Able to use other attn layer in EfficientNet now. Create test ECA + GC B0 configs. Make ECA more configurable.

tests/test_models.py

@@ -24,7 +24,7 @@ NUM_NON_STD = len(NON_STD_FILTERS)
 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',
+        '*efficientnet_l2*', '*resnext101_32x48d', '*in21k', '*152x4_bitm', '*101x3_bitm', '*50x3_bitm'
         '*nfnet_f3*', '*nfnet_f4*', '*nfnet_f5*', '*nfnet_f6*', '*nfnet_f7*', 
         '*resnetrs350*', '*resnetrs420*']
 else:

timm/models/efficientnet.py

@@ -91,6 +91,12 @@ default_cfgs = {
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/spnasnet_100-048bc3f4.pth',
         interpolation='bilinear'),
 
+    # NOTE experimenting with alternate attention
+    'eca_efficientnet_b0': _cfg(
+        url=''),
+    'gc_efficientnet_b0': _cfg(
+        url=''),
+
     'efficientnet_b0': _cfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/efficientnet_b0_ra-3dd342df.pth'),
     'efficientnet_b1': _cfg(
@@ -1223,6 +1229,24 @@ def efficientnet_b0(pretrained=False, **kwargs):
     return model
 
 
+@register_model
+def eca_efficientnet_b0(pretrained=False, **kwargs):
+    """ EfficientNet-B0 w/ ECA attn """
+    # NOTE experimental config
+    model = _gen_efficientnet(
+        'eca_efficientnet_b0', se_layer='eca', channel_multiplier=1.0, depth_multiplier=1.0, pretrained=pretrained, **kwargs)
+    return model
+
+
+@register_model
+def gc_efficientnet_b0(pretrained=False, **kwargs):
+    """ EfficientNet-B0 w/ GlobalContext """
+    # NOTE experminetal config
+    model = _gen_efficientnet(
+        'gc_efficientnet_b0', se_layer='gc', channel_multiplier=1.0, depth_multiplier=1.0, pretrained=pretrained, **kwargs)
+    return model
+
+
 @register_model
 def efficientnet_b1(pretrained=False, **kwargs):
     """ EfficientNet-B1 """

timm/models/efficientnet_builder.py

@@ -278,9 +278,9 @@ class EfficientNetBuilder:
         self.norm_layer = norm_layer
         self.se_layer = get_attn(se_layer)
         try:
-            self.se_layer(8, rd_ratio=1.0)
+            self.se_layer(8, rd_ratio=1.0)  # test if attn layer accepts rd_ratio arg
             self.se_has_ratio = True
-        except RuntimeError as e:
+        except TypeError:
             self.se_has_ratio = False
         self.drop_path_rate = drop_path_rate
         if feature_location == 'depthwise':

timm/models/layers/eca.py

@@ -38,6 +38,9 @@ from torch import nn
 import torch.nn.functional as F
 
 
+from .create_act import create_act_layer
+
+
 class EcaModule(nn.Module):
     """Constructs an ECA module.
 
@@ -48,20 +51,27 @@ class EcaModule(nn.Module):
             refer to original paper https://arxiv.org/pdf/1910.03151.pdf
             (default=None. if channel size not given, use k_size given for kernel size.)
         kernel_size: Adaptive selection of kernel size (default=3)
+        gamm: used in kernel_size calc, see above
+        beta: used in kernel_size calc, see above
+        act_layer: optional non-linearity after conv, enables conv bias, this is an experiment
+        gate_layer: gating non-linearity to use
     """
-    def __init__(self, channels=None, kernel_size=3, gamma=2, beta=1):
+    def __init__(self, channels=None, kernel_size=3, gamma=2, beta=1, act_layer=None, gate_layer='sigmoid'):
         super(EcaModule, self).__init__()
-        assert kernel_size % 2 == 1
         if channels is not None:
             t = int(abs(math.log(channels, 2) + beta) / gamma)
             kernel_size = max(t if t % 2 else t + 1, 3)
-
-        self.conv = nn.Conv1d(1, 1, kernel_size=kernel_size, padding=(kernel_size - 1) // 2, bias=False)
+        assert kernel_size % 2 == 1
+        has_act = act_layer is not None
+        self.conv = nn.Conv1d(1, 1, kernel_size=kernel_size, padding=(kernel_size - 1) // 2, bias=has_act)
+        self.act = create_act_layer(act_layer) if has_act else nn.Identity()
+        self.gate = create_act_layer(gate_layer)
 
     def forward(self, x):
         y = x.mean((2, 3)).view(x.shape[0], 1, -1)  # view for 1d conv
         y = self.conv(y)
-        y = y.view(x.shape[0], -1, 1, 1).sigmoid()
+        y = self.act(y)  # NOTE: usually a no-op, added for experimentation
+        y = self.gate(y).view(x.shape[0], -1, 1, 1)
         return x * y.expand_as(x)
 
 
@@ -86,27 +96,35 @@ class CecaModule(nn.Module):
             refer to original paper https://arxiv.org/pdf/1910.03151.pdf
             (default=None. if channel size not given, use k_size given for kernel size.)
         kernel_size: Adaptive selection of kernel size (default=3)
+        gamm: used in kernel_size calc, see above
+        beta: used in kernel_size calc, see above
+        act_layer: optional non-linearity after conv, enables conv bias, this is an experiment
+        gate_layer: gating non-linearity to use
     """
 
-    def __init__(self, channels=None, kernel_size=3, gamma=2, beta=1):
+    def __init__(self, channels=None, kernel_size=3, gamma=2, beta=1, act_layer=None, gate_layer='sigmoid'):
         super(CecaModule, self).__init__()
-        assert kernel_size % 2 == 1
         if channels is not None:
             t = int(abs(math.log(channels, 2) + beta) / gamma)
             kernel_size = max(t if t % 2 else t + 1, 3)
+        has_act = act_layer is not None
+        assert kernel_size % 2 == 1
 
         # PyTorch circular padding mode is buggy as of pytorch 1.4
         # see https://github.com/pytorch/pytorch/pull/17240
         # implement manual circular padding
-        self.conv = nn.Conv1d(1, 1, kernel_size=kernel_size, padding=0, bias=False)
         self.padding = (kernel_size - 1) // 2
+        self.conv = nn.Conv1d(1, 1, kernel_size=kernel_size, padding=0, bias=has_act)
+        self.act = create_act_layer(act_layer) if has_act else nn.Identity()
+        self.gate = create_act_layer(gate_layer)
 
     def forward(self, x):
         y = x.mean((2, 3)).view(x.shape[0], 1, -1)
         # Manually implement circular padding, F.pad does not seemed to be bugged
         y = F.pad(y, (self.padding, self.padding), mode='circular')
         y = self.conv(y)
-        y = y.view(x.shape[0], -1, 1, 1).sigmoid()
+        y = self.act(y)  # NOTE: usually a no-op, added for experimentation
+        y = self.gate(y).view(x.shape[0], -1, 1, 1)
         return x * y.expand_as(x)