Tweak some comments, add SKNet models with weights to sotabench, remove an unused branch

sotabench.py

@@ -56,8 +56,7 @@ model_list = [
            model_desc='Trained from scratch in PyTorch w/ RandAugment'),
     _entry('efficientnet_es', 'EfficientNet-EdgeTPU-S', '1905.11946',
            model_desc='Trained from scratch in PyTorch w/ RandAugment'),
-    _entry('fbnetc_100', 'FBNet-C', '1812.03443',
+
-           model_desc='Trained in PyTorch with RMSProp, exponential LR decay'),
     _entry('gluon_inception_v3', 'Inception V3', '1512.00567', model_desc='Ported from GluonCV Model Zoo'),
     _entry('gluon_resnet18_v1b', 'ResNet-18', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
     _entry('gluon_resnet34_v1b', 'ResNet-34', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
@@ -82,14 +81,22 @@ model_list = [
     _entry('gluon_seresnext101_64x4d', 'SE-ResNeXt-101 64x4d', '1812.01187', model_desc='Ported from GluonCV Model Zoo'),
     _entry('gluon_xception65', 'Modified Aligned Xception', '1802.02611', batch_size=BATCH_SIZE//2,
            model_desc='Ported from GluonCV Model Zoo'),
+
     _entry('mixnet_xl', 'MixNet-XL', '1907.09595', model_desc="My own scaling beyond paper's MixNet Large"),
     _entry('mixnet_l', 'MixNet-L', '1907.09595'),
     _entry('mixnet_m', 'MixNet-M', '1907.09595'),
     _entry('mixnet_s', 'MixNet-S', '1907.09595'),
+
+    _entry('fbnetc_100', 'FBNet-C', '1812.03443',
+           model_desc='Trained in PyTorch with RMSProp, exponential LR decay'),
     _entry('mnasnet_100', 'MnasNet-B1', '1807.11626'),
+    _entry('semnasnet_100', 'MnasNet-A1', '1807.11626'),
+    _entry('spnasnet_100', 'Single-Path NAS', '1904.02877',
+           model_desc='Trained in PyTorch with SGD, cosine LR decay'),
     _entry('mobilenetv3_rw', 'MobileNet V3-Large 1.0', '1905.02244',
            model_desc='Trained in PyTorch with RMSProp, exponential LR decay, and hyper-params matching '
                       'paper as closely as possible.'),
+
     _entry('resnet18', 'ResNet-18', '1812.01187'),
     _entry('resnet26', 'ResNet-26', '1812.01187', model_desc='Block cfg of ResNet-34 w/ Bottleneck'),
     _entry('resnet26d', 'ResNet-26-D', '1812.01187',
@@ -103,7 +110,7 @@ model_list = [
     _entry('resnext50d_32x4d', 'ResNeXt-50-D 32x4d', '1812.01187',
            model_desc="'D' variant (3x3 deep stem w/ avg-pool downscale). Trained with "
                       "SGD w/ cosine LR decay, random-erasing (gaussian per-pixel noise) and label-smoothing"),
-    _entry('semnasnet_100', 'MnasNet-A1', '1807.11626'),
+
     _entry('seresnet18', 'SE-ResNet-18', '1709.01507'),
     _entry('seresnet34', 'SE-ResNet-34', '1709.01507'),
     _entry('seresnext26_32x4d', 'SE-ResNeXt-26 32x4d', '1709.01507',
@@ -114,8 +121,9 @@ model_list = [
            model_desc='Block cfg of SE-ResNeXt-34 w/ Bottleneck, deep tiered stem, and avg-pool in downsample layers.'),
     _entry('seresnext26tn_32x4d', 'SE-ResNeXt-26-TN 32x4d', '1812.01187',
            model_desc='Block cfg of SE-ResNeXt-34 w/ Bottleneck, deep tiered narrow stem, and avg-pool in downsample layers.'),
-    _entry('spnasnet_100', 'Single-Path NAS', '1904.02877',
+
-           model_desc='Trained in PyTorch with SGD, cosine LR decay'),
+    _entry('skresnet18', 'SK-ResNet-18', '1903.06586'),
+    _entry('skresnext50_32x4d', 'SKNet-50', '1903.06586'),
 
     _entry('tf_efficientnet_b0', 'EfficientNet-B0 (AutoAugment)', '1905.11946',
            model_desc='Ported from official Google AI Tensorflow weights'),

timm/models/layers/cbam.py

@@ -2,6 +2,9 @@
 
 Experimental impl of CBAM: Convolutional Block Attention Module: https://arxiv.org/abs/1807.06521
 
