 +1089,63 @@ def ecaresnet50d(pretrained=False, **kwargs):
     return _create_resnet('ecaresnet50d', pretrained, **model_args)
 
 
+@register_model
+def resnetrs50(pretrained=False, **kwargs):
+    model_args = dict(
+        block=Bottleneck, layers=[3, 4, 6, 3], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
+        avg_down=True,  block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
+    return _create_resnet('resnetrs50', pretrained, **model_args)
+
+
+@register_model
+def resnetrs101(pretrained=False, **kwargs):
+    model_args = dict(
+        block=Bottleneck, layers=[3, 4, 23, 3], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
+        avg_down=True,  block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
+    return _create_resnet('resnetrs101', pretrained, **model_args)
+
+
+@register_model
+def resnetrs152(pretrained=False, **kwargs):
+    model_args = dict(
+        block=Bottleneck, layers=[3, 8, 36, 3], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
+        avg_down=True,  block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
+    return _create_resnet('resnetrs152', pretrained, **model_args)
+
+
+@register_model
+def resnetrs200(pretrained=False, **kwargs):
+    model_args = dict(
+        block=Bottleneck, layers=[3, 24, 36, 3], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
+        avg_down=True,  block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
+    return _create_resnet('resnetrs200', pretrained, **model_args)
+
+
+@register_model
+def resnetrs270(pretrained=False, **kwargs):
+    model_args = dict(
+        block=Bottleneck, layers=[4, 29, 53, 4], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
+        avg_down=True,  block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
+    return _create_resnet('resnetrs270', pretrained, **model_args)
+
+
+
+@register_model
+def resnetrs350(pretrained=False, **kwargs):
+    model_args = dict(
+        block=Bottleneck, layers=[4, 36, 72, 4], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
+        avg_down=True,  block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
+    return _create_resnet('resnetrs350', pretrained, **model_args)
+
+
+@register_model
+def resnetrs420(pretrained=False, **kwargs):
+    model_args = dict(
+        block=Bottleneck, layers=[4, 44, 87, 4], stem_width=32, stem_type='deep', replace_stem_max_pool=True,
+        avg_down=True,  block_args=dict(attn_layer='se', reduction_ratio=0.25), **kwargs)
+    return _create_resnet('resnetrs420', pretrained, **model_args)
+
+
 @register_model
 def ecaresnet50d_pruned(pretrained=False, **kwargs):
     """Constructs a ResNet-50-D model pruned with eca.