+WARNING: Results with these attention layers have been mixed. They can significantly reduce performance on
+some tasks, especially fine-grained it seems. I may end up removing this impl.
+
 Hacked together by Ross Wightman
 """
 

timm/models/layers/cond_conv2d.py

@@ -1,4 +1,7 @@
-""" Conditional Convolution
+""" PyTorch Conditionally Parameterized Convolution (CondConv)
+
+Paper: CondConv: Conditionally Parameterized Convolutions for Efficient Inference
+(https://arxiv.org/abs/1904.04971)
 
 Hacked together by Ross Wightman
 """
@@ -28,7 +31,7 @@ def get_condconv_initializer(initializer, num_experts, expert_shape):
 
 
 class CondConv2d(nn.Module):
-    """ Conditional Convolution
+    """ Conditionally Parameterized Convolution
     Inspired by: https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/condconv/condconv_layers.py
 
     Grouped convolution hackery for parallel execution of the per-sample kernel filters inspired by this discussion:

timm/models/layers/eca.py

@@ -42,7 +42,7 @@ class EcaModule(nn.Module):
     """Constructs an ECA module.
 
     Args:
-        channel: Number of channels of the input feature map for use in adaptive kernel sizes
+        channels: Number of channels of the input feature map for use in adaptive kernel sizes
             for actual calculations according to channel.
             gamma, beta: when channel is given parameters of mapping function
             refer to original paper https://arxiv.org/pdf/1910.03151.pdf

timm/models/layers/mixed_conv2d.py

@@ -1,4 +1,6 @@
-""" Conditional Convolution
+""" PyTorch Mixed Convolution
+
+Paper: MixConv: Mixed Depthwise Convolutional Kernels (https://arxiv.org/abs/1907.09595)
 
 Hacked together by Ross Wightman
 """

timm/models/layers/selective_kernel.py

@@ -1,4 +1,6 @@
-""" Selective Kernel Convolution Attention
+""" Selective Kernel Convolution/Attention
+
+Paper: Selective Kernel Networks (https://arxiv.org/abs/1903.06586)
 
 Hacked together by Ross Wightman
 """

timm/models/sknet.py

@@ -1,3 +1,9 @@
+""" Selective Kernel Networks (ResNet base)
+
+Paper: Selective Kernel Networks (https://arxiv.org/abs/1903.06586)
+
+Hacked together by Ross Wightman
+"""
 import math
 
 from torch import nn as nn
@@ -47,19 +53,11 @@ class SelectiveKernelBasic(nn.Module):
         outplanes = planes * self.expansion
         first_dilation = first_dilation or dilation
 
-        _selective_first = True  # FIXME temporary, for experiments
+        self.conv1 = SelectiveKernelConv(
-        if _selective_first:
+            inplanes, first_planes, stride=stride, dilation=first_dilation, **conv_kwargs, **sk_kwargs)
-            self.conv1 = SelectiveKernelConv(
+        conv_kwargs['act_layer'] = None
-                inplanes, first_planes, stride=stride, dilation=first_dilation, **conv_kwargs, **sk_kwargs)
+        self.conv2 = ConvBnAct(
-            conv_kwargs['act_layer'] = None
+            first_planes, outplanes, kernel_size=3, dilation=dilation, **conv_kwargs)
-            self.conv2 = ConvBnAct(
-                first_planes, outplanes, kernel_size=3, dilation=dilation, **conv_kwargs)
-        else:
-            self.conv1 = ConvBnAct(
-                inplanes, first_planes, kernel_size=3, stride=stride, dilation=first_dilation, **conv_kwargs)
-            conv_kwargs['act_layer'] = None
-            self.conv2 = SelectiveKernelConv(
-                first_planes, outplanes, dilation=dilation, **conv_kwargs, **sk_kwargs)
         self.se = create_attn(attn_layer, outplanes)
         self.act = act_layer(inplace=True)
         self.downsample = downsample
@@ -222,7 +220,7 @@ def skresnet50d(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
 @register_model
 def skresnext50_32x4d(pretrained=False, num_classes=1000, in_chans=3, **kwargs):
     """Constructs a Select Kernel ResNeXt50-32x4d model. This should be equivalent to
-    the SKNet50 model in the Select Kernel Paper
+    the SKNet-50 model in the Select Kernel Paper
     """
     default_cfg = default_cfgs['skresnext50_32x4d']
     model = ResNet